318 | Breaking Analysis | Forget AGI…The Prize is Enterprise AGI

We believe much of the artificial intelligence industry is chasing the wrong prize. Frontier model vendors such as Anthropic PBC and OpenAI Group PBC, they may have shifted their commercial focus toward enterprise customers, but they’ve not changed their fundamental architecture.

Specifically, they’re still trying to concentrate ever more intelligence inside a generalized model. We agree with Databricks Inc. Chief Executive Ali Ghodsi that the practical definition of artificial general intelligence has actually been achieved. Moving the goalpost to superintelligence — or what we’ve called Messiah AGI in a prior Breaking Analysis — does little to create differentiation for enterprise customers.

The real prize as we see it is what we call enterprise AGI. What do we mean by that? Specifically, we’re talking about intelligence that is unique to and owned by each enterprise.

Enterprise AGI is all about harmonizing proprietary data, business processes, policies and this tacit human knowledge that we frequently discuss on Breaking Analysis. The idea is to then turn those artifacts into persistent assets over which models and agents can reason and ultimately act. The frontier model is an important ingredient in this equation, but the enterprise system of intelligence, what we call the SoI — think enterprise ontology or digital twin) is the linchpin of achieving enterprise AGI. This is where true business value is derived for enterprises. This is distinct from “data communism,” a term we’ve introduced before. Data communism is where everyone consumes essentially the same embedded intelligence. Rather, we’re advocates of “data capitalism,” in which each company controls, governs and advances its own differentiated intelligence.

In this week’s Breaking Analysis, we build on our previous works and further discuss the implications of what we heard at Databricks’ Data + AI Summit and put these findings in the context of our SoI framework.

The Real Prize Is Enterprise AGI

Our opening visual uses George’s Wile E. Coyote metaphor to frame the central argument. The graphic shows Sam Altman as racing past the enterprise opportunity and off the superintelligence cliff, while Ali Ghodsi and Satya Nadella look on from the edge. The joke carries a more serious point in that Databricks, Microsoft and other enterprise platforms are competing in a different race – one centered on owning the system of intelligence layer that connects enterprise data, business meaning and agentic action.

Key Points

• The AI industry has pivoted commercially toward the enterprise, but not all vendors have pivoted architecturally.
• While increasingly focused on enterprise markets, frontier model vendors are still primarily building ever-smarter generalized models.
• Enterprise AGI is not just a model plus enterprise data – it is the ability to turn proprietary data, business processes and tacit knowledge into governed assets.
• Those assets become the foundation that allows models to reason and agents to act with enterprise-specific context.
• The system of intelligence is the linchpin of Enterprise AGI because it captures how an individual enterprise actually works.
• Databricks, Microsoft, Snowflake, Google, AWS, Palantir, Salesforce, SAP and others are converging on this opportunity from different starting points.
• This episode will use Databricks’ latest announcements, especially Genie Ontology, to assess where the market is heading and what remains missing.

The premise of this Breaking Analysis is that much of the AI industry has focused on the wrong finish line. For the past several years, the dominant narrative was AGI, then superintelligence (i.e. Messiah AGI). In our view, that race has already moved beyond the point of practical enterprise relevance. The real economic prize is not generalized intelligence in the abstract – it is Enterprise AGI.

This distinction is important. Many frontier model vendors have clearly pivoted to the enterprise because that is where the money is. But as we often emphasize, that does not mean they have pivoted to the architecture required to win the enterprise. Selling smarter models into companies is not the same as building enterprise intelligence. The frontier labs are still largely trying to concentrate more intelligence inside the model. Enterprise AGI requires something different in our view. Specifically, capturing a company’s proprietary data, business processes and tacit knowledge as governed assets that models can reason over and agents can act upon.

That is the point of the opening visual. Sam Altman, portrayed as Wile E. Coyote, is racing past the enterprise opportunity and off the superintelligence cliff. Ali Ghodsi and Satya Nadella are positioned like the Road Runner, standing at the edge and smirking because they understand that the real battle is not simply about having the smartest frontier model. It is about owning the enterprise intelligence layer – the layer that harmonizes data, encodes business meaning, incorporates tacit knowledge and provides the guardrails for agents to operate safely and effectively.

The traditional enterprise vendors and modern data platforms increasingly understand that the path to durable advantage is to turn data, processes and institutional knowledge into assets. Those assets become both the context and the control plane for AI. They allow agents to reason over the business, act within enterprise-specific constraints and coordinate with humans to achieve collective outcomes. This cannot be done with siloed models alone.

Palantir pioneered much of this hard work by connecting proprietary data, operational processes and domain knowledge into a model of how the enterprise runs. Now the broader market is converging on the same destination. The frontier model vendors want to rule the enterprise. The SaaS vendors are under pressure to move beyond systems of record. The hyperscalers see the control point forming above infrastructure. And the modern data platforms – represented by Databricks and Snowflake – are trying to turn governed data foundations into systems of intelligence.

This Breaking Analysis will define Enterprise AGI, examine Databricks’ latest moves through the lens of our system of intelligence framework, assess the maturity of Databricks’ Genie Ontology and show how this emerging architecture could reshape enterprise software.  The important question we’ll address is which vendors can help enterprises turn their unique data, processes and tacit knowledge into compounding, governed intelligence assets.

Key Takeaway:
The enterprise AI prize will not be won by generalized intelligence alone. It will be won by platforms that convert each enterprise’s proprietary data, processes and tacit knowledge into a system of intelligence that agents can reason over and act through.

Approach #1: Data Communism

The graphic below introduces the first approach to building enterprise intelligence – what we call data communism. The visual deliberately combines the frontier-model worldview with a Marxian metaphor – i.e. the world’s smartest people contribute their reasoning traces to frontier models, and the resulting intelligence is packaged into the model and distributed back to everyone. The promise is a powerful common intelligence layer. The problem is that everyone gets essentially the same built-in intelligence.

Key Points

• Frontier model vendors have moved toward enterprise customers, but their core architecture still concentrates intelligence inside the model.
• The first generation of models largely learned from broadly available internet data.
• The next phase requires far more specialized reasoning traces – for example, how an investment banking analyst values an acquisition.
• Those traces are expensive, narrow and domain-specific, but once bottled into the model they become shared intelligence available to all customers.
• This creates a common capability layer, not enterprise-specific differentiation.
• Selling a smarter frontier model to the enterprise is not the same as creating enterprise intelligence.
• The Enterprise AGI prize requires intelligence that is particular to each enterprise – grounded in its data, processes and tacit knowledge.

The first architecture for enterprise intelligence is what we call data communism. The phrase is intentionally catchy, but it captures an important market dynamic. The frontier-model view assumes that the world’s best intelligence should be absorbed into ever-smarter models and then distributed back broadly to users and enterprises. In theory, this creates a powerful shared intelligence layer. In practice, it means every enterprise receives access to essentially the same embedded intelligence.

That is the nuance we believe the market is under-appreciating. A frontier model can become extraordinarily capable and still fail to understand how a specific enterprise operates. The model may know general finance, general sales, general software engineering, general healthcare workflows and general customer support patterns. But general intelligence is not the same as specific enterprise intelligence.

The first several generations of frontier models were trained largely on broadly available internet data. That approach produced remarkable general-purpose capabilities because the models absorbed enormous amounts of publicly available knowledge. But as frontier model vendors push deeper into enterprise use cases, the next frontier of training data becomes much more specialized. It is not enough to learn from the public web. The models need reasoning traces from expert work.

An investment banking example makes the point. To automate or augment the work of an analyst valuing an acquisition, a model needs more than generic financial knowledge. It needs examples of the analyst’s workflow, assumptions, judgment calls, data sources, intermediate reasoning and preferred outputs. That kind of reasoning trace is expensive to generate, narrow in scope and deeply specialized. The same pattern applies across other domains – insurance underwriting, supply chain planning, fraud investigation, clinical operations, field service, procurement and countless other enterprise workflows.

But there’s is the catch. Once that specialized knowledge is captured and bottled into a frontier model, it becomes part of a shared model capability. It may improve the model for everyone, but it does not create proprietary advantage for any one enterprise. Everyone gets the same built-in intelligence. The model becomes smarter, but the enterprise does not necessarily become more differentiated.

This is why we believe the “all intelligence in one model” architecture is insufficient for Enterprise AGI. It can create a powerful baseline, but it cannot fully encode the unique way a company operates – its proprietary data, internal processes, decision rights, risk controls, customer definitions, pricing logic, approval flows, tribal knowledge and operating constraints. Those are not generic assets. They are enterprise-specific assets.

The critical distinction is that frontier models can supply generalized reasoning, but Enterprise AGI requires each company to capture its own operating knowledge as data – and to treat that knowledge as a governed, reusable, compounding asset. Without that enterprise-specific layer, companies are simply renting generalized intelligence that competitors can access as well.

That is the limitation of data communism. It lifts the floor for everyone, but it does not raise the ceiling for a particular enterprise. The real prize is not merely a smarter model. It is a system that learns how each enterprise actually works.

Key Takeaway:
Data communism creates a common intelligence layer, but common intelligence is not competitive advantage. Enterprise AGI requires each organization to turn its own data, processes and tacit knowledge into proprietary intelligence assets.

