AI··1 min read
Most agent projects that stall do so for boring reasons: unclear scope, messy data, nobody owning the edge cases. Here is the checklist we run first.
An AI agent sounds like one thing, but it is really three: a model that reasons, the tools it is allowed to use, and the data it can see. When a project goes sideways, it is almost never because the model was not smart enough. It is because one of the other two was never pinned down.
Before writing any code, we ask these five questions.
1. What exact job is it doing?
"Handle customer support" is not a job. "Answer order-status questions and hand everything else to a human" is. The narrower the first version, the faster you find out whether it works.
2. What does it need to read?
List every source the agent needs: help docs, order database, CRM notes. Then check how current and how clean each one is. An agent will repeat outdated docs with total confidence.
3. What is it allowed to do?
Reading data is low risk. Issuing refunds, sending emails or changing records is not. Decide up front which actions need a human to approve them, and build that approval step in from day one.
4. What happens when it is unsure?
Every agent will hit questions it cannot answer. The good ones say so and route the conversation somewhere useful. Design the hand-off before you design the happy path.
5. How will you know it is working?
Pick a handful of real conversations, write down what a good answer looks like, and test against them every time something changes. Without this, every prompt tweak is a guess.
The model is the easy part. The scope, the data and the hand-off are where the work is.
If you can answer all five, you are ready to build. If you cannot, that is fine too, and it is exactly what a discovery call is for.
Written by Agentive Team