Your AI doesn’t know what your company knows

I built an agent, not too long ago, that watches how customers use a feature I own. It runs on its own schedule and pushes its findings to me each morning. It’ll report if a customer’s workload is starting to see performance worsen or suggest actions to help a company with a low “health score.” Its first real run returned 15 insights, tiered according to severity: three Critical, four High and the rest Medium. One of the Criticals was genuinely useful. The rest were unobjectionable and inert (save for two Highs, which were flat-out wrong).

I say this as someone who builds these systems for a living, who chose the data and who wanted this to work. But the result was still just one usable finding out of 15. If you're an executive who has funded an analytics agent program this year and wondered why the output feels like homework, this is why.

The explanation I hear most often is volume, that agents generate far more output than any organization can absorb. I've heard many people argue that this is the central problem of enterprise AI right now. The MIT Technology Review Insights report, "Raising the Quality of Insights in Agentic Analytics," found that fewer than half of the insights generated by analytics agents lead to actions that improve the business. That figure is up from a year ago, but it still means most output is wasted.

But frankly, I don’t believe that volume, on its own, is the issue. Agents certainly produce more output, and eventually that will need managing, but treating volume as the problem right now leads organizations somewhere expensive: building routing, triage and prioritization machinery for a stream of “insights” that mostly don’t deserve it. You end up with a well-governed pipeline for material that was never worth governing.

The reason my agent gave me 14 mediocre (or worse) findings is not that it was too productive. It simply didn’t know what mattered. It had the data, but what it lacked was everything that gives that data meaning.

For example, take a phrase every company uses but no two define the same way: active customer.

I am an active customer at the place where I get my car washed. I go every six or eight weeks, which is exactly the rhythm they expect, and by their definition I’m in good standing. But if I went to, say, my local grocery store once every six to eight weeks, there is no way I would be considered an active customer. One definition does not fit all, and trying to treat them as such would lead companies to draw the wrong conclusions about their own business and act on faulty insights. This even happens within an organization; finance’s definition of an active customer can be meaningfully different from the way marketing defines it, for instance.

This holds back the vision of an agentic enterprise because agents don’t know which definition applies; none of it is encoded in the database. There’s no column called active customer; there are behaviors, dates and identifiers. What ties them together is a judgment that reflects the product, the sales motion and the market. In most organizations that judgment has never been written down anywhere. It tends to live in queries, BI reports, Excel sheets and teams’ memories.

Usually, the people who hold that knowledge are the analysts. They understand the business; they understand the data; and until recently the third thing they did — write SQL — was the thing we measured them on most. That third thing is now close to free; none of the analysts on my team write code anymore. What remains is exactly what our agents are missing.

That is ambient context, and it’s the most valuable analytical asset that most enterprises own. It is also, at the moment, almost entirely invisible to the agentic systems acting on it. An agent handed a well-modeled data set and no context will produce fluent, plausible, unusable analysis. Worse yet, it will do it quickly and at scale.

Where these definitions exist at all, they were built for a different job. For 20 years, the semantic model has lived inside BI tools, and that made sense. The definition existed to produce a report for a person to apply their own judgment and decide what to do. The tool never had to be more precise than the human reading it.

AI systems, however, break that arrangement in a very specific way: They don't always act consistently. If you ask the same question twice, you aren't guaranteed the same response, and that demands a kind of precision that the old semantic model can’t satisfy. You need verified queries for the questions that matter enough to have exactly one right answer. You need synonyms, because people use vocabulary your columns may not. You need relationships and metrics stated explicitly rather than left for someone to infer.

And when this is available where AI is actually running, not sitting in a reporting tool outside it, customers have true AI governance.

Most large enterprises already run several systems capable of generating analysis, and the truth is they are going to run more. If context is something you author separately inside each one, then every tool you buy carries a second cost: recreating institutional knowledge, by hand, in a new place, and then keeping that copy honest. That cost recurs, and it scales with your vendor count. So if you feel as though you’re paying repeatedly for the same work, chances are you’re not imagining it.

That is precisely why I’d argue the state of your organization’s context is even more important than which model you standardize on. It needs to exist in a form that can be shared across the systems you have chosen, without relying on someone to rebuild it again and again.

When I think about the rate of change and transformation happening right now, my mind goes back to the Industrial Revolution or the Renaissance. I think about what happened after Gutenberg. The printing press made information dramatically cheaper to produce, and as a result, there was a sudden deluge of content available, much of it unreliable, disorganized and difficult to navigate. The real breakthrough came not from the press but from what people built around it over the next century: indices, tables of contents, catalogs, editorial standards, libraries. The technology that produced the abundance was available long before the infrastructure that made it useful.

Buying the model is buying the press. The harder, less exciting work is getting what your organization knows into a form that the system can use.

Do that, and you arrive at the volume problem for real: hundreds of high-quality findings, genuinely worth acting on, more than any team can pursue. That’s a good problem, and one that almost nobody has earned yet.