Approach #2: Data Capitalism – Enterprise Intelligence as a Proprietary Asset

This graphic below presents the alternative to data communism – i.e. data capitalism. Frontier models remain fundamental, but generalized intelligence will be widely available. The differentiating attribute is the intelligence that is particular to each enterprise – its proprietary data, business processes, policies, tacit knowledge and operating model. The stack below shows how those assets flow from data platforms and systems of record into a system of intelligence, then into systems of agency and engagement.

Key Points

• Generalized intelligence raises the floor for everyone, but it does not create unique enterprise advantage.
• Competitive advantage comes from proprietary data, business processes, corporate rules and tacit knowledge.
• The system of intelligence is the critical layer because it becomes the digital representation of how the enterprise actually operates.
• Data platforms and systems of record tell us what happened; the system of intelligence explains why it happened, what is likely to happen and what should be done next.
• The system of agency allows agents to act through that intelligence, not around it.
• The system of engagement becomes the new work surface where people, agents, intent and outcomes come together.
• The more profound insight is that the system of engagement and system of intelligence must be co-designed – the client teaches the back end, and the back end improves the client.
• The next era of enterprise scale will not be organized only around physical assets or labor hierarchies, but around intelligence as a governed corporate asset.

The alternative to data communism is what we call data capitalism. In this model, intelligence is not merely absorbed into a generalized frontier model and redistributed to everyone. Instead, each enterprise captures its own data, processes, policies and tacit knowledge, governs them according to its own corporate requirements and turns them into durable assets that models and agents can reason over.

This distinction is important. Frontier models are still fundamental. They provide generalized reasoning, language understanding, code generation and multimodal capabilities that will become increasingly powerful. But if that generalized intelligence is broadly available, it cannot be the basis of sustainable enterprise differentiation. The source of advantage moves to what is unique – i.e. the company’s customers, products, workflows, operating rules, institutional memory, regulatory constraints, decision rights and culture.

The stack on the slide above shows where that value concentrates. At the bottom, data platforms and systems of record remain essential. They tell the enterprise what happened. They store transactions, events, records, documents and snapshots of the business. But by themselves, they do not explain the business. They do not know why something happened, what is likely to happen next or what action should be taken.

That is the role of the system of intelligence. We believe this is the critical layer in the emerging enterprise AI software stack. It is a digital representation of the enterprise – a live model of the state of the business. It connects governed data with business meaning, metrics, policies, processes, relationships and tacit knowledge. In the industrial age, companies organized around physical assets such as railroads, warehouses, factories and assembly lines. In the AI age, we believe companies will increasingly organize around intelligence assets – the modeled representation of how the enterprise works and specifically its processes.

This is why the management analogy is so relevant in our view. The current debate often imagines a future of billion-dollar companies run by one person and an army of agents. That framework captures individual productivity, but it misses the larger organizational opportunity. Agents will not simply make individuals more productive. When grounded in a system of intelligence, agents will have expertise that can support planning, control, coordination, resource allocation and organizational alignment. Those are the core functions of management. The more complete the underlying intelligence layer, the more humans and agents can coordinate at scale.

In our view, this points to a future of larger and more complex forms of economic organization, not merely smaller companies with fewer people. The lesson from companies such as Amazon is that scale advantage increasingly comes from an operating platform that lets humans, software and data coordinate across many domains. Enterprise AGI extends that logic by adding agents that can reason and act through the enterprise model.

The top of the stack shows the system of agency. This is where agents use the system of intelligence to answer questions, analyze options, plan actions and operationalize decisions. The key is that agents should not act independently of enterprise context. They should act through the system of intelligence, where rules, constraints, metrics and trusted data provide the guardrails.

The left side of the stack shows the system of engagement – the new work surface for human and agent interaction. This is where users express intent, ask questions, resolve ambiguity, approve actions and interact with insights, decisions and data. The system of engagement is not just a front end. It becomes a learning surface. It captures the language, questions, corrections and decisions that help the system of intelligence understand how the enterprise actually works.

That co-design between the system of engagement and the system of intelligence is one of the most important insights in this architecture. The intelligent client and intelligent back end must reinforce each other. The engagement layer collects user intent and behavioral signals. The intelligence layer turns those signals into governed context. The agency layer then acts with increasing confidence because it is operating on a richer model of the enterprise.

There is also a key design tension worth noting. A purely top-down model of the enterprise can take too long to build and may become obsolete before it is complete. A purely bottom-up model can learn quickly from users, queries, workflows and behavioral signals, but it risks producing silos and inconsistency. The most promising architecture will combine both. It will infer what it can from the bottom up, govern what must be standardized from the top down and continuously reconcile the two.

That is why data capitalism is not simply about owning data. It is about owning the intelligence production system. Enterprises will create advantage by turning their proprietary data, processes and tacit knowledge into governed, reusable and compounding assets. The more those assets are used, refined and connected to decisions, the more valuable they become.

Key Takeaway:
Data capitalism treats enterprise knowledge as a proprietary asset. The system of intelligence becomes the new organizing layer of the firm – connecting data, process, policy and tacit knowledge so humans and agents can coordinate actions at enterprise scale.

Databricks Moves Up the Enterprise Intelligence Stack

Slide setup:
The slide below maps our enterprise intelligence software framework onto the Databricks stack. While Databricks make a laundry list of important announcements at its user conference (30,000+ attendees), the real story is Databricks moving up from data infrastructure into the higher-value layers of the emerging AI stack – the system of engagement, the system of intelligence and the system of agency. Genie is the agentic client layer, Genie Ontology is the emerging intelligent back end, and Agent Bricks plus Unity AI Gateway begin to operationalize and govern agentic work.

Key Points

• Databricks is no longer positioning only as data infrastructure – it is moving up into enterprise intelligence.
• Genie and the broader family of Genies represent the system of engagement – intelligent clients for different roles.
• Genie One is aimed at business users and anchors the experience in business data, dashboards, apps and governed assets.
• Other Genies – such as Code, ZeroOps, Agents, App Builder and Flow – are role-specific coworkers that understand the Databricks environment.
• Genie Ontology sits through the middle of the stack as Databricks’ emerging system of intelligence – a map of the enterprise’s data and business meaning.
• Agent Bricks, Omnigent, Unity Catalog and Unity AI Gateway form the basis of the system of agency and governance layer.
• The key question is whether and how these pieces reinforce each other in a closed loop – learning from users, applications, governed data and agent activity.
• The data layer remains essential, but increasingly looks like infrastructure. The new platform is the enterprise ontology or digital twin, and the new applications are agents.

Databricks’ Data and AI Summit unveiled many product announcements, as always. In our view, the more important story is architectural. Databricks is moving up the enterprise AI stack from data infrastructure into the layers where the next era of value capture is likely to occur – i.e. engagement, intelligence and agency.

This pattern is not unique to Databricks. We saw similar motion from Snowflake, however Databricks’ vision appears more complete. Modern data platforms understand that data infrastructure, while necessary, is not sufficient. The higher-value opportunity is to turn governed data into business context, then turn that context into agentic action. That requires a front-end experience where users express intent, an intelligent back end that understands enterprise meaning and a governance fabric that controls how agents act.

In the Databricks stack, the system of engagement is represented by Genie and the broader family of Genies. Genie One is the business-user client. It is not simply another dashboarding interface. It is a data-aware coworker designed to connect business users to the assets available across the enterprise – dashboards, apps, Genie Spaces, tables, governed metrics and other analytic assets. The important point is that the experience is anchored in business data. This means the user interface becomes a source of signals about how people ask questions, resolve ambiguity and consume insights.

Databricks is also extending the Genie concept to other roles. Code, ZeroOps, Agents, App Builder and Flow are examples of role-specific intelligent clients. These products are not bespoke copilots dropped into an enterprise environment. They are post-trained or tuned around the Databricks world, which means they understand Databricks assets, workflows and operational patterns. A coding assistant that understands how to build data pipelines in Databricks, or a ZeroOps assistant that understands how to manage Databricks infrastructure, can operate with less supervision and fewer manual adjustments than a more generic tool.

That role-specific design is a clue to a broader Enterprise AGI architecture. Just as Databricks can post-train agents on its own environment, an enterprise can build its own system of intelligence so its business agents understand the company’s environment. The enterprise needs a map – a representation of its data, entities, relationships, policies, metrics and workflows. That is where Genie Ontology becomes strategically important.

Genie Ontology runs through the middle of the Databricks stack because it is the emerging system of intelligence layer. In our framework, the system of intelligence is the map of the enterprise. It is what allows the agentic clients to navigate business context and what allows agents to act with confidence. In its current form, Genie Ontology is an early step toward a digital twin – not yet a complete operational model of the enterprise, but a deliberate move toward a governed representation of enterprise data and business meaning.

Above the ontology sits the system of agency. Databricks has several pieces here. Agent Bricks is evolving into an agent development platform for building, deploying, optimizing and governing agents. Omnigent is an open-source harness that can connect agentic clients and coding assistants into the Databricks governance layer. Unity Catalog and Unity AI Gateway provide the policy, access control, routing, model governance, tracing and cost management capabilities that enterprises will need as agentic workloads expand.

This is where the stack begins to come together. The system of intelligence provides the map. The Genies use that map to help users interact with enterprise data and business context. Agent Bricks and related agency tooling allow agents to operationalize decisions. Unity Catalog and Unity AI Gateway govern what is allowed. The more these layers reinforce one another, the more Databricks can move from data platform to enterprise intelligence platform.

