The assumption your agent never questioned

A mid-size retailer built an agent to score leads by purchase intent. The logic was clean: any customer who visited a product page more than three times in seven days got top priority. It worked well for months. Then the team restructured the website and the product page became the default landing page for all paid campaigns. The agent kept scoring leads with the same formula. The sales team started complaining that the "hot leads" weren't converting. It took six weeks for anyone to connect the dots.

That is the problem of inherited assumptions. It is not a technical failure. The agent was working perfectly. What had aged was the business assumption buried inside the scoring rule.

How an assumption enters an agent

Business assumptions reach automated workflows through three main paths.

The first is the prompt. When you write "prioritize customers with an average deal size above $20K," you are encoding a strategic decision — which segment matters most — directly into the instruction. If that decision changes, the prompt becomes outdated with no visible warning.

The second path is explicit business rules: conditionals, filters, thresholds. "If the lead scores X, do Y." These rules typically live in configuration files that no one revisits after go-live.

The third is training data or historical data used to calibrate the agent. A model trained on 2022 data carries the behavioral patterns of that period. If the market has shifted, the model does not know. It keeps responding with the logic of the past.

Why assumptions age without warning

Human systems have natural friction. When a strategy changes, someone needs to communicate, align, and convince. That friction, costly as it is, acts as a propagation mechanism: the change moves through the organization in a visible way.

Agents in production have no such friction. They execute quietly, consistently, and deliver results that look normal until the drift accumulates enough to show up in a business number.

The risk compounds with scale. An agent processing a hundred interactions a day with a wrong assumption generates a hundred errors a day. An agent processing ten thousand generates ten thousand. Automation does not surface the problem. It multiplies it.

What to audit before scaling

Before increasing the volume of an agent already in production, three questions need answers.

What is the core assumption of this agent? Write it in one sentence — what the agent assumes to be true about the business. "This agent assumes that qualified leads are those who visit the product page more than three times." If you cannot write that sentence, the agent has an assumption no one has mapped.

Is that assumption still true? Check with whoever made the original decision. Compare against recent data. An assumption from six months ago may have been invalidated by a product change, a new campaign, or a market shift.

What happens when it fails? Simulate the scenario where the assumption is wrong. Does the agent produce a neutral result, a bad result, or a result that looks good but is not? The third case is the most dangerous, because it triggers no alert.

The minimum review ritual

I am not proposing a bureaucratic process. I am proposing a simple cadence: at each strategic planning cycle, or whenever a relevant business decision changes, someone with business context — not just technical context — reviews the agents in production using those three questions.

In practice, this means keeping a living document with the assumptions behind each agent: what it assumes, when that assumption was last validated, and who is responsible for keeping it current. That is it. No special tooling required.

Governing an agent starts with governing the assumptions that feed it.

AI agents are excellent at executing. They are poor at noticing when what they are executing no longer makes sense. That awareness is human work — and it needs to be on the agenda before any decision to scale.

If you have agents in production today, the practical question for this week is: can you write, in one sentence, the core assumption behind each of them? If the answer is no, that is where to start.