Customer Discovery Interviews for an AI Startup

Customer discovery interviews should help an AI startup understand a real workflow before proposing a product. Ask what people do today, where it breaks and what the consequences are. Enthusiasm for an impressive demo does not establish that someone will adopt or pay for the service.

The US Small Business Administration's planning guidance describes market research and direct customer research as ways to understand demand and competition. The interview method below is an independent practical approach, especially useful when a new model capability makes it tempting to build a solution before identifying the problem.

Pick one workflow and audience

Choose a specific group and task. “Small businesses that want AI” is too broad. “Small importers checking supplier invoices against purchase orders” gives you a concrete process to investigate.

Write the assumptions you hope to test: frequency, current effort, cost of mistakes, buying authority and existing alternatives. Keep them as assumptions until you obtain evidence. Do not turn them into statements simply because several founders repeat them online.

Recruit people who actually perform or own the workflow. A person who likes technology but never handles the task may offer interesting opinions without helping you assess demand.

Ask about a recent real example

Start with “Tell me about the last time you did this.” Then follow the sequence: what triggered the task, what information was available, which tools were used and how the result was checked.

Avoid leading questions such as “Would an AI assistant save you lots of time?” They encourage agreement while hiding the details needed to design a useful product.

Ask for a walkthrough rather than sensitive documents. If a participant chooses to show material, obtain appropriate permission and minimize collection. You can often learn the workflow from a sanitized example without keeping customer identities or financial records.

Understand the cost of the current process

Ask how long the task takes, who else becomes involved and what happens when it goes wrong. Distinguish estimates from recorded measurements. A participant saying “probably an hour” is not an audited time study.

Explore the existing workaround. Spreadsheets, manual review and outsourced help are competitors too. A new application must improve enough of the workflow to justify switching, training and trust.

For a fictional invoice-review service, saving ten minutes may be less important than making the final approval explainable. If a finance manager still has to redo the work because the output is opaque, a faster draft may not create meaningful value.

Map the buying process

The user, approver and payer may be different people. Ask who can authorize a trial, who approves data access and how software purchases are evaluated.

Do not treat a friendly conversation as a purchasing commitment. Separate “interested in hearing more” from “agreed to a defined pilot with an owner and success criteria.” Both are useful signals, but they carry different weight.

Ask about restrictions early. A customer may have requirements around hosting, auditability or integrations that change the product significantly. These are part of the market, not inconvenient details to postpone until after development.

Keep a structured interview record

Write notes soon after the conversation. Attribute statements accurately and do not fabricate quotes for a pitch deck. When summarizing with AI, check the summary against the original notes and preserve uncertainty.

Duck Cloud's text tools can help organize sanitized notes. Keep confidential interview material inside the consented storage process rather than distributing it through public examples.

Compare patterns without pretending the sample is a survey

A small number of interviews can reveal repeated obstacles. It does not establish a population percentage or market size. Report “several interviewees described this problem,” not an invented statistical claim.

Look for contradictions as well as agreement. If one group needs strict review and another wants fully automatic processing, they may be different markets. Combining them into one average requirement can create a product that satisfies neither.

Choose a next experiment that addresses the largest uncertainty. That might be a manual pilot, a prototype or a willingness-to-pay conversation. It need not be a full application.

At the end, ask permission for a specific follow-up. A clear next step respects the participant's time and makes future contact easier to manage.

Conclusion

Good discovery interviews focus on recent behavior, existing alternatives and the buying process. Keep evidence separate from enthusiasm and use patterns to choose the next experiment. An AI capability becomes a business opportunity only when it solves a specific problem people care enough to address.