A Practical Guide to Problem Interviews for AI SaaS

This is the pillar guide for AI SaaS teams running problem interviews. The framing is intentionally tilted toward the AI-specific patterns that other interview guides miss: trust boundaries, hallucination tolerance, the difference between hopes and adoption, and how to talk to skeptics. Read through and you have everything you need to start the round this week.

Step One: Define the User by Two Filters

Before any outreach, define the user. Two filters: a job-or-role and a recent behavior. Examples: legal ops leads at companies between 200 and 1000 employees who reviewed at least one contract last week. Customer support managers at SaaS companies who use a ticketing tool. Marketing operators at B2B firms who shipped at least one campaign last month.

Avoid filters like "people interested in AI." That selects for ideology, not workflow. The two-axis filter selects for actual past behavior, which is what you need.

Step Two: Reach Fifteen of Them

Plan to land fifteen calls. Reply rates of 25 to 40 percent are normal, so plan to send 40 to 60 messages. LinkedIn is the most reliable channel for B2B AI SaaS audiences. Niche communities work for specific roles.

The message is three sentences. One on why specifically them, one on what you are researching and that you are not selling, one asking for thirty minutes. Do not include "AI" in the subject line. It self-selects for AI enthusiasts and skews your sample.

Step Three: Build a Three-Question Spine

Three story prompts about specific past events. The rest of the call is follow-up.

Question one: walk me through the last time you did X. What were you trying to do, what did you actually do, what did the output look like?

Question two: have you ever tried an AI tool for this? What happened?

Question three: where in this workflow do you check the output by hand? Why there specifically?

Question two and three are AI-SaaS-specific. They surface the trust boundary that determines whether your product can succeed.

Step Four: Run the Call

Open with a low-stakes intro. "Thanks for the time. I am researching how teams handle X right now, I am not selling anything today, I just want to learn from people who have lived through it." Notice the intro does not mention AI.

Ask question one. Then shut up. After every answer, count to three before responding. The silence often produces the most useful sentence of the call.

If they ask what you are building, deflect: "Still figuring it out, that is why I am asking these questions. Can we come back to that at the end?" They will say yes.

Step Five: Capture AI-Specific Vocabulary

When the customer says something specific or vivid about AI, write it in quotes. "Hallucinated." "Sounded confident but was wrong." "Got it almost right." "Cooked the result." "Better than I expected for the easy stuff." "Could not be trusted with the financial part."

These phrases are messaging gold. They will end up in your landing page, your sales emails, your product copy. The customer is handing you the words they use among themselves about AI tools. Use them.

Step Six: Map the Trust Boundary

Specific to AI SaaS. For each interview, write down where in the workflow the customer trusts AI output and where they do not. Why there? What changed? What would have to be true for them to extend the trust boundary?

The trust boundary is the design constraint your product has to respect. Most failed AI products failed because they crossed it. Most successful AI products succeeded because they stayed inside it and earned their way out gradually.

Step Seven: Synthesize Every Five Calls

Reread the previous five sets of notes back-to-back. Look for repeats. Same trigger event. Same workaround. Same vocabulary. Same trust boundary. Same competing tool.

Patterns are signal. Idiosyncrasies are interesting but not actionable. After fifteen calls you should be able to write one paragraph that names the user, the workflow, the trust boundary, the cost, and the smallest version of an AI-augmented better solution.

Step Eight: Decide What You Have

Three questions. Is the workflow real, frequent, and painful enough that people are doing something ugly to handle it? Is the trust boundary clear and respectable - can you build a product that operates inside it without faking trust the customer has not granted? Is the cost data quantifiable enough to anchor pricing?

If yes, yes, and yes: build the smallest possible version and put it in front of the same fifteen people. If any is no, you have learned something equally valuable.

Common Failure Modes

Opening with the AI question. Resets every call to ideology mode. Skip it.

Demoing your prototype mid-call. Contaminates the data. Save demos for separate solution interviews.

Treating future hopes as adoption signals. Hopes do not predict use. Past spend does.

Skipping the failure stories. People who tried AI tools and quit are some of your highest-signal interviewees.

Avoiding the skeptics. The skeptics' objections are your highest-value content.

When You Should Skip This

Two cases. One: you are the user, you have lived in this AI-augmented workflow yourself for a year, your gut already encodes the patterns. Two: you are running a one-week throwaway prototype to learn a new model, not a real launch attempt.

Most AI SaaS founders are in neither case. The cheap test: spend a week on interviews before deciding you are exempt. If they surface nothing surprising, you might be. If they surface anything at all, you were not.

The Trade

Twenty-five to thirty hours over two to four weeks. At the end, you have a one-paragraph spec grounded in real conversations, a list of fifteen people to launch to, and a clear sense of where AI belongs in the workflow versus where the human review remains. The product that follows is narrower, more focused, and aimed at a real trust boundary.

The alternative is shipping into a market shaped by your hypothesis about AI use, not the market's actual workflow. The cheaper test goes first. Place the bet better and let the AI engineering you are good at compound on a target you actually got right.