A one-hour interview becomes many pages of text. Asking AI for “the key takeaways” produces a polished paragraph: the user likes a feature, the team will evaluate it, and the discussion continues next week. What a project needs is different: what was decided, where people disagree, and what evidence is still missing.
AI is useful for cleaning speech, finding topics and drafting follow-up material. It should not turn an ambiguous remark into a decision. Build a record that can be checked first, then let the model organise it.
Keep context with the transcript
Retain speaker, timestamp, interview goal and question list. Without the question context, “that does not work” could mean price, workflow or interface. Sample-check names, numbers, product terms and negative words after transcription.
Use four fixed outputs
- Confirmed facts: current state and observed behaviour.
- Explicit decisions: who agreed to do or not do what.
- Disagreements and hypotheses: differing views and claims without evidence.
- Questions to confirm: what needs another question, test or source.
“We may consider it” is a hypothesis, not a decision. “The customer says the process is slow” is feedback, not proof of a product defect.
Ask for evidence in the prompt
You are an interview-record assistant. From the transcript, output confirmed facts, explicit decisions, disagreements or hypotheses, and questions to confirm. Attach a speaker and timestamp or source quote to every item. Do not turn intent, suggestions or inference into decisions. Mark insufficient evidence as “cannot confirm.” Then propose no more than five follow-up actions, each with an owner suggestion and validation signal.
Before using ChatGPT or Claude, remove phone numbers, emails, identity data, unpublished pricing and other sensitive material.
Make actions closable
“Research the need” is not a task. A useful action states owner, deadline, evidence and trigger: review twenty comparable tickets by Wednesday; if more than five show the same failed step, bring samples to review. AI can surface candidates; people own the decision.
Post-call check
- Does every decision have a speaker and source?
- Did any hypothesis become a fact?
- Was a disagreement flattened by the summary?
- Can each action be verified and closed?
A good interview record is not merely shorter. It lets the next reader distinguish what is known, disputed and still worth asking. AI lowers noise; people make judgement and commitments.