AI customer interview questions for B2B marketers

Search results for AI customer interview questions are mostly hiring guides, because recruiters publish more than marketers. For B2B teams the phrase means discovery questions about how customers use AI tools, plus a prompt that writes an unbiased script. This page covers that sense and points the other two elsewhere.

What are AI customer interview questions?

AI customer interview questions split into three senses, and the one you mean decides which section you read.

• Questions you ask customers about AI: Discovery questions for B2B customers about how they use ChatGPT, Claude, Copilot or the AI features inside tools they already pay for. Start with the question bank below. Sample question: “Walk me through the last time an AI output was wrong in a way that mattered.”
• Questions an AI asks customers: An AI moderator runs the interview from your guide and probes answers in real time, with no human on the call. Skip to the AI-moderated sections. Sample probe: “You said the rollout stalled. What happened in the week before that?”
• Questions for AI-role candidates: Hiring loops for ML engineers, AI product managers and customer-facing AI roles. Go to the hiring section near the end. Sample question: “When would you choose retrieval over fine-tuning, and what would change your mind?”

The first sense has the least published material and the highest value for a growth team, because the answers feed the VoC artifact your content and positioning run on.

What should you ask customers about their AI use?

Every question in this bank asks about a specific past event, never a hypothetical, because hypothetical answers are the ones customers get wrong. The Mom Test’s three rules are to talk about the customer’s life rather than your idea, ask about specifics in the past, and talk less than you listen. Teresa Torres adds the mechanism: people answer questions about future behaviour based on what they know they should do, so the goal of each interview is one detailed story about something that already happened.

You will not ask all 32 questions in one call. NN/g’s funnel guidance budgets 5–8 open-ended questions per interview guide, broad first and narrow later, with the rest held as probes.

The bank covers current workflow and tool switching. It also covers trust and hallucination tolerance, with questions on willingness to pay and data privacy objections.

What does their current AI workflow look like?

Start here because the baseline is messier than any procurement record suggests. The tool on the invoice is often not the tool doing the work, because people mix company seats with personal accounts.

1. Walk me through the last time you used an AI tool for work. What was the task, and what did you do with the output?
2. What happened right before you opened it? What were you trying to avoid doing by hand?
3. Which AI tools did you open last week, and which one did you open most?
4. For that task, did you use a company-provided tool or a personal account? Why that one?
5. Who else on your team touched the output before it shipped?
6. What did you change in the output before you used it, if anything?
7. What is the last task you tried in an AI tool and then gave up on?

When did they last switch AI tools?

Switching is more common and more painful than vendors admit. In a survey of 542 US decision-makers, 66% had attempted to migrate between AI platforms, and 58% of those said it failed or took far more effort than expected, while 44% keep multiple vendors to spread risk.

1. When did you last change which AI tool or model you use for a task? What triggered it?
2. What did that switch cost you in time, rework or lost context?
3. How many AI vendors does your team pay for today, and why more than one?
4. What else did you try before the tool you use now?
5. The last time a vendor added an AI feature to a product you already pay for, did you use it? What happened?
6. What would have to break for you to drop the tool you use most?

How do they catch AI mistakes?

Ask about the last error, not about trust in general, because self-reported trust and actual checking behaviour diverge. In the KPMG and University of Melbourne study of 32,352 employees, 66% had used AI output without evaluating it and 56% had made work mistakes because of AI.

1. Tell me about the last time an AI output was wrong in a way that mattered. How did you catch it?
2. What do you check before you send AI-generated work to someone else?
3. For which tasks do you accept the output as-is, and for which do you always rewrite?
4. When an AI tool gives you a number or a citation, what do you do with it?
5. Has an AI mistake ever reached a customer or an executive? Walk me through it.
6. How did your team’s checking process change after that?
7. Which vendor claims about accuracy have you tested yourself, and what did you find?

What have they actually spent on AI?

Never ask how much they would pay. Fitzpatrick’s warning is that a pricing number from a hypothetical question feels rigorous precisely because it is a number, and it is fluff. Ask about past spend decisions instead. The spread is wide: in G2’s 2025 survey of 1,169 buyers, 88% of AI power users would pay a premium for AI features against 31% of non-adopters, and only when the vendor showed value first.

1. What did your team spend on AI tools last quarter, and which budget did it come from?
2. The last time a vendor raised prices because of an AI feature, what did you do?
3. Which AI line item did you cut or negotiate down in the last year, and why?
4. Who signed off on your most recent AI purchase, and what did they ask you to show?
5. Seat, usage or credits: which model did you pick last time you had a choice, and what drove it?
6. What have you built in-house instead of buying, and what made building the better option?