Key Questions Remain

There are still open issues that require more research. The first is how deeply the system of engagement and the system of intelligence are co-designed. The intelligent client must not merely query the ontology. It should help teach it. User questions, corrections, clarifications, accepted answers and rejected answers can become signals that improve the enterprise map. That learning loop is much stronger when Databricks controls the client experience. It becomes more complicated when third-party clients – such as coding assistants, enterprise copilots or other agentic work surfaces – sit in front of the Databricks back end.

The second question is how far Genie Ontology can evolve. Today, it appears strongest as a semantic and contextual layer over governed data. That is valuable, but Enterprise AGI requires more than semantic harmonization. It requires a governed, executable model of how the business operates – including actions, preconditions, effects, policies, workflows and live state. Databricks has important ingredients, but the full system of intelligence is still emerging.

The third question is where the value line gets drawn. Databricks and others are trying to make data formats and data infrastructure less visible to customers. That is encouraging progress at witnessed by Ryan Blue’s short on-stage banter with Ali Ghodsi during the day 1 keynote. Customers should not have to care whether the underlying format is Delta, Iceberg or something else. The infrastructure should just work. But that also means infrastructure becomes more like hardware – necessary, powerful and expensive, but increasingly standardized. The differentiating platform is the ontology or digital twin. The applications on top are agents.

This is the strategic leap forward Databricks is attempting. It is trying to turn its governed data foundation into an enterprise intelligence platform, then use agents as the application layer above it. If successful, Databricks will not simply help enterprises store, govern and analyze data. It will help them model how the business works and let humans and agents act through that model.

Key Takeaway:
Databricks is trying to move from data platform to enterprise intelligence platform. The critical test is whether Genie, Genie Ontology, Agent Bricks and Unity governance become a reinforcing system that learns how each enterprise operates and turns that knowledge into agentic action.

Genie One – The Agentic Client for Business Users

This next slide zooms in on the system of engagement. Genie One is Databricks’ business-user agentic client – a data-smart AI coworker that brings dashboards, Genie Spaces, apps and governed data into a common experience. It may look like a new analytic workspace, but we believe the intent is deeper. Specifically, this is where business users express intent, ask questions, clarify ambiguity and create the signals that can make the intelligent back end smarter.

Key Points

• Genie One is Databricks’ front door for business users – a data-aware AI coworker rather than just another BI interface.
• The experience brings together dashboards, Genie Spaces, Databricks Apps, governed data assets, metrics, semantic objects, notebooks and queries.
• The initial focus is native Databricks assets, but the ambition extends to external files, documents, unstructured content and SaaS data.
• The key architectural advantage is grounding the experience in trusted, verified “gold” data assets.
• Ground truth gives the ontology something authoritative to connect other assets to.
• Connectors are necessary but insufficient – the hard work is mapping meaning across systems.
• The strategic reason to own the user experience is that user questions, corrections and clarifications can teach the system of intelligence.

Genie One is best understood as Databricks’ business-user agentic client. It is a data-smart AI coworker designed to give business users a single place to work with data, analytics, applications and AI. On the surface, it looks like a more modern way to interact with BI and analytic assets. In our view, its importance is much larger because it becomes the enterprise work surface where intent, context and feedback are captured.

A year ago, Genie was primarily a way to access BI assets such as dashboards. The scope is now expanding. Genie One brings together AI/BI dashboards and visualizations, Genie Spaces, Databricks Apps, governed tables, views and datasets, metrics and semantic objects, notebooks, queries and other workspace assets. In other words, it is moving from a dashboard front end toward an intelligent business workspace.

That implies the system of engagement is not only a user interface. It is the place where users ask questions in business language, resolve ambiguous definitions, choose among possible answers and validate whether the response is useful. Those interactions create behavioral signals. Over time, those signals can help build and refine the system of intelligence.

The next phase is connecting Genie One to a broader universe of enterprise assets. That includes local files uploaded into Genie Spaces – such as Excel or CSV files – and blending those with Unity Catalog governed data. It also includes documents and unstructured content from systems such as SharePoint, Google Drive, Confluence and Glean. In addition, Databricks has highlighted more than 80 connectors into SaaS and enterprise applications, including common systems across CRM, support, productivity and file storage.

But connectors are only the first step. The hard part is mapping meaning. It is one thing to connect to Salesforce or SAP. It is another thing for Genie One to know that “customer” in one system maps to “account” in another system, that a customer should be treated differently under certain credit conditions, or that a revenue metric must be calculated consistently across multiple operational systems.

This is where Databricks has a potentially important advantage – i.e. in the ground truth. The company starts from governed, verified data assets – what one would call “gold” data. A certified dashboard, governed metric or trusted table provides an authoritative stake in the ground. That gives the ontology something to connect other assets to. Without this anchor, the enterprise graph becomes much harder to build because there is no agreement on what the trusted reference point should be.

The strategic significance of Genie One, therefore goes beyond giving business users a better front end. More importantly, in our view, Databricks is trying to create a learning surface for the enterprise. Every query, accepted answer, rejected answer, clarification and correction can become input into the intelligent back end. The system of engagement teaches the system of intelligence, and the system of intelligence makes the engagement layer more useful.

That feedback loop is central to Enterprise AGI. The enterprise does not simply need a chatbot over dashboards. It needs an intelligent client that can learn the company’s language, assets, definitions and workflows. Genie One is an early expression of that architecture. Its success will depend on how well Databricks can expand from native analytic assets into the messier world of files, documents, SaaS applications and cross-system business meaning.

Key Takeaway:
Genie One is more than a business-user interface. It is Databricks’ attempt to own the engagement layer where enterprise intent is captured and fed back into the system of intelligence. Its long-term value depends on turning user interaction into governed enterprise context.

Databricks’ Many Genies – Role-Specific Coworkers Built on a Shared Enterprise Map

This next graphic below expands the Genie concept beyond the business-user surface. Genie One is the front door for business users, but Databricks’ broader ambition is to create data-smart AI coworkers for many roles – developers, data engineers, data scientists, analysts, app builders and agent creators. The slide’s message is “Everyone. Everywhere. Everything.” The strategic point is that these role-specific experiences can draw from, and contribute back to, the same governed enterprise context through Unity Catalog and Genie Ontology.

Key Points

• The market has no shortage of agents, copilots and agent development tools.
• Differentiation comes from integrating those agents with a company-specific data and action space.
• Genie Ontology and Unity Catalog form the emerging enterprise map and governance layer.
• Role-specific Genies can understand the Databricks environment and operate with less supervision.
• Genie Code can help build Databricks applications and data pipelines.
• ZeroOps can understand the Databricks environment, diagnose issues and suggest or perform remediation.
• Data science and analyst experiences can operate from governed metrics and predefined semantic objects.
• Databricks’ ambition has three vectors: give Genie to everyone, connect it to everything and let users access it everywhere.
• The more pervasive the engagement surface, the stronger the learning loop between users and the intelligent back end.
• The risk is that if third-party clients own too much of the engagement layer, Databricks may lose some of the interaction signals needed to improve the ontology.

Genie One is the business-user surface, but Databricks’ ambition is much broader in our assessment. The company is extending Genie across roles and workflows, creating data-smart AI coworkers for developers, data engineers, data scientists, analysts, app builders, agent builders and business users. In our view, this is the right architectural direction because Enterprise AGI will not be delivered through a single generic chatbot. It will require role-specific experiences that understand the tools, data, permissions, workflows and business language of the enterprise.

The market is already crowded with copilots and agent tools. There are plenty of agents, and there will be many more. The real question is what differentiates them. We believe the answer is not the interface alone, and not the model alone. The differentiator is whether the agent is integrated with the company’s specific data and action space.

This is where Genie Ontology and Unity Catalog become important. Together, they begin to create a company-specific environment – a map of the enterprise and a governance layer around it. If that map is rich enough, role-specific agents can understand how the company works, which data assets are trusted, which metrics are most important, which permissions apply and which actions are allowed.

This is why the role-based approach is so relevant. Genie Code can help developers build Databricks applications or data pipelines because it understands the Databricks environment. ZeroOps can help diagnose and remediate operational issues because it understands the infrastructure and pipeline context. Data science and analyst experiences can use governed metrics and semantic objects, making it easier to build models and analyses from trusted enterprise definitions. Business users interact through Genie One, but the broader pattern is the same in that each role gets a specialized coworker that is grounded in the same intelligent back end.

The slide above frames the strategy around three vectors.

• First, Databricks wants to give Genie to everyone – business users, developers, analysts, data engineers and agent creators. \
• Second, it wants to connect Genie to everything – the lakehouse, governed metrics, federated applications, queries, search, tools, actions and unstructured data through mechanisms such as MCP.
• Third, it wants users to access Genie everywhere – desktop, mobile, Slack, Teams, AI productivity tools, agentic coding tools and eventually through the user’s own agents.

That “everywhere” ambition is notable. The system of engagement is more than a consumption layer. It is also a learning surface. Every user question, correction, clarification, accepted answer, rejected answer and workflow interaction becomes potential signal for the system of intelligence. The more surface area Databricks controls or observes, the more opportunity it has to improve the ontology. The better the ontology becomes, the more useful the role-specific Genies become. That is the virtuous cycle.

This also explains why the system of engagement is becoming one of the most contested layers in enterprise AI. Microsoft, Snowflake, Google, Amazon, the frontier model vendors and the SaaS players all want their own agentic client to become ubiquitous. Whoever owns the engagement layer has privileged access to the behavioral exhaust that teaches the intelligent back end how the enterprise actually operates.

