An agent is only as good as the definitions underneath it

Most companies evaluating AI agents believe they have a technology decision to make: which model, which framework, which vendor. Then the pilot starts, and the agent confidently produces a number that finance disputes within the hour. The post-mortem never blames the real culprit, because the real culprit is invisible: nobody in the company ever wrote down what the words mean.

What counts as an active customer? When does a lead become a conversion? Is revenue recognized at charge date or subscription date? Every organization runs on such definitions, and in most of them the definitions live in individual SQL queries, Excel files, and the heads of specific analysts. Researchers call the accumulated cost semantic debt — inconsistent business logic hiding beneath the surface, invisible until two reports disagree.

Most "AI agent" failures are definition failures. The agent didn't hallucinate — it guessed, because guessing was the only option you left it.

Why agents make it worse before they make it better

A human analyst who inherits an ambiguous metric asks around, finds the one person who knows, and quietly patches the gap. An agent doesn't ask around. It re-derives the meaning from whatever it can see — and every new agent you add re-derives it again, independently, from a slightly different angle. Semantic debt that was an annoyance at human speed becomes a contradiction machine at agent speed.

The boring fix

The fix has no glamour and enormous leverage: definitions as governed artifacts. Written once. Versioned like code. Reviewed when they change. Read by every workflow — human or agent — instead of re-invented inside each one.

In practice, taking one metric out of tribal knowledge looks like this:

1. Name the owner. One team owns what "qualified lead" means. Disagreement is resolved there, once — not in every dashboard.
2. Write the definition down — plain-English intent plus the exact logic, side by side, so both a person and a system can read it.
3. Version and reference it. The definition lives in one governed file; every report, workflow, and agent points at it. Change it in one place, and everything downstream updates together.

qualified_lead:
  description: "Contact with a booked call, a company
    domain, and no prior lost deal in 12 months"
  logic: booked_call AND company_domain IS NOT NULL
         AND NOT lost_deal_within_12m
  owner: revenue-ops
  version: 4        # changed 2026-08: excludes re-engaged lost deals

That's it. Not a platform migration — a discipline. And it is exactly the groundwork that decides whether agents compound or embarrass you: an agent reading governed definitions inherits the company's agreed truth; an agent without them inherits the argument.

A rule of thumb

Before automating any workflow, run the two-teams test on every noun in it. Each word where two teams give two answers is a definition you owe the system before the system can owe you results. In our experience the list is shorter than teams fear — a dozen definitions usually cover the workflows that matter — but skipping them is why the last pilot died.