To price an AI feature with no benchmark, interview 20 to 30 of the people who would pay for it before you launch. Ask what they spend today on the job the feature does, which unit of value they would accept being billed on (seats, usage, outcomes or a flat add-on), and where the price starts to need sign-off. Do not ask "what would you pay?" Run those conversations as AI-moderated interviews and you can have them done in a week, with a report that shows which pricing model your buyers will accept and why, before the price goes live and gets hard to change.
Why AI feature pricing is harder than normal pricing
Pricing a normal feature is mostly a packaging question: which tier does it sit in? An AI feature adds two problems. Each use costs you money in inference, and buyers have no settled sense of what the thing is worth. Recent data shows how unsettled the market is.
- Nobody has picked a model yet. ICONIQ's 2026 State of AI report found that subscription and platform fees are still the most common way to charge for AI (57% of companies). Consumption-based pricing rose from 35% to 42% in about six months, and outcome-based pricing from 18% to 23%. Companies now combine 1.7 pricing models on average, and 37% plan to change their AI pricing model in the next 12 months.
- Buyers want AI priced differently from the rest of your product. Futurum's 1H 2026 software buyer research found 42.9% of buyers prefer consumption pricing for GenAI features, while preference for consumption pricing on core software fell to 30.1%. The same customer may want your seat licence and your AI add-on billed in two different ways.
- Most AI launches are not moving revenue much. Simon-Kucher reports that 96% of companies plan to launch new AI features in the next two years, yet for most the revenue impact stays modest, typically below 10%.
Put those together. Everyone is shipping AI, few are getting paid well for it, and a third of companies expect to change how they charge within a year. A PM who launches at a guessed price is likely to reprice once, in public, and repricing an AI feature after launch costs more than a week of interviews up front.
What you are really trying to learn
A willingness-to-pay study for an AI feature has to answer four questions. None of them is "what is the right number?"
- What is the value unit? Does the buyer think of the feature as more seats' worth of work, as a number of tasks done, or as an outcome such as a resolved ticket or a qualified lead? The unit they already use in their head is the one they will accept on an invoice.
- What does the job cost today? Staff hours, an agency, another tool. Money already being spent tells you a budget exists, and roughly how big it is.
- How much unpredictability can they live with? Usage pricing sounds fair until a finance team cannot forecast it. Ask how they would explain a variable bill internally, and whether they would want a cap.
- Who signs off, and at what amount? A PM who loves the feature and a procurement team that will not approve it add up to no sale. Find the amount where the purchase stops being something one person can decide.
Questions to ask (and the ones to avoid)
Anchor every question to what people did, not what they imagine.
- "Walk me through the last time you did [the job the AI feature does]. Who did it, and how long did it take?"
- "What do you spend on that today, in