Databricks’ advantage is that its engagement layer is anchored in governed data. Its challenge is distribution. Microsoft has the productivity surface. SaaS vendors have application workflows. Frontier model vendors have the broadest AI mindshare. Databricks has to make Genie pervasive enough that it can capture the interaction signals needed to strengthen Genie Ontology, while still accommodating third-party clients that enterprises will inevitably use.

The “Many Genies” strategy therefore can’t be about product sprawl. It must be about turning role-specific engagement into a shared learning system. If the Genies remain connected to the same governed enterprise map, they can reinforce one another. If they fragment into disconnected copilots, the architecture loses its flywheel advantage.

Key Takeaway:
Databricks’ Many Genies strategy differentiates from other copilots. Its focus is creating role-specific engagement surfaces that share one governed enterprise map – and continually teach the intelligent back end how the business works.

Genie Ontology – The Core of the Intelligent Back End

This next slide moves to the heart of the intelligent back end…Genie Ontology. This is where Databricks starts turning governed data into shared enterprise meaning by connecting business terms, metrics, authoritative data sources, semantic definitions and expertise. The current capability is best understood as semantic harmonization – helping organizations agree on what their data actually means. That is a major step toward a system of intelligence, but it is not yet a complete digital model of how the enterprise operates.

Key Points

• Genie Ontology is Databricks’ emerging system of intelligence layer.
• Its purpose is to create a map of the enterprise’s data and business meaning.
• The ontology extracts and organizes “snippets” of knowledge from tables, queries, dashboards, documents, metric views, pipelines and connected apps.
• Examples include metric definitions, authoritative data sources, entity relationships and rule-like business semantics.
• Its current strength is semantic harmonization – defining what terms, metrics and entities mean across the enterprise.
• This bottom-up learning approach is a meaningful shift from older semantic layers that were largely authored top-down.
• The major advantage is ground truth – tying inferred knowledge back to certified, governed, “gold” data assets.
• The hard part becomes encoding business process logic, policy, preconditions, effects and operational rules.
• Whoever owns the business definitions may be able to generate the dashboard, the app and the agent experience.
• That is why ontology ownership could become a major enterprise software control point.

Genie Ontology is the most important piece of Databricks’ enterprise intelligence strategy in our view because it begins to answer the central question in Enterprise AGI – i.e. how does the system know what the business actually means?

Metadata and data in rows and columns is not enough. A table can store revenue, customers, opportunities, orders or invoices, but it does not automatically know which revenue number is authoritative, how “active customer” should be defined, how a metric should be calculated, which source is trusted or which business process rule governs a decision. Enterprise AGI requires that meaning to be captured, governed and made available to agents.

This is what Genie Ontology is designed to do. It builds a map of the enterprise’s data and business context by extracting knowledge from assets such as tables, queries, dashboards, pipelines, metric views, documents and connected applications. Those knowledge snippets can include metric definitions, authoritative-source pointers, entity relationships, synonyms, business terms, SQL expressions, join relationships, formatting rules and domain-specific instructions. The ontology then ranks and uses those snippets based on signals such as provenance, authority, usage frequency and freshness, while enforcing Unity Catalog permissions.

In practical terms, Genie Ontology can encode statements such as the following examples: revenue should come from a specific certified finance table; an active user is a distinct user deduplicated across platforms; a particular metric like NRR is calculated using an approved formula; or a certain dashboard is the single authoritative source of truth. These examples are non-trivial. They are the semantic plumbing that determines whether an AI system produces trusted answers or merely plausible ones.

Our assessment is that Databricks’ current strength is semantic harmonization. It is helping enterprises agree on what the data means across different users, dashboards, metrics and domains. That is a big step forward because most enterprises still struggle with inconsistent definitions, duplicated metrics, disconnected reports and competing versions of the truth.

The architectural nuance is how much of this can be learned or inferred from the bottom up. Historically, enterprise semantic layers and ontologies were mostly authored top-down. Experts defined the model, governed the terms and manually curated the meaning. That approach can produce consistency, but it is slow, brittle and often fails to keep up with how the enterprise changes. Databricks and Snowflake are showing that at least part of the enterprise map can be learned from real usage patterns – queries, dashboards, metric definitions, documents, user behavior and accepted answers.

That does not mean bottom-up learning solves the entire problem. It works well for many entities, relationships, measures, definitions and analytic results. It is much harder when the system must understand business process logic. For example, “qualified lead” might only count after a demo is booked. Credit extension may depend on customer tier, payment history, region, product type and risk policy. Revenue recognition may depend on contract terms, delivery milestones and local regulation. These are not merely definitions – they are operating rules. Some can be inferred and even hard-coded. But as things change these must be explicitly taught, approved and governed across the enterprise.

This is where the distinction between a descriptive ontology and an executable ontology becomes important. Genie Ontology today appears strongest as a descriptive or read-context layer over governed data. It helps agents answer accurately by grounding them in enterprise semantics. But a full system of intelligence must eventually go further. It must represent typed objects, actions, preconditions, effects, workflows, live state and policy constraints. That is what allows agents not only to answer questions, but to act with confidence.

The strategic implication is that whoever owns the ontology owns the business definitions. And in the AI era, owning the definitions will matter more than owning the dashboard in our opinion. If the system knows the authoritative metrics, entities and relationships, then it can generate the visualization, the narrative, the app or the agent workflow dynamically. The dashboard becomes an output, not the control point.

This explains why the semantic layer is becoming a competitive battleground. Microsoft’s reported blocking of Databricks Unity metrics from being consumed in Power BI underscores the stakes. The stated rationale may be consistency, but the deeper issue is control. If Power BI owns the definitions, Microsoft can push those definitions into Fabric and make them authoritative. If Databricks owns the definitions through Unity Catalog and Genie Ontology, then Power BI becomes less central because the visualization can be generated from Databricks’ governed semantic layer.

In our view, this is the next major platform fight. The system of record captured transactions. The BI layer captured reporting. The data platform captured storage, governance and analytics. The system of intelligence aims to capture enterprise meaning itself. That is a higher-value control point because it becomes the foundation for agents, applications and business action.

Genie Ontology is not yet a complete enterprise digital twin. It does not yet fully model how the business operates in real time. But it is an important step toward the intelligent back end, and it gives Databricks a credible path from governed data platform to enterprise intelligence platform.

Key Takeaway:
Genie Ontology turns governed data into business meaning. Its current strength is semantic harmonization, but the strategic prize is larger. It’s owning the enterprise definitions, rules and context that agents will need to reason, decide and act.

How Genie Ontology Learns – The Living Context Graph

This slide explains how Genie Ontology is built and kept current. Rather than relying only on a traditional top-down semantic-modeling exercise, Databricks is trying to infer enterprise meaning from the bottom up – from KPIs, SQL queries, business vocabulary, user interactions, corrections, accepted answers, Unity Catalog schemas, Lakeflow pipelines and unstructured workplace content. The result is a living context graph that ranks knowledge by provenance, authority, frequency and freshness before using it to ground agent responses.

Key Points

• Traditional semantic layers (e.g. AtScale, GoodData) are mostly authored top down; Genie Ontology is designed to learn much more from the bottom up.
• The system ingests “human signals” such as KPIs, SQL queries, business vocabulary and user-defined context.
• It also harvests system signals from Unity Catalog schemas, Lakeflow pipelines and unstructured workplace content.
• User interactions become behavioral exhaust – questions, accepted answers, corrections and clarifications can all improve the ontology.
• Databricks uses a mechanism they call OntoRank (not sure why they didn’t just call it “OntologyRank”) to weigh knowledge based on provenance, authority, usage frequency, freshness and links to certified assets.
• The key advantage is grounding – tying inferred context back to trusted, governed, “gold” data assets.
• This approach can reduce hand curation and help solve the enterprise context problem.
• Bottom-up learning works best for entities, relationships, measures, vocabulary and BI-style semantics.
• It is much harder for business process rules – what must happen, what must not happen and why.
• The system of engagement is not just a client – it is a teaching surface for the system of intelligence.

The critical question for any enterprise ontology is not just what it knows today. It is how it learns, how it stays current and how it resolves conflicts when the enterprise changes. Genie Ontology is important because Databricks is not trying to build this primarily as a static, top-down semantic layer. The design center is a bottom-up learning loop.

That is a major shift. Historically, enterprise semantic layers were authored by experts. Business analysts, data teams and governance committees defined metrics, dimensions, hierarchies, joins and business terms. That model can produce consistency, but it is slow and difficult to keep current. By the time a top-down enterprise model is complete, the business has often moved on.

Genie Ontology takes a different path. It learns from a combination of human artifacts and system behavior. The slide above shows three major input streams. The first is the Genie Space workshop, where users define or expose KPIs, SQL queries and business vocabulary. The second is the live interaction loop, where successful chat responses, user feedback and clarifications can be captured as permanent knowledge. The third is the automated ecosystem harvesters, which scan Unity Catalog schemas, Lakeflow pipelines and unstructured workplace content such as Slack messages and documents.

The purpose is to infer the meaning of enterprise objects from how people and systems actually use them. The ontology looks at snippets from tables, queries, dashboards, pipelines, metric views, documents and connected apps. It then ranks those snippets using a more sophisticated Google PageRank-style filter that weighs provenance, authority, usage frequency, freshness and connections to certified assets. The goal is to determine which definitions, relationships and sources are most likely to represent ground truth.