What privacy objections have they raised?

Privacy objections are specific, so your questions should be too. Cisco’s 2025 benchmark of 2,600 security and privacy professionals recorded what people put into generative AI tools: employee information at 46%, non-public company information at 42% and customer information at 31%.

1. What information is your team not allowed to put into an AI tool, and how is that enforced?
2. Tell me about the last time security or legal reviewed an AI vendor. What did they ask for?
3. What did the vendor’s training-data policy say, and did it change the decision?
4. Have you pasted customer or company information into a personal AI account? What happened next?
5. Which certifications or documents did procurement require before approving the last AI tool?
6. What happened to the last deal where the vendor could not explain how the model used your data?

How do you sequence questions across a 30-minute discovery call?

Thirty minutes is enough for one full story plus two probes, and no published framework gives you a minute-by-minute split, so the budget below is mine. Story-based interviews often run 20 to 30 minutes. JTBD interviews run longer and Fitzpatrick’s shorter, so 30 minutes sits in the usable middle.

• Opening, 3 minutes: Set a casual tone and say there are no right answers. Treat the interview like filming a documentary of their decision, not an evaluation. Confirm recording and any AI note-taker, then ask question 1.
• Workflow, 10 minutes: Stay on the one story from question 1. Build the timeline with “what happened first” and “what happened next” rather than jumping to question 3. Use questions 2, 4 and 5 as probes only when the story stalls.
• Friction points, 8 minutes: Move to questions 7 and 14. When you hear “usually”, “would” or “might”, redirect to the last specific instance. This is where the quotable VoC lines come from.
• Willingness to pay, 5 minutes: Ask questions 21 and 24, both about past decisions. If the customer volunteers a future price, note it and ask what they paid for the closest thing they already bought.
• Close, 4 minutes: Ask question 13 or 30 if time allows, then ask who else on their team you should talk to and what you should have asked but did not. Fitzpatrick’s test is that a meeting succeeded when it ends with a commitment to a next step, even if that step is an intro to a colleague.

Cut the willingness-to-pay block before you cut friction. A 30-minute call that produces one detailed story about an AI failure that reached a customer is worth more than one that touches all five themes shallowly.

How do you use ChatGPT to generate an interview guide?

The prompt formula works because it pulls the interview rules from a context artifact rather than retyping them, so every guide your team generates obeys the same constraints.

Vendor prompt guides typically order a developer message as identity, instructions, examples and context, with reference documents wrapped in XML tags so the model knows where they begin and end. They also recommend a handful of examples and telling the model what to do rather than what to avoid. The formula below follows both.

The raw prompt, with the context artifact pasted in where the tags indicate:

<role>
You are a B2B customer discovery researcher writing an interview guide. You follow The Mom Test and story-based interviewing: every question asks about a specific past event, never a hypothetical, never an opinion of our product.
</role>

<audience_persona>
[paste the audience persona artifact: role, company stage, tools they run, vocabulary they use]
</audience_persona>

<voc_rules>
Never write questions that start with "Would you", "Do you think", "How much would you pay" or "How likely". Never pair two questions in one line. Never name a feature of our product. Never suggest an answer inside the question.
</voc_rules>

<examples>
<example>Walk me through the last time you used an AI tool for work. What was the task?</example>
<example>Tell me about the last time an AI output was wrong in a way that mattered. How did you catch it?</example>
<example>The last time a vendor raised prices because of an AI feature, what did you do?</example>
</examples>

<task>
Write a 30-minute discovery guide for the persona above on the topic of how they use AI tools for [topic]. Output: an opening line, eight numbered questions grouped under workflow, friction, spend and close, and two follow-up probes per question. Plain prose, no tables, no commentary.
</task>

Expect a first pass that looks like this (illustrative sample, shortened to the lines that need editing):

Opening: Thanks for joining. There are no right or wrong answers. I want to hear your story.

Workflow
1. Walk me through the last time you used an AI tool for a reporting task. What was the task?
2. How often do you use AI tools in a typical week, and which ones?

Friction
3. Do you find that AI outputs need a lot of editing before you can use them?
4. Tell me about the last time an AI output was wrong. How did you catch it, and what did you do next?

Spend
5. Would you be willing to pay more for a tool that was more accurate?
6. Who approved your most recent AI purchase, and what did they ask to see?

