Disagreement can be the market map
Conflicting customer interviews become useful when the context behind each claim stays visible.
When customer interviews disagree, the disagreement is often more useful than the average.
On September 25, I compiled an Insuveo research memo from conversations across broking, underwriting, claims, servicing, distribution, and carrier contexts. Insuveo is still an exploration into how AI might improve commercial insurance workflows. There is no validated product, no revenue, and no completed pilot behind the memo.
The memo could have become a clean list of recurring problems. There were real repetitions. Information arrives incomplete. People follow up across organizations. Decisions and documents move through several hands before the official system records the result. Yet the more useful pattern was variation.
A problem that feels urgent from one position can look secondary from another. Commercial underwriting itself is highly heterogeneous. Risk engineering, market conditions, wording, exposures, portfolio composition, broker quality, and pricing can all change how a case is judged. A single phrase such as "the underwriting workflow" compresses too much.
That variation changed how I read contradiction in customer research.
Disagreement can define a boundary
Two accounts can appear inconsistent because the context that makes each one true is missing.
A broker may experience the work as a coordination problem because they are moving information between the client and insurer. An underwriter may experience the same case through the quality of the risk information and the limits of their authority. A claims or servicing team enters at another point, with different handoffs and consequences. Combining these accounts into one general pain point would make the research sound clearer while making the market less accurate.
The disagreement may be showing a segment boundary. It can reveal that the workflow changes by product, organization, role, or decision. That boundary matters because software built for a repeated coordination task is different from software involved in judgment.
I am still testing where those boundaries are. The memo does not establish which workflow should come first, who the buyer is, or whether the work is standardized enough for a useful product. Willingness to pay, integration requirements, and the right level of human review remain open questions.
A summary can remove the evidence I need
AI is useful for organizing a large body of notes. It can group themes, retrieve related passages, and help turn scattered observations into a working document. The risk is that a polished summary can make uneven evidence look settled.
Most summaries reward repetition. The common themes move upward. Exceptions become caveats, then disappear. That may be acceptable when the purpose is a quick recap. It is dangerous during product discovery because the exception may explain why a proposed workflow works in one setting and fails in another.
This is a concern I am carrying into the tools I build, rather than a result the memo proved. A reliable research system should preserve the claim alongside its setting. It should be possible to see which workflow produced an observation, where the evidence conflicts, and what remains inference. Retrieval should help me return to the underlying material instead of replacing it with a smoother conclusion.
Granveo taught me a related lesson. A saved conclusion loses value when its sources and revisions disappear. In customer discovery, the consequence is sharper. A synthesis without provenance can turn an uncertain market into an imaginary consensus.
Better synthesis changes the next question
Preserving disagreement should make the next decision more precise. It helps me ask smaller questions.
Instead of asking where AI can help insurance, I can ask where incomplete information creates repeated coordination work within a specific case. I can examine what happens before a referral, how people know that a submission is ready, or where context is lost between underwriting and claims. Each question is narrow enough to meet a real workflow.
One working hypothesis from the research is that insurance may need autonomous coordination before autonomous decisions. An AI workflow could structure incoming material, identify missing information, and support follow ups while people keep responsibility for judgment. That is still a hypothesis. The next useful evidence would come from access to a live workflow and observation of where the idea breaks.
The memo left me with unresolved contradictions and a better sense of what they contain. Some may be noise. Others may mark the boundary of a product, a buyer, or a workflow. While Insuveo remains research, keeping that context visible is more useful than producing a cleaner average.