Ground truth is the critical advantage. Many enterprise knowledge graphs fail because they can collect information but cannot determine what is authoritative. Databricks has a stronger starting point because it can anchor inferred knowledge to governed assets in Unity Catalog – certified dashboards, approved metric views, trusted tables and permissioned data. That gives the ontology a reference layer against which other signals can be evaluated.

This is why the engagement is so critical. Genie is not merely a user interface. It is part of the learning system. Users express intent in natural language. They ask questions using company-specific vocabulary. They accept or reject answers. They correct ambiguous definitions. They clarify terms that the system does not understand. Over time, those interactions become behavioral exhaust that can teach the intelligent back end.

The power of this architecture is that it can solve part of the context problem without massive hand curation. It can learn that certain people are experts in certain domains. It can infer that two fields are often joined together. It can learn that a certain metric is the preferred measure for a given business question. It can discover vocabulary, usage patterns, authoritative dashboards and common analytic paths. For BI, analytics, entity resolution and semantic harmonization, this bottom-up approach can go surprisingly far.

But it also has a ceiling. Behavioral signals show what people do. They do not always reveal what the business requires. Query logs can expose types, grain, joins and vocabulary. They can show that users often calculate a metric in a certain way. But they cannot reliably enforce policy, obligation, prohibition or intent. They cannot always tell the system what must happen, what must not happen, which rule has priority, or why a process exists.

Enterprise AGI requires more than a descriptive map of data meaning. It needs an operational model of the business. It must understand not only that “platinum customer” means accounts above a certain revenue threshold, but also what actions are allowed for those customers, which approvals are required, what policies constrain the action and what downstream effects the action creates.

That is why this slide above is both encouraging but also revealing in its limits. The living context graph is a meaningful step toward a system of intelligence because it lets the enterprise learn from usage rather than depending entirely on top-down modeling. But to reach higher levels of ontology maturity, the learning mechanism has to evolve. Bottom-up inference must be combined with explicit teaching, governance and promotion of business logic into shared enterprise assets.

Key Takeaway:
Genie Ontology’s bottom-up learning loop is a breakthrough for semantic harmonization. But behavior can only show what happened – it cannot fully define what is required. To become Enterprise AGI, the ontology must combine inferred context with governed business rules and explicitly taught operating logic.

The Clarification Loop – How the Client Teaches the Ontology

This next slide shows the learning loop in a simple form. A business user asks a question using company-specific language: “Show me lost revenue from platinum customer segment.” Rather than guessing, Genie recognizes that “platinum customer segment” is ambiguous and asks for clarification. The user supplies a definition – platinum customers are accounts that generated more than $10K in revenue in any given month – and that clarification can become reusable enterprise context for future users and agents.

Key Points

• The example may look simple, but it captures why the system of engagement is strategically vital.
• Genie does not merely answer questions – it can ask clarifying questions when the ontology lacks confidence.
• User clarifications become behavioral exhaust that can teach the system of intelligence.
• Company-specific language is a major source of enterprise context – terms like “platinum customer” often have local meaning.
• Owning the business-user client is key because that is where ambiguity is surfaced, clarified and captured.
• The clarification loop works well for business definitions, metrics, semantic intent and common enterprise vocabulary.
• The loop becomes harder when the clarification implies policy, permissions, process rules or cross-functional governance.
• Every major vendor wants this engagement layer because it becomes the new work surface and the training surface for the intelligent back end.
• Without the client, a vendor may own data or models but miss the interaction signals needed to improve the ontology.

The platinum customer example looks almost trivial, but it reveals one of the most important design principles in Enterprise AGI in that the intelligent client has to teach the intelligent back end.

A user asks Genie to show lost revenue from the platinum customer segment. The phrase “platinum customer segment” is not self-evident. It could mean top-tier accounts by annual revenue, customers above a lifetime value threshold, accounts in a loyalty program, customers with premium support, or any number of company-specific definitions. A generic model might guess. A governed enterprise system should not.

In this example, Genie recognizes the ambiguity and asks for clarification – i.e. how should “platinum customer segment” be defined? The user responds that platinum customers are accounts that generated more than $10K in revenue in any given month. That answer is more than a static one-time instruction. It can become enterprise context – a reusable snippet of meaning that helps future users and agents interpret the same term consistently.

This is why the system of engagement is so crucial in our model. The front-end client is not merely where users consume answers. It is where the enterprise expresses intent. It is where ambiguous terms are surfaced. It is where competing definitions are reconciled. It is where users confirm, reject, correct and refine what the system believes. In other words, the client is a teaching instrument.

That teaching function depends on tight co-design between the system of engagement and the system of intelligence. The ontology must be able to identify a gap in its knowledge, route a clarification question to the user, capture the response, attach provenance and authority, and then determine whether the new definition should remain local, be routed for approval or be promoted into broader enterprise context.

This is also why owning the business-user experience is strategically valuable. If the vendor owns the client, it can capture the natural-language query, the ambiguity, the clarification and the subsequent usage pattern. If another vendor owns the client, those signals may be incomplete or unavailable. The back end may still answer questions, but it loses some of the behavioral exhaust needed to improve the ontology.

This also explains why so many vendors are fighting for the new agentic work surface. Snowflake needs a business-user client because its intelligent back end improves when users teach it. Microsoft is investing heavily in Microsoft 365 Copilot because the productivity surface is where much enterprise intent is expressed. AWS, Google, OpenAI, Anthropic, Salesforce and others all have versions of the same ambition. The prize is not just user interface real estate. The prize is the learning loop.

There is a deeper platform implication. As the agentic client becomes the center of gravity, traditional applications risk becoming tools that are invoked by the new surface rather than destinations users visit directly. Office documents, dashboards, BI reports, CRM screens and workflow apps may increasingly become editable artifacts or tool endpoints inside a broader agentic experience. The vendor that owns the client can shape how those tools are invoked – and can capture the feedback that trains the intelligent back end.

The clarification loop works especially well for company-specific English, metric definitions, entity labels and semantic ambiguity. It can learn that “platinum customer” has a local definition, that a particular dashboard is authoritative, or that a certain metric should be calculated a specific way. But it also has limits. When a clarification crosses into policy, permissions, compliance or operational rules, the system needs governance. A single user’s answer cannot automatically become enterprise truth in every context.

That is the boundary between learning and governance. The system should learn from users, but it must also know when to ask, when to route for approval, when to keep context local and when to promote knowledge into the enterprise ontology. This is how bottom-up learning begins to meet top-down control.

Key Takeaway:
The agentic client is not just the front end for Enterprise AGI. It is the teaching surface. Whoever owns the clarification loop can capture the company-specific language, definitions and intent that make the system of intelligence smarter over time.

Ontology Maturity – From Semantic Context to Agent Coordination

This next graphic places Genie Ontology on our maturity curve. Our assessment is that Databricks currently sits around levels 5 to 6 – the transition from diagnostic intelligence toward agent coordination. The core principle is that the richer and more faithful the enterprise model becomes, the more sophisticated the analytics can be, and the greater the scope and confidence of agentic action.

Key Points

• Databricks Genie Ontology is meaningful, but it is not yet a fully executable enterprise ontology.
• We place it roughly between level 5 and level 6 on the maturity model.
• Levels 1 to 5 are primarily diagnostic – they improve reporting, correlation, behavioral analysis and prediction.
• Level 6 begins the move toward agent coordination by connecting people, resources, entities and relationships in an enterprise knowledge graph.
• Level 7 requires a semantic action layer – actions become modeled data, with preconditions, effects and guardrails.
• Level 8 is the real-time digital twin – the live state of the business becomes the shared source of truth.
• Level 9 is an autonomous operations platform – workflows themselves become editable data and humans set goals.
• The jump from level 5-6 to level 7 and above cannot be achieved by bottom-up inference alone.
• To advance, the ontology must incorporate more explicit teaching, business-process modeling, governance and human-in-the-loop validation.
• Forward-deployed AI engineering becomes part of the bridge from descriptive ontology to executable operations.

The ontology maturity model helps clarify where Databricks is today and what must come next. In our assessment, Genie Ontology sits roughly between levels 5 and 6. That is a strong position relative to where most enterprise data environments are, but it is not yet the upper end of Enterprise AGI.

The governing principle of the model is that as the enterprise representation becomes richer, the analytics become more sophisticated, and agents can act with broader scope and greater confidence. At the bottom of the model, siloed reporting can answer only narrow questions from individual systems. The enterprise may have multiple customer records across CRM, credit, KYC and support systems, but no unified representation of the customer. The output is a report that a human reads and acts upon.

As maturity increases, the enterprise moves from isolated reports to data warehouses, event hubs, behavioral analytics and predictive analytics. By level 5, the system can begin to more accurately answer questions such as what is this customer’s churn risk, and why? The system can inform a recommended action – for example, shifting sales coverage to reduce churn or increase expected revenue – but a human still decides. This is where we believe Databricks Genie Ontology is increasingly relevant.

Level 6 is the enterprise knowledge graph. At this level, the system begins to connect people, resources, accounts, products, transactions and relationships in a more structured way. Instead of simply calculating a churn score, the system can reason over paths. For example, which person can reach this investor, which account is tied to which relationship, which business unit owns which resource and which governed rule constrains a decision. Agents can act on intent, but only within narrow, governed lanes.

