You Bought an AI Agent. Great! Can Your Data Support It?

Summary

Adoption of AI is nearly universal, but very few organizations have scaled it, and the reason isn’t the model. It’s the data underneath it. This piece looks at why agents expose a governance and data-definition problem that most companies never solved, using McKinsey and Gartner research, a recent client conversation, and Databricks’ newest product bet as evidence.

[Estimated read time: 6 minutes]

The pilot always works

I’m pretty sure each AI pitch this year has the same shape. There’s a demo, an agent reads a ticket, checks a system, drafts a response, maybe takes another action. It’s impressive. Then the pilot ends, the team tries to put it into production, and it stops working because the agent needed one clear answer to “who is this customer?” and the company has six murky ones.

Models got very good, very fast. The data underneath them didn’t, and that work isn’t in the budget.

The numbers make the case

McKinsey’s most recent State of AI survey found 88 percent of organizations now use AI in at least one business function, up from 55 percent in 2023.

Adoption is basically universal. But only a third of organizations are scaling AI anywhere beyond that first function, and just seven percent say they’re fully scaled across the business. For agents specifically, only 23 percent have scaled one, despite nearly everyone experimenting. Gartner predicted in early 2025 that 60 percent of AI projects would be abandoned by the end of 2026 for lack of AI-ready data, and I’m seeing that every day in conversations with business leaders.

And the value isn’t showing up where boards expect it. McKinsey’s own March 2025 survey found more than 80 percent of companies report no material contribution to earnings before interest and taxes (EBIT) from their generative AI, even as adoption has climbed.

None of that is a story about AI not working. It’s a story about AI being asked to run on a foundation that was never built for it, and it’s a repeat of the BI story we all went through three years ago.

Agents make the data problem worse, not better, and most pitches skip that part

A chatbot that gives a wrong answer is embarrassing. An agent that gives a wrong answer usually does something with it (updates a record, sends an email, kicks off a workflow) based on data never governed for machine consumption.

The governance model most companies run today assumes a human is reading the data and using judgment before taking action. In 2026, agents query enterprise systems millions of times a day and make the call themselves. Only about 30 percent of organizations have reached a governance maturity level that accounts for that. The other 70 percent are running agents on rules written for a slower, more human-paced world.

Retrieve, or compute? That is the question

A conversation I had this week made the retrieve-versus-compute problem concrete. About a year and a half ago, a client built a governed financial reporting layer on a modern cloud data platform, pulling numbers from several different business systems into one consistent format. Their tech strategy is now shifting toward an all-SaaS stack, and the instinct is to skip rebuilding that middle layer and let AI handle the cross-system reporting directly. On the surface, that sounds like progress…but really, it’s the BI mistake all over again manifesting as an agent instead of a dashboard.

The question that matters is whether AI retrieves an answer or computes one. If you ask a model “how profitable was this product last quarter?” and it recalculates that number fresh every time by reasoning over raw data, you’ll get a slightly different answer depending on the day, the prompt, and which context the model happens to pull in. That’s fine for a first draft of a memo. It’s disqualifying for a number going in front of a board.

The fix isn’t a better model, it’s making sure AI queries a governed semantic layer where “profit,” “customer,” and “quarter” already mean one specific thing, so the model retrieves a trusted figure instead of re-deriving one on the fly. Building that layer takes real work up front, more than pointing an agent at raw exports ever will. AI can help build it, though. And once it exists, every use case in that domain reuses it instead of solving the same definitional problem all over again.

And success doesn’t generalize the way people expect. Companies often choose a finance use case first, and it works because the business logic and definitions mostly already exist. The obvious next move is to do the same thing for customers, except there’s usually no equivalent master data management for customer identity. The second pilot doesn’t get to reuse the finance foundation. Each new domain needs its own version of the same unglamorous work: one agreed set of definitions, one governed way to get to them.

Whether AI retrieves an established answer or computes one, ask how the output gets validated. Start with a human checking the output before it goes anywhere important. If the team can’t explain how the output gets checked, that’s the tell that the foundation underneath it isn’t done yet either.

The Databricks Data + AI Summit surprise

At Databricks’ Data + AI Summit in June 2026, the most telling reveal wasn’t a smarter model. It was the preview of Genie Ontology: a context layer for agents unveiled alongside the release of Genie One.

Ontology draws from Unity Catalog’s governed semantic layer, business glossary terms, metric views, domains a company defines once, plus signals from connected systems and the Databricks platform. An OntoRank algorithm surfaces the most authoritative definition when more than one exists. Strip away the product name and this echoes the reasoning of the retrieve-versus-compute argument: an agent queries a governed metric instead of reasoning over a raw table.

What’s new is the mechanism Databricks uses to determine which sources carry the most authority when the available context isn’t uniform. That’s worth paying attention to. A platform with every incentive to sell “just point AI at your data and let it figure things out” instead used its biggest stage of the year to lead with governance, glossary, and metric definitions; agents came second. If that’s where the platform is putting the engineering effort, that’s a decent signal for where a company’s own effort should go, too.

Stop treating “which agent” or “which model” as the first decision

It’s usually the fourth or fifth, not the first. The real sequence is: first, know what the business is trying to accomplish. Second, name and size the specific use cases worth funding. Third, understand what your current data and systems can currently support. Fourth, build the platform those use cases require. Most companies run it backwards. They buy the platform or the agent first and back into a use case when IT asks what it’s for.

That’s exactly how you end up with a data platform nobody asked for, or an agent demo running on data your teams don’t trust yet.

The companies getting return this year are the ones who did the unglamorous work first: building one place where systems of record agree instead of eight conflicting exports, choosing one definition of what a customer or a claim or a margin means instead of five, and standing up a data platform with someone accountable for its quality and freshness. Everything downstream—copilots, agents, dashboards—is only ever as good as that foundation. It’s the least interesting slide in any AI strategy deck and the only one that determines whether the rest of the deck ships.

The AI conversation in most companies is still stuck on the demo

The conversation that matters this year, the one that decides whether anything downstream holds up, is whether your data can survive contact with something that acts on its own. For most organizations, their data can’t yet. That’s the work to do now. Knowing where your data stands and doing something about it is what separates this year’s real progress from this year’s demos.