“Add AI” is not a product requirement. It becomes a useful decision when you can name the task, the input, the output, and what happens if the output is wrong.
Describe the job without naming a tool
Start with the workflow: “A founder reviews 30 customer notes each week and needs to spot repeated objections.” That statement identifies a task and a possible measure. “We need an AI assistant” identifies neither.
Ask whether the work is repeated, whether the input is available and appropriate to use, and whether a better result would change a real decision. A manual or simpler software step may be enough for a first test.
Put a person at the right review point
Define which errors are acceptable, which are costly, and who checks the output. A summary used as a prompt for a founder’s own review is different from an unsupervised decision sent to a customer. The second needs much stronger controls.
Do not feed client-confidential or sensitive material into a tool simply because it is convenient. Check permissions and data handling before the workflow is tested.
Run a small comparison
Measure time, quality, and failure rate against the current way of working. If the AI-assisted version does not improve a defined outcome, remove it. If it helps only after extensive correction, count that correction time.
The MVP Scope-Down Worksheet gives this choice a place in the build/defer decision. Keep AI only when it improves the user or founder workflow you have chosen to test.