This is why we describe Databricks as being around levels 5 to 6. Genie Ontology is more than a BI semantic layer because it learns business terms, metrics, entities, authoritative sources and relationships from governed data and usage signals. But it does not yet fully meet the stricter definition of level 6 if typed relationships are expected to exist as first-class declarative objects across the enterprise. Nor does it yet implement the higher levels of the model as the ontology itself.

The higher levels are where the architecture becomes truly operational. Level 7 is the semantic action layer. Here, actions such as “submit credit memo,” “approve discount,” “change shipment priority” or “reassign sales coverage” become modeled data. The system understands the preconditions, effects, permissions and guardrails around each action. Agents no longer merely recommend what could be done – they can choose and run actions within governed boundaries.

Level 8 is the real-time digital twin. In that model, the ontology is not simply querying operational systems for the live state of the business. The digital twin becomes the shared source of truth for the operating state of the enterprise. Analyzing becomes equivalent to operating because the system can answer what is true right now and agents can coordinate off a shared live state rather than passing messages through disconnected applications.

Level 9 is the autonomous operations platform. At this point, the workflow itself becomes editable data. Agents can plan, optimize and adjust operations, while humans set objectives, constraints and goals. This is the fully operational version of Enterprise AGI – not a model that knows generic business concepts, but a system that understands and can help run the specific enterprise.

The key point is that moving from level 5-6 to the higher levels requires a different learning mechanism. Bottom-up inference can take the enterprise a long way. It can learn types, texture, joins, vocabulary, common metrics, usage patterns and even some rule-like semantics. But behavior shows what happened. It does not reliably specify what is required.

The nuance is a system may infer that users often calculate churn in a certain way, but it cannot know from behavior alone which remediation actions are allowed, which approvals are mandatory, which policy takes precedence, or why a process exists. Those higher-order rules must be taught, authored, governed and validated. This is where human-in-the-loop design becomes more important, not less.

We expect forward-deployed AI engineering to become part of this bridge and is something Ali Ghodsi referenced in his day 1 keynote (evidently Databricks has this capability now). The shift from descriptive ontology to executable ontology requires humans who can help customers capture operating logic, clarify business rules, connect processes to governed data and promote local knowledge into enterprise assets. Automation will reduce the amount of manual modeling required, but the upper levels of the maturity model require explicit business-process meaning.

The implication is that Databricks has important ingredients, but the path upward is not automatic. To own the SoI, Genie Ontology must evolve from a semantic knowledge layer into a governed operational model. It must move from understanding metrics and entities to understanding actions, policies, state, workflows and decision rights. Adjacent products such as Agent Bricks, MCP tools, Lakebase and Unity AI Gateway may help, but the key question is whether these capabilities converge into a unified system of intelligence rather than remaining separate product islands.

Key Takeaway:
Databricks is at a meaningful point on the maturity curve – around levels 5 to 6 – but the next jump is harder. To reach the higher levels of Enterprise AGI, Genie Ontology must evolve from a descriptive semantic layer into a governed, executable model of business actions, policies, workflows and live state.

Constructing a Governed, Executable Ontology

This slide shows the bridge from a descriptive ontology to an executable one. Bottom-up learning can reveal structure, enterprise vocabulary and recurring work patterns. But behavior alone cannot tell the system what must happen, what must not happen, which policy has priority or why a rule exists. To move beyond levels 5 and 6, the ontology has to combine bottom-up skill harvesting with top-down governance.

Key Points

• Bottom-up inference works well for structure – business objects, relationships, grain, shared naming and commonly used metrics.
• It is much harder to infer operating logic – each process step, its conditions, its outcomes and how teams actually execute work.
• It is harder still to infer authority – company policy, regulations, contracts, required actions, deadlines and prohibitions.
• The next step is to harvest user-authored agent skills from local work and convert them into reusable enterprise assets.
• Those skills must be abstracted, checked for conflicts, routed by risk, governed and promoted into the ontology.
• A purely bottom-up model risks becoming a Tower of Babel – lots of local automation with no shared enterprise logic.
• A purely top-down model risks being too slow and brittle.
• The winning architecture meets in the middle – bottom-up contribution, top-down approval and continuous promotion or demotion of shared logic.
• Databricks will need this hybrid model if Genie Ontology is to evolve from semantic harmonization into an executable system of intelligence.
• This transition will take time because it requires humans in the loop, governance committees and explicit teaching of business-process meaning.

The next step in ontology maturity is harder than semantic harmonization. Genie Ontology can learn a great deal from usage patterns, queries, dashboards, metric definitions and end-user clarifications. That bottom-up approach can expose structure, shared vocabulary and recurring patterns of work. But it cannot fully capture what the enterprise requires.

The graphic above frames the issue as three layers.

Layer 1 is structure. This is the foundation. It includes the key business objects – customers, orders, invoices, products, accounts, opportunities, suppliers and assets. It also includes how those objects connect, the level of detail, the grain of the data and shared naming conventions. This is where bottom-up learning is strongest. Query logs, dashboards, schemas, joins and user vocabulary can reveal a lot about how the enterprise describes itself.

Layer 2 is operating logic. This is where the problem becomes more difficult. Operating logic includes metric calculations, each step’s conditions and outcomes, team workflows and how a process actually runs today. This is not just “what does platinum customer mean?” It is “what happens after a platinum customer misses a payment?” or “under what conditions should a sales coverage change be recommended?” or “which workflow should be triggered when risk exceeds a threshold?”

Layer 3 is authority and cross-cutting norms. This is the highest and hardest layer. It includes company-wide policy, regulatory obligations, contractual constraints, required actions, deadlines, prohibitions and the rationale behind each rule. These rules cannot be learned safely from behavior alone because behavior may be incomplete, inconsistent or wrong. The system needs explicit governance to determine what is allowed, what is required and what must be prevented.

This is where skill harvesting becomes important. The key finding is that agents are emerging first through personal productivity. Users will author local skills, prompts, workflows and automations to help them do their jobs. Those skills contain valuable tacit knowledge about how work actually gets done. But if every user builds skills independently, the enterprise gets fragmentation. It gets lots of useful local automation, but no shared system of intelligence.

The approach should not to suppress local authoring. The idea is to create a promotion pipeline. The slide above lays out that pipeline: author locally, abstract, route by risk, govern and promote. A user or team creates a useful skill. An LLM or tooling layer abstracts it, identifies duplicates and conflicts, and makes it legible. The system routes it based on risk. A governance process reviews it. If approved, the skill is promoted into shared operating logic. If it proves wrong, stale or too narrow, it can be demoted or re-fragmented.

That is how bottom-up contribution can become enterprise logic. The skill starts as local knowledge. It becomes readable. It is compared with other skills and policies. It is governed. Then it becomes a reusable corporate asset that agents can rely on.

This is the critical transition from behavioral signal to executable ontology. Behavioral exhaust can show what people do. User-authored skills can reveal how people think the work should be done. Governance determines whether that logic should become enterprise truth.

We believe this is where many Enterprise AGI strategies will either mature or stall. A purely bottom-up system becomes a Tower of Babel. A purely top-down system takes too long and cannot keep up with the business. The winning architecture meets in the middle – local innovation at the edge, abstraction and harmonization in the middle, and policy, regulation and governance from the top.

For Databricks, this is the next major challenge we expect them to tackle. Genie Ontology has a credible path as a bottom-up, inferred semantic layer. But to move into levels 7, 8 and 9, it must support more explicit teaching of actions, rules, policies, live state and business-process logic. Adjacent capabilities in agent development, governance, Unity Catalog, Unity AI Gateway and partner ecosystems can help. But the architecture has to converge into a governed, executable ontology rather than remain a collection of useful but separate features.

This also explains why the timeline important. Enterprises will not fully arrive at this model overnight. Capturing tacit knowledge, abstracting local skills, resolving conflicts, routing risk and governing shared logic is hard organizational work. Models will become more capable, and they will automate more of the translation and abstraction. But our research indicates that the higher levels of Enterprise AGI will still require humans in the loop because the system must learn not only what happens, but what should happen.

There are two caveats to our scenario: 1) Competition, inertia and platform affinity will likely create silos of intelligence, injecting friction into the new enterprise operating model; and  2) some in the community believe that this approach is entirely too complex for enterprises to adopt and that systems will emerge – perhaps from research labs or other startups – that simplify the adoption of these complex capabilities. This approach could occur perhaps via partnerships, led by the leading LLM players which could include an integration layer across the impending silos.

Key Takeaway:
The path from semantic ontology to executable Enterprise AGI requires a hybrid architecture. Specifically, a bottom-up skill harvesting, top-down governance and a promotion pipeline that turns local know-how into shared business logic. Without that middle layer, agents may become useful personal tools, but they will not become a governed operating system for the enterprise.

Omnigent – Governing a Heterogeneous Agent Estate

The next graphic below shifts from Databricks’ native Genie experience to the broader reality of enterprise AI. Specifically, companies will not have one agentic client. They will have many. Omnigent is Databricks’ open-source harness for connecting third-party systems of engagement – coding assistants, enterprise copilots and external agents – into Unity AI Gateway, where the enterprise can apply common governance across models, agents, MCP tools, skills and telemetry. The value proposition is openness with control. The open question is whether third-party clients can contribute the same rich semantic feedback that native Genie clients use to improve Genie Ontology.

