Find the assumption behind your proposed feature. An AI assistant can help identify it; a dated customer record can help you check it.
I have mentored more than 30 people who now have businesses at different stages. A question from that work belongs in any product review: what could a customer do that would change our mind? For a manager deciding whether to spend another week on a feature, the answer should fit on one page.
Give the assistant a claim you can check
Suppose your team is considering a tool that flags mismatched line items during invoice reconciliation. A colleague says finance teams will use it every week. That sentence bundles a claim about the problem with a claim about demand for your product. Start with the problem. Define your target group as small wholesalers whose finance teams reconcile invoices each week. Ask whether at least half of those teams find a mismatched line item in at least three of four weeks. The person who checks invoices and the person who can approve spending may be different people. Save questions about purchase for a separate test.
Ask your assistant
“Read this proposed feature and the approved notes below. Find the assumption that could make the feature unnecessary. Write one observation that would support it and one that would weaken it. For each existing claim, give the note or record reference, or mark the claim unverified. Suggest a check that fits our time limit. Leave the decision to the reviewer.”
Give the assistant only material you have permission to use. Keep customer names and invoice details in the approved system. If the assistant says several firms have the same issue, open the cited records. Check the dates, how the firms were recruited and whether those notes came from separate teams. A plausible objection is useful for planning a check. Its wording gives it no extra weight.
Update your estimate when a record comes in
Write down your starting probability before you look at the record. Bayes’ theorem gives you a way to update it when evidence arrives. Then ask how likely that observation would be if the claim held and how likely it would be if the claim did not hold. A record that is about as likely under either possibility tells you little about which one holds.
Here is a teaching example. Suppose your starting probability for the wholesaler claim is one in four. Select a team and week by a defined rule. Your evidence is a reviewer verifying a dated record of a mismatch for that team and week. A team can have a mismatch in one week even when fewer than half the teams meet the recurrence threshold. For this example only, suppose the record would appear in three out of five cases where the claim holds and in one out of five cases where it does not.
In the example the record raises the probability from 25% to 50%. The theorem is a mathematical identity. The prior and the two probabilities of observing the record were chosen to demonstrate it. A founder cannot use the result as a measured probability of customer demand. Two people describing the same invoice also give you one underlying event. Keep the shared source visible before you try to combine their accounts.
Decide what the next week is worth
Before looking for records, agree who qualifies for the test. Recruit from an agreed list of eligible firms and record how you chose them. A team drawn from a queue of unresolved invoice complaints will answer a different question from a team drawn from that wider list. Distinguish a checked week with no mismatch from a week for which nobody could supply a record. Use this first set of records to decide whether a broader test is worth the effort. Leave probabilities out of the decision memo when you have no defensible basis for them.
Ask a colleague who did not write the proposal to check each observation against the source. Put the claim, the observation that would change your mind, the spending limit, the review date and that reviewer’s name on one page. For the reconciliation example, spend at most one week checking dated records from the previous four reconciliation weeks. Leave the software build for a later decision. Count only teams with four documented weeks in the recurrence comparison. Report how many teams that excludes and why. If at least half the teams with complete records meet the threshold, ask the named budget holder to compare the cost of the current workaround with the cost of changing it. If the records show occasional mismatches that staff already handle easily, stop the feature. If too few teams have complete records, hold the decision and record what evidence is missing.
Record why your team made its decision. Include the evidence that changed your mind or justified another test. Link to the records in the approved workspace so someone can revisit the decision later. For a wider team trial, see how to measure the work of an AI pilot.
Sources & further reading
MIT OpenCourseWare, Introduction to Probability and Statistics, Class 3: Bayes’ rule ↗NIST, Generative Artificial Intelligence Profile, source verification action MS-2.5-003 ↗Sources checked 2026-11-13.
