The Number Was Never Missing. The Definition Was

Are you getting the most out of your AI agents? Here’s a test you can run on any AI agent your company is piloting right now. Ask it, “How many qualified leads did paid social drive last quarter?”

Likely, you’ll get a nicely formatted answer, delivered fast and presented with total confidence. Here’s the catch. There’s more than a decent chance it’s wrong.

That’s because within that innocent-looking question there are at least three choices, and since no one created the single definition, there is no way for the AI agent to know the answer.

For example, a “qualified” lead could be one that cleared marketing’s scoring threshold. It could also be one that sales actually accepted. “Drive” could mean first touch, last touch, or some multi-touch attribution model. And “last quarter” could be fiscal or calendar.

Your marketing ops lead knows their definition. Your head of sales knows their definition. Most likely those definitions are different. And the agent doesn’t know either, so it guesses. Ask again next week, and I’m willing to wager that it guesses differently.

Your Pilots Are Stalling

The industry isn’t staying quiet on this topic. According to IDC, for every 33 AI prototypes companies build, only 4 reach production. That’s an 88% failure rate that the firm blames on the readiness of data and processes, not the models.

Gartner is even more specific. In its 2026 Market Guide for Agentic Analytics, it predicts that by 2028, 60% of agentic analytics projects relying solely on MCP (Model Context Protocol) will fail due to a lack of a consistent semantic layer.

In August, Gartner went further, naming AI context platforms as their own category and positioning them as software that builds and governs business meaning, then delivers it to AI agents. It expects most enterprises running agents in production to standardize on one by 2028. When an analyst firm creates a category, a problem has moved from theory to budget line.

The Layer Everyone is Missing

In his post, he pointed to Gartner’s AIR Layers framework that splits the enterprise stack into three parts:

• System of Action, where agents sit

• System of Intelligence, where context sits

• System of Record, where data sits

While that middle layer is the key, most AI budgets invariably go to the shiny agents on one end (the top) and warehouses on the other (the bottom). The layer in between, the one that tells an agent what your data means, is usually empty, which is why agents are often left to guess.

More Horsepower is Not the Answer

The answer for many businesses? Purchase bigger models or better prompts. But here’s the issue with these approaches. They treat a definition problem like a capability problem.

Companies continue to dedicate enormous budgets to Snowflake, Databricks, BigQuery, and marketing stacks with more logos than a NASCAR hood. While that money gave them storage and speed at scale, it cannot tell an agent what a “qualified lead” is or assign attribution accurately. The data exists, but the meaning was never codified in the system.

A company can hire the sharpest marketing analyst in the world. If you ask them on day one what counts as a qualified lead, they would ask you to provide the definition of “qualified.”

Why Mined Definitions Plateau

There are tempting shortcuts available today. One requires you to point an AI at your query history and old dashboards, and let it reverse-engineer what your terms probably mean.

It’s a clever idea, but it comes with a built-in ceiling. These tools can only learn from questions somebody already asked. Ask something new, like how a campaign you launched last month performed, and they’re back to guessing.

Typically, when it comes to accuracy, these tools plateau in the mid-80% range, with no signal about which answers land in the other 15%. That figure may be fine for a demo, but it’s not acceptable for a pipeline number in the board deck.

Even worse, mined definitions have no owner, so when the CRO asks, “Who decided what counts as a qualified lead?” The honest answer is nobody. The model simply inferred it from old SQL.

Ask First, Then Verify

Pure automation is the mining shortcut I touched on above, while armies of consultants don’t scale, and what they learn walks out the door when the engagement ends. What works is automating the capture of expert knowledge, not replacing the experts.

We call it forward engineering and not reverse engineering. With forward engineering, you      go to the domain experts who actually know what the data means, treat what they say as ground truth, and only then generate the semantic models, dbt code, test and documentation, reconciling every generated artifact against the warehouse before it’s allowed to ship.

Automation without an accountable human checking the results isn’t a feature; it’s the risk the industry keeps rediscovering the hard way.

What a Wrong Guess Can Cost You

Picture budget season and the CFO wants to cut 20% from the marketing budget. You walk into the board meeting with an AI-generated slide touting how your paid social sourced 40% of qualified pipeline.

Then a member of the CFO’s team asks who defined “qualified” and which attribution model counted as “sourced.” You don’t know, because the agent picked it for you and didn’t say.

That goes beyond an awkward moment. Your budget is walking out the door, and maybe your credibility with it. With a governed definition, you’d have the answer in one sentence and a name to back it up.

Define The Terms

If your AI pilots keep stalling, don’t start by shopping for a smarter model. Ask whether anyone at your company owns the definitions your agent depends on.

If the answer is no, your System of Intelligence, the “I” in the AIR Framework, is empty, and that’s your bottleneck. And as the executive who has to stand behind the pipeline number, it’s on you to fix it. Define the terms first and then deploy the agent. You’ll find the number was never missing. The definition was.