Key Points

• Omnigent is designed to wrap or connect third-party agent clients and bring them into Databricks governance.
• Unity AI Gateway becomes the control point for access, policy, routing, budgets, tracing and agent registry.
• This is Databricks’ attempt to govern a heterogeneous AI estate rather than assume all work happens inside Genie.
• The upside is openness – enterprises can use outside agents while applying common policy and observability.
• The limitation is feedback quality – external clients may not expose the same natural-language queries, clarifications and user interactions needed to improve Genie Ontology.
• Agent traces become strategically important because they are the new behavioral exhaust for AI systems.
• Observability for agents could become as important to the AI era as clickstreams were to big data.
• Execution quality is key because long-running agents need recovery, rollback and continuity of reasoning state.
• Agentic business continuity will require protecting not just data, but process state, intent and reasoning context.
• The broader industry risk is fragmentation – multiple clients, multiple ontologies and multiple systems of intelligence competing inside the enterprise.

Omnigent addresses one of the most practical problems in enterprise AI – i.e. the agent landscape will be heterogeneous. Enterprises will use Databricks Genie, Microsoft 365 Copilot, Claude, ChatGPT, coding assistants, SaaS agents, hyperscaler agents and custom-built agents. No single vendor should assume it will own every system of engagement.

This creates a governance problem. If every agentic client connects to models, tools, data sources and workflows independently, the enterprise gets fragmentation, inconsistent permissions, uncontrolled spend and weak observability. Omnigent is Databricks’ answer to that problem. Omnigent is an open-source “harness of harnesses” that can sit around or connect external agent harnesses and route them into Unity AI Gateway and Unity Catalog governance.

The slide above highlights the core governance functions including agent registry, access control, contextual policies, budgets, smart routing and agent tracing. It also shows the breadth of what Databricks wants to govern; specifically models, agents, MCP tools, skills, external agents and AI coding tools. This is the right direction because the enterprise needs a unified control plane for its AI estate, even if the clients and agents come from many vendors.

The strategic leverage is Omnigent lets Databricks participate even when it does not own the front end. If an enterprise wants to use Claude, ChatGPT/Codex, Microsoft 365 Copilot, Cursor, Replit, Salesforce, Amazon AgentCore or another agentic surface, Databricks can still provide governance, routing, policy, permissions, cost controls and observability through Unity AI Gateway. That gives customers openness without abandoning enterprise control.

But there is an important limitation. Governing a third-party client is not the same as co-designing the client with the system of intelligence. A native Genie experience can ask a clarifying question when the ontology is uncertain. It can capture the user’s response. It can convert that response into persistent enterprise context. That is the feedback loop that teaches Genie Ontology.

It is not yet clear that Omnigent can fully replicate that loop for third-party clients. It may be able to govern the agent, trace its activity and route its access through enterprise policy. But if the ontology needs to push a clarification back to a user inside a third-party interface – for example, “what do you mean by platinum customer?” – that may require deeper integration than a governance harness alone provides. Our take is that Omnigent can administer and observe external agents, but it may not capture the same semantic feedback that Databricks gets from native Genie clients.

That distinction means the next platform fight is not only about governance. It is about learning. Whoever captures the user’s intent, ambiguity, correction and clarification has an advantage in improving the system of intelligence. Governance lets Databricks stay in the flow of third-party agent usage. Native engagement lets Databricks learn more directly.

The observability point is equally important. Agent traces are becoming a new class of enterprise data. In the web era, clickstreams became the behavioral signal that powered digital analytics, personalization and big data. In the AI era, reasoning traces, tool calls, action paths, failures, retries, prompts, intermediate decisions and human overrides may become the equivalent signal. They will help enterprises diagnose agent behavior, improve agent performance and train the broader system.

The data volumes could be enormous. Agent traces may be orders of magnitude larger than traditional application telemetry because agents generate reasoning paths, tool interactions and intermediate state over many steps. That makes observability a strategic capability that goes well beyond troubleshooting and alerts.

Consistent and durable execution is another key ingredient. As agents move from short tasks to long-running processes that unfold over hours or days, enterprises need a source of truth for the agent’s state. If an agent fails mid-process, the system must know where it was, what it had already done, what intent it was pursuing, what reasoning path it followed and how to recover or roll back safely. This is not a simple checkbox. It is a maturity journey, similar to the long evolution of transactionality, recovery and consistency in databases.

This also expands the meaning of business continuity and resilience. Historically, business continuity focused on recovering systems and data. In an agentic enterprise, resilience must include process state, reasoning state and intent. The enterprise must be able to recover not only the data, but the operating context of the work in motion. Over time, we expect a new form of data protection and resilience to emerge around agent state, ontology state, process state and enterprise context.

The bigger industry implication is fragmentation. The technology industry almost always fragments when multiple powerful vendors try to capture value. That pattern is already visible in AI. SaaS vendors are embedding intelligence into their applications. LLM vendors are pushing agentic clients and plug-in ecosystems. Hyperscalers are building their own agent platforms. Data platforms are building ontologies and governance layers. Enterprises will make bets across several of these layers, often in different parts of the organization.

Omnigent does not eliminate fragmentation. It manages it. It gives Databricks a way to bring third-party clients and agents back into a governed environment. But it does not fully solve the deeper issue of multiple systems of intelligence emerging inside the same enterprise. That is the next major challenge. In other words, how to govern, reconcile and ultimately rationalize competing intelligence layers so the enterprise does not end up with a new generation of AI silos.

Key Takeaway:
Omnigent gives Databricks a pragmatic way to govern third-party agents in a fragmented enterprise AI landscape. But governance is not the same as learning. The key question is whether Databricks can capture enough traces, feedback and clarification from external clients to improve Genie Ontology – or whether the richest learning loop remains reserved for native Genie experiences.

The Race to Enterprise AGI

Slide setup:
This slide zooms out from Databricks and frames the broader race to Enterprise AGI. The major vendor camps are entering from different positions: frontier labs and copilots through the system of engagement; data platforms through governed enterprise data; SaaS and process vendors through systems of record, workflows and domain process models. But all roads lead to the same control point – the system of intelligence, where business logic, skills, rules, relationships and tacit knowledge become governed corporate assets.

Key Points

• The highest-value real estate in the emerging enterprise AI stack is the system of intelligence – the digital twin or enterprise ontology.
• Frontier labs are entering through agentic clients and copilots, but will likely try to move into the system of intelligence.
• Data platform players such as Databricks and Snowflake are moving up from infrastructure into governed business semantics and enterprise context.
• SaaS and process vendors such as Salesforce, SAP, Palantir, Celonis, Blue Yonder and RelationalAI start closer to business process and systems of record.
• The enterprise begins with thousands of islands – operational apps, packaged apps, custom apps, analytic systems and unstructured content.
• Harmonizing portions of the estate reduces the number of islands and makes the broader enterprise map easier to build.
• A list of connectors is not a map – access to apps and data does not equal understanding of how the business operates.
• Traditional BI clients are being displaced by agentic clients that capture intent, clarification and human-in-the-loop semantics.
• Frontier vendors may evolve through memory and skills – memory becomes state, and skills become logic.
• Personal memory and personal skills must eventually become governed workgroup and enterprise assets.
• Enterprise AGI requires both adaptive intelligence from LLMs and deterministic intelligence from business rules, policies, state and tacit knowledge.
• Volume, distribution and brand matter – the best architecture does not always win.

The race to Enterprise AGI is a wide-open jump ball. Every major vendor camp sees the prize, but each is entering from a different starting point.

Frontier model vendors and copilots are coming in through the system of engagement. They have the user interface, the model intelligence and, in some cases, massive consumer and developer volume. Microsoft has business ubiquity through Microsoft 365 Copilot. OpenAI has ChatGPT and Codex. Anthropic has Claude and emerging coworker patterns. These players are closest to user intent because they sit at the interaction layer.

Data platform companies are coming from governed enterprise data. Databricks and Snowflake have historically lived in the infrastructure layer, but both are moving upward into semantics, governance, catalogs, metrics, ontologies and agentic clients. That move is significant. We no longer think of these companies only as data platforms. They are trying to become enterprise intelligence platforms.

SaaS and process vendors are entering from systems of record and business workflows. Salesforce Data Cloud, SAP Business Data Cloud, Palantir, Celonis, Blue Yonder, RelationalAI and others each have a claim on some part of enterprise process, domain knowledge, operational data or business logic. Their advantage is that they often begin closer to how work actually gets done. Their challenge is that no single application vendor owns the entire enterprise.

The slide makes the convergence point clear: the system of intelligence. This is the layer where data, business logic, skills, relationships, policies and tacit knowledge become a reusable corporate asset. It is the enterprise ontology or digital twin. It is the map that agents use to understand the state of the business, answer questions and operationalize decisions through systems of record.

The starting point, however, is messy. Enterprises have thousands of islands. They have packaged applications, custom applications, operational systems, analytic systems, spreadsheets, documents, workflows, collaboration platforms and unstructured content. Each island has its own definitions, permissions, process logic and data model. Enterprise AGI cannot emerge from that fragmentation unless the estate becomes more harmonized.

This is why the motion from data platforms and enterprise application vendors matters. Salesforce Data Cloud, SAP Business Data Cloud, Snowflake, Databricks and others are each harmonizing part of the estate. They do not solve the whole problem, but they reduce the number of islands. Fewer islands are easier to bridge. An overlay map, a shared ontology or a set of governed connections becomes more plausible when large portions of the enterprise are already organized into coherent domains.