Three edits a researcher makes to that output, and the rule each one produces:

• Questions 2 and 3 need specific, neutral wording: In question 2, you ask for “A typical week”, inviting the “I usually” answer the guide is built to avoid. Rewrite as “Which AI tools did you open last week?” The rule: replace every “typical” or “often” with a dated window. In question 3, you suggest the answer: “Do you find that outputs need a lot of editing” tells the customer what you expect. Rewrite as “What did you change in the last output before you used it?” The rule: no yes/no question may contain the hypothesis.
• Question 5 is the hypothetical price question the rules banned: The model wrote it anyway because the ban sat in the prompt, not in an example. Delete it and add question 22 from the bank. The rule: a banned pattern needs a matching good example, not only a prohibition.

Move those three rules into the voc_rules block of your context artifact so your team can check each new guide against them. The prompt is one invocation point.

Your team stores the persona and rules outside the chat, alongside the example library, and reuses that durable layer across sessions. Five marketers can generate guides from the same shared context instead of relying on one-off prompts.

What does an AI-moderated customer interview look like?

An AI-moderated interview is your guide run by software that reads each answer and decides whether to probe, within limits you set before launch. The participant receives a link, opens it on their own schedule, and answers by text, voice or video depending on the platform.

The mechanics vary by platform, and the differences matter for B2B work:

• Researchers set follow-up depth before launch: In Outset, the researcher sets the probing depth and the AI asks follow-ups in real time. In Listen Labs, follow-ups are set per question as none, on short answers, one, or two to three.
• Session length varies by platform: Outset’s guidance is to aim for 20 to 30 minutes or less. Maze runs a live video call where the participant can interrupt the moderator, and the feature is Enterprise-only. Most platforms present a voice or text interface rather than an avatar, and participants complete the session before a deadline instead of booking a slot.

Independent evaluations are more cautious than the vendor pages. A January 2026 test of two AI interviewers with 10 participants found the moderators followed the script rather than the insight, never created, reordered or skipped a question, and only 3 of 10 participants said the conversation felt natural. Read those numbers as the current ceiling: an AI moderator will run your 32-question bank consistently, and it will miss the detour where a customer mentions the deal that collapsed over data policy.

How do AI-moderated interviews compare to human-led ones?

Researchers can reduce cost and turnaround using AI moderation, while human interviewers earn higher ratings on follow-up depth and trust. Use the comparison below to separate independent evidence from vendor data:

On follow-up depth and trust, human interviewers still win. On candour the picture is mixed, because some people share more when no one is watching. On cost and time to a first report, AI moderation is cheaper and faster. Treat vendor case studies that claim dozens of interviews in two weeks as vendor data, not as an independent ceiling.

The practical split for a growth team: run AI-moderated sessions when you need volume on a narrow, well-defined topic such as tool-switching triggers across 60 customers, and run human calls when the topic is the one you do not understand yet. No vendor publishes AI-versus-human completion rates, so budget for drop-off on voice-only AI sessions until you have your own numbers.

How do you analyse AI interview answers?

AI handles transcription and first-pass coding well enough to use, and it misses the latent themes that make VoC useful for positioning, so the workflow is AI first, human second, on every transcript. The worked example shows what AI coding captures and how a human researcher adds themes to a structured VoC entry.

Transcription error rates now sit low enough for English business calls on independent speech-to-text leaderboards. Product names and jargon are the weak point, so load competitor and feature names into the custom vocabulary before the first call.

Coding accuracy depends on how much inference the code requires. Models can match human agreement on descriptive tags and still miss interpersonal or inferential themes. Every quote the model pulls gets checked against the transcript. Adding explicit inclusion and exclusion boundaries to each code is the fix that moves inferential tags.

What does coding one transcript look like?

Take this excerpt from a discovery call with a marketing ops lead (illustrative, built from the question bank above, not a saved transcript):

“We moved the reporting workflow off the vendor’s assistant to a Claude project in March. Honestly the outputs were fine. It was more that I couldn’t see what it was doing with our pipeline data, and our security lead kept asking. So I rebuilt it where I could control the inputs. It took about two weeks of evenings. Now I rewrite maybe a third of what it gives me before it goes to my VP, which is still faster than the old way.”

A codebook prompt applied to that excerpt returns three codes, and all three are correct: tool switching with a March date, a data-visibility trigger tied to a security review, and partial rewriting of outputs before use. That is the descriptive layer, and the AI gets it on the first pass.

