By Scott Sumner | AI at Work
AI customer research can turn a folder of interviews into a confident summary. The danger is losing the condition that made a customer’s answer useful.
AI can help analyze customer interviews by extracting quotes and grouping themes. Keep every finding linked to its source, separate observation from interpretation, and review contradictory evidence before changing your offer or website.
A faster summary is useful. A summary that quietly changes what customers meant is an expensive brief for your next campaign. Here is how I would use AI to do the sorting without letting it settle the argument.
The small word that changes the finding
Consider this fictional customer interview:
“I’d pay more if someone handled the migration. I don’t want another monthly subscription.”
Now imagine the summary says: “Customers are willing to pay more for a premium plan.”
The sentence sounds plausible. It also changes the decision. The customer described paying for a particular piece of work, with a condition attached. The summary turned that into enthusiasm for an ongoing product.
This is an illustrative failure, not a result from a client project or a measured error rate. It shows why checking for invented facts is not enough. You also need to check whether a summary preserves the meaning of real statements.
The UK Government’s research-analysis guidance makes a useful distinction: record what you saw or heard first, then work out what it means, then decide what to do. It recommends keeping observations, findings, and actions separate. The guidance was published in 2016. AI did not make that discipline obsolete.
Build an evidence sheet before a summary
My recommendation is to ask AI for an evidence sheet before asking for an executive summary. That gives a reviewer something concrete to inspect instead of a finished argument to accept or reject.
Give each interview a stable identifier, such as Interview 04, and preserve timestamps or paragraph numbers. A useful row contains the exact quote, its source location, the proposed theme, and any condition that changes its meaning.
| Field | Fictional example |
|---|---|
| Observation | “I’d pay more if someone handled the migration.” |
| Source | Interview 04, paragraph 18 |
| Interpretation to check | Migration effort may be a buying obstacle. |
| Important qualifier | The next sentence rejects another monthly subscription. |
| Possible next step | Test a clearly scoped migration service. |
Notice what the sheet does not say: launch a premium subscription. That decision needs more evidence.
Also distinguish participants from mentions. One person who repeats an objection six times is still one person. If you report a count, state the unit, the sample size, and whether each participant had a comparable opportunity to discuss the topic. A theme appearing in a handful of interviews is not automatically a population estimate.
What citations can and cannot prove
Source-linked output is a real product capability, not just a better-sounding instruction. Anthropic’s citations documentation describes how Claude’s API can reference specific locations in supplied documents, including character ranges in plain text and page ranges in PDFs.
That can make checking a claim easier. It does not make the claim correct.
A citation can point to the right passage while the interpretation misses a condition, merges two different speakers, or gives a minor complaint too much weight. Read the surrounding exchange, including the question that prompted the answer. If the transcript looks doubtful, check the original recording where available and permitted.
There are product limits too. The cited documentation says image citations are not supported and that scanned PDFs without extractable text are not citable through this feature. That is a limit of this citation mechanism, not a claim that all AI tools cannot read scans.
For an off-the-shelf research tool, ask to see the actual path from a finding back to the transcript. A decorative source badge is not enough.
A practical interview-analysis workflow
Start with a decision, not “find insights”
“What prevents a buyer from booking an initial consultation?” is a useful research question. “Tell us what customers want” leaves too much room for a polished but directionless report.
Choose material relevant to that question. Keep prospects, current customers, and lost deals distinguishable. Their answers may describe different stages of the buying process, not a contradiction to average away.
Prepare the material responsibly
Use a tool approved for the information involved. Check your organization’s permissions and the supplier’s data-handling settings before uploading recordings or transcripts. Remove identifying details that the analysis does not need. Keep the restricted originals available to authorized reviewers.
This is an operational safeguard, not a substitute for your organization’s privacy or legal review.
Extract before interpreting
Here is a starting instruction to adapt:
Using only these transcripts, extract passages relevant to why a buyer does or does not book a consultation. Preserve the exact wording. Include interview ID and paragraph number. Separate each observation from your interpretation. Preserve conditions, uncertainty, and disagreements. If the source does not support a conclusion, say so. Do not invent quotes or turn paraphrases into quotations.
This prompt sets expectations. It is not a guarantee. Check the output against the source documents.
Ask what argues against the leading theme
Once themes emerge, ask for the strongest contradictory passage and the material that does not fit. An exception may reveal a different customer group, a missing follow-up question, or a reason your preferred campaign message will fail.
I would review every passage behind a recommendation that could change pricing, positioning, or a public promise. For exploratory sorting, start with a manageable batch and check enough rows to understand the errors before expanding the workflow. Do not label the whole dataset verified because a few examples looked good.
Make the final decision outside the summary
Bring the evidence to someone who understands the service and, where possible, someone who heard the interviews. Decide whether the finding supports a change, a small test, or another research question. GOV.UK’s guidance recommends involving research observers in analysis to reduce individual bias. AI output is another input to that discussion, not a replacement for it.
Turn research into a test, not a claim
Customer language can improve a service page, but not every quote belongs on a website. Private research is not automatic permission to publish a testimonial. And one person’s desired outcome is not proof that you deliver it.
In the fictional migration example, a sensible website test might explain what a migration includes, what the buyer must supply, and whether the work is separate from an ongoing subscription. Measure whether that explanation resolves the question in subsequent conversations. Do not jump straight to “Customers love our premium plan.”
This connects to AI Findability: clear, accurate service information helps people and search systems understand the business. The purpose of research is to improve that information, not manufacture stronger-sounding proof.
For the broader workflow, the context behind an AI task matters as much as the instruction. Here, the context includes who was interviewed, why they were selected, and what decision the research is meant to inform.
One thing to do next: take one AI-generated customer finding and trace it back to the complete answer. Does the recommendation still hold when you restore the qualifiers?
If your service pages are built around assumptions rather than the questions buyers actually ask, talk with Scott about the website and AI workflow behind them. The useful starting point is the evidence you already have.
Sources and scope
- GOV.UK: Analyse a research session. Published and last updated May 24, 2016. Supports the distinction between observations, findings, and actions, plus collaborative review.
- Anthropic: Citations. Live product documentation checked September 18, 2026. Supports the described citation formats and limitations.
This is an evergreen operating guide, not a report of a new research study. The interview, evidence row, and website test are hypothetical. The recommended workflow is my application of the cited principles, not a vendor-validated accuracy claim.