By contrast, a frontier-model agent that exposes a list of connectors is not enough. Connectors give access, but they do not create meaning. A connector to Salesforce does not explain how the enterprise defines a qualified lead. A connector to SharePoint does not identify which policy is authoritative. A connector to Gmail or Slack does not tell the system which customer escalation rule applies. Without a system of intelligence underneath, the agent sees endpoints, not a business map.

This is why traditional BI clients are no longer the center of the system of engagement. Dashboards and reports remain useful, but the new engagement layer is agentic. It captures natural language intent, clarifications, corrections, approvals and work in progress. Those human-in-the-loop interactions are critical because they teach the system of intelligence. The client and the intelligent back end must learn together.

We believe the frontier vendors will move into the system of intelligence because they have to. Their likely path begins with memory and skills. Memory becomes state. Skills become logic. At first, these are personal – a user’s preferences, personal workflows, prompt patterns and saved automations. Over time, frontier vendors will try to expand this into workgroup memory and workgroup skills. Eventually, they will need a way to harmonize those skills, reconcile conflicts, govern promotion and turn them into shared enterprise assets.

That is the moment when personal productivity becomes enterprise architecture. A collection of user skills is not yet a system of intelligence. Without governance, it becomes a Tower of Babel. To become Enterprise AGI, those skills must be abstracted, harmonized, approved and promoted into business logic. That requires something that looks much more like an ontology, rules engine or digital twin than a generic chatbot.

The key architectural point is that Enterprise AGI needs two kinds of intelligence. LLMs provide adaptive intelligence – reasoning, language, summarization, code generation, planning and flexibility. The system of intelligence provides deterministic intelligence – business rules, policies, permissions, state, relationships, workflows and tacit knowledge. The enterprise needs both. A very smart model cannot fully substitute for capturing a company’s own business rules and operating knowledge as corporate assets.

That said, volume matters. The largest foundation-model vendors have enormous usage volume, brand affinity and model-training feedback loops. OpenAI’s consumer footprint is a strategic advantage. Anthropic, Google, Meta and xAI have model-scale ambitions. Microsoft has massive business distribution. Many standalone agents and copilots may struggle to reach those levels of usage. The history of enterprise technology also reminds us that the best architecture does not always win. OS/2 versus Windows remains a useful cautionary tale.

The market will not resolve into one clean, harmonized layer imposed from the top down. Partnerships will form. Alliances will shift. Friends will become competitors. Model vendors will move down into enterprise context. Data platforms will move up into intelligence. SaaS vendors will defend process control. Hyperscalers will try to own the cloud-scale substrate. New players will emerge around ontology, governance, agent observability and business-process intelligence.

Our view is that the winners will be those that connect all three layers – engagement, intelligence and agency – into a closed loop. The system of engagement captures intent and feedback. The system of intelligence turns enterprise knowledge into governed assets. The system of agency lets agents act through that context and through systems of record. Any vendor missing one of those layers will need to partner, acquire or build.

Key Takeaway:
The race to Enterprise AGI is not a race to the smartest standalone model, the best data platform or the most popular copilot. It is a race to own the system of intelligence – the governed enterprise map that combines adaptive LLM intelligence with deterministic business logic, state, rules and tacit knowledge.

Enterprise Software Splits Into Two Categories – Above the Ice and Below the Ice

This slide shows where architecture meets economics. Below the ice sits infrastructure – data platforms, procedural applications and the lower layers of the stack. These layers remain essential, but they increasingly behave like plumbing, with utility-style pricing. Above the ice is where the modeled enterprise lives – the system of intelligence, system of agency and system of engagement working together to learn the business, coordinate humans and agents, and move the market toward outcome-based economics.

Key Points

• Every software era has infrastructure, platforms and applications.
• In the AI era, data becomes the new infrastructure – important, but not where the highest-margin value necessarily concentrates.
• The new platform is the business-process model – the system of intelligence that makes data actionable.
• Data tells us what happened and who was involved; the system of intelligence explains why it happened, what is likely to happen and what should happen next.
• As the ontology or digital twin becomes richer, agents can take broader actions with greater confidence.
• The system of intelligence becomes the shared coordination substrate for humans and agents.
• The new application layer is agentic – agents act through the system of intelligence and systems of record.
• Above-the-ice vendors can move toward value-based or outcome-based pricing.
• Below-the-ice vendors remain more exposed to utility pricing, consumption pricing and margin compression.
• Outcome pricing will be debated, but the ability to measure contribution to business outcomes improves vendor pricing power.
• The long-term prize goes well beyond storing data and gets into modeling the business well enough to improve measurable outcomes.

The final implication is economic. The enterprise AI stack is beginning to split into two broad categories: above the ice and below the ice as shown by the James Bond movie image above.

Below the ice is infrastructure. This includes data platforms, storage, compute, formats, pipelines, procedural applications and the technical plumbing required to make the enterprise run. These capabilities are vital. They do not disappear. But as the market matures, much of this layer becomes more standardized, more interchangeable and more exposed to utility-style economics.

Above the ice is where differentiation occurs. This is where the system of intelligence, system of agency and system of engagement come together to learn the business from modeled data, encode how the enterprise operates and enable humans and agents to coordinate around shared outcomes.

The old saying was that data is the new oil. We think that metaphor is increasingly incomplete. Data by itself is not necessarily the source of value. Data tells us what happened, who was involved, which transaction occurred and which state was recorded. But data becomes valuable when a model makes it actionable. In our view, data is closer to the new hardware – foundational infrastructure that must be present, but whose value is unlocked by the platform above it.

That platform is the business-process model – the system of intelligence. It captures the enterprise’s operating logic, business rules, tacit knowledge, relationships, state and decision context. It allows the enterprise to move from “what happened?” to “why did it happen?”, “what is likely to happen?” and “what should we do next?”

This connects directly to the ontology maturity model. As more of the business is captured in an ontology or digital twin, agents can act with greater confidence and across a wider range of scenarios. The richer the model, the more the system understands context, constraints, cause and consequence. At the lower levels, the system produces better reports and recommendations. At the higher levels, agents can coordinate with one another and with humans through a shared representation of the live business.

That shared representation is the key. Enterprise AGI is not about a single agent completing a single task. It is about creating a coordination substrate for the organization. Humans and agents need to share signals, state, goals, policies and operating context. The system of intelligence becomes the place where those elements are represented and governed. That is where collective outcomes are directed.

This is why the economic model changes above the ice. If a vendor is only providing infrastructure, it is generally priced like infrastructure – consumption, usage, utility, seats or capacity. If a vendor can model the business and contribute directly to business outcomes, it has a stronger claim on value-based pricing. The closer a platform gets to measurable business output, the more pricing power it can command.

Outcome pricing will not be simple. Customers may resist paying software vendors what feels like a royalty on their own business results. They may not want vendors embedded in cost of goods sold or claiming a direct share of revenue improvement, margin expansion, risk reduction or productivity gains. But even if the market does not move all the way to pure outcome pricing, the ability to measure contribution to outcomes changes the negotiation.

A vendor that can credibly show it helped reduce churn, improve win rates, optimize inventory, accelerate collections, lower fraud, shorten cycle time or improve customer retention is in a much stronger pricing position than a vendor selling generic usage units. The value conversation changes from “how much did you consume?” to “what business result did we help produce?”

That is the prize for enterprise software vendors. The data platform alone is not enough. The copilot alone is not enough. The model alone is not enough. The value sits in the modeled enterprise – the system that turns data, processes and tacit knowledge into a governed business-process platform, then lets agents operate through that platform.

This also explains why the industry is moving beyond debates about table formats, data lakes and warehouse architectures. Those matter less and are below the ice. The strategic battle is moving upward. The new platform layer is the system of intelligence. The new application layer is agents. The new pricing frontier is business outcomes.

For customers, the implication is equally important. The question is not just which model is smartest or which data platform is fastest. The question is which platform can help the enterprise capture its unique operating knowledge as an asset. That asset becomes harder to migrate over time because it contains business definitions, policies, skills, process logic, tacit knowledge and live state. The switching costs are technical, operational and organizational.

For vendors, the message is value accrues above the ice. Infrastructure will remain large and important, but the best economics will belong to the platforms that learn how the business operates, improve measurable outcomes and make themselves part of the enterprise operating model.

Key Takeaway:
Enterprise software value is moving above the ice. Data becomes infrastructure, the business-process model becomes the platform, and agents become the applications. The vendors with the strongest pricing power will be those that turn enterprise knowledge into a governed system of intelligence and tie that system to measurable business outcomes.

Action Item for Business Technology Executives

Don’t treat AI as a model-selection exercise. Rather treat AI as an enterprise-intelligence construction project. Within 90 days, every executive team should assign a single accountable owner to start building its owned governed system of intelligence – a living enterprise map that captures data, metrics, business rules, process logic, skills and tacit knowledge as corporate assets that agents can reason over and act through.

The mandate should be to pick one high-value business domain, model the critical objects, define authoritative metrics, harvest user-authored skills, govern the rules and connect agents only where the enterprise has enough context to act safely. The companies that do this will compound proprietary intelligence; those that simply plug frontier models into fragmented systems will rent generic intelligence and call it transformation.