Two themes a human coder added:

• The stated reason and the operating reason diverge: The lead says outputs were fine and then describes rewriting a third of them before her VP sees them. The real constraint is reputational exposure to a senior reader. When interviewees state a claim first and give the example last, a coder often captures only the claim.
• The cost of switching was unpaid time: Two weeks of evenings is a willingness-to-pay signal, not a workflow detail. She spent personal time to avoid a vendor she could not audit. As a researcher, you can treat that time as a signal for data transparency rather than ask her to name a hypothetical premium.

The researcher adds both themes to the VoC artifact as structured entries, which the team supplies to downstream AI sessions. A compact entry looks like this:

Theme: data visibility as switching trigger
Verbatim from the ops lead: they could not see what the assistant did with pipeline data
Context: marketing ops lead, Series B SaaS, moved reporting off vendor assistant, March
Signal: rebuilt workflow on personal time rather than accept opaque data handling
Use: objection-handling copy, security page, pricing-page FAQ
Source: interview 07, 14:20

The VoC artifact is the durable layer. Researchers add entries after each coded interview, and marketers use the same file for every content session and positioning draft. The team builds on those themes instead of leaving them in one researcher’s notes.

Once those themes live in a shared VoC file, the next job is an AI content workflow every marketer writes from, then prompts that summarize interview transcripts without inventing quotes, and a case-study prompt that turns one transcript into a draft.

What are the 80/20 rule and the 5 C’s of interviewing?

The 80/20 rule in interviewing means the customer talks for roughly 80% of the call and you talk for 20%, and the 5 C’s is a hiring heuristic that transfers only loosely to discovery. Both come up in search results for this topic, so both get a short answer here.

The talk ratio has one explicit source. The US Digital Service discovery guide budgets the interviewer at about 20% of the airtime and the stakeholder closer to 80%.

Fitzpatrick and Torres both say listen more than you talk without giving a number, and the ratio has nothing to do with the Pareto principle of 80% of effects coming from 20% of causes.

On a 30-minute call, 20% is six minutes of your voice. Use that time for the opening and close, with probes in between and nothing left for pitching.

The 5 C’s has no standard version. The most repeated set is competency, character, communication skills, culture fit and career direction, which is a candidate-evaluation list.

If you want a customer-call equivalent, keep the C’s that describe the interviewer rather than the interviewee. Bring curiosity about the specific story and clarity to each question. Keep interviews consistent so you can code the transcripts against one codebook.

What if you meant hiring questions for AI roles?

Hiring loops for AI roles follow the same stages as general technical hiring, with one added round that tests judgment about when a model is guessing. A typical process runs resume review, intro calls, a skills assessment that varies by team, and final interviews with several people across one or two days. Ask every candidate the same questions and score them on one rubric. The competency questions below are paired with what a strong answer contains.

• LLM fundamentals: Ask “Explain what happens between a prompt and a response, and where the output can go wrong.” A strong answer covers tokenization, context limits, sampling, and names factual inconsistency as a property of the model rather than a bug to patch.
• RAG vs fine-tuning: Ask “When would you choose retrieval over fine-tuning?” A strong answer uses retrieval to add recent or external knowledge and fine-tuning to get better at a specific task, and flags the common mistake of insisting on fine-tuning when prompting already works.
• Bias vs variance: Ask “Your model scores well on training data and poorly in production. What do you check first?” A strong answer identifies high variance from a training error far below the test error and proposes more data or lower complexity.
• Guardrails: Ask “What is the difference between a guardrail and an evaluation?” A strong answer states that guardrails block clear failures before a user sees them while evaluators measure quality after the response, then gives an input and an output example, such as refusing database writes without human approval.
• Behavioural round for customer-facing AI roles: Ask “Tell me about a time an AI output reached a customer and was wrong.” A strong answer names the detection lag, the customer impact, the process change that followed, and what the candidate would now check before shipping, in STAR order without being prompted for it.
• AI product sense: Ask “Here is a feature spec. Where will the model guess, and what is the minimum quality you would ship at?” The strong answer maps failure modes and defines a quality bar with an eval, which matches the product-sense round Meta added for AI roles, judged on noticing when the model is guessing.

For customer-facing roles such as AI solutions consultants or support leads, weight the behavioural and guardrail rounds over bias-variance. The person explaining a hallucination to an enterprise buyer needs the judgment questions more than the derivation.

The interviews are the input. The system that turns 12 transcripts into a VoC artifact every marketer on your team writes from is the part that compounds, and it is the part most teams never build because each researcher keeps their own notes. The weekly operator notes on that system land in The Messy Middle newsletter.