Writing requirements an AI agent can follow
AI agents fail for the same reason projects fail: vague requirements. Here is how a business analyst would brief an agent.

When an AI agent does something odd, the first instinct is to blame the model. Most of the time, the real problem is the brief. We asked it to “handle customer emails” without saying what handling means.
Write it like a process spec
The same discipline a business analyst uses for software requirements works for agents:
- Trigger. What starts the work? A new email, a form, a status change.
- Inputs. Which systems and fields can the agent read?
- Steps and rules. The order of work, the decisions and the rules behind each decision.
- Outputs. What exactly should exist when it finishes: a record, a draft, a message.
- Exceptions. What to do when information is missing or the case does not fit.
Add acceptance criteria
For each task, write how you will judge success. For example: “Every invoice is matched to a purchase order, or flagged with a reason.” This becomes your test plan and your monitoring rule.
Test with real cases
Collect twenty real examples, including messy ones, and run the agent against them before go-live. Note where it fails and tighten the rules. It is the same user acceptance testing we have always done, just for a new kind of worker.
Clear requirements turn an unpredictable agent into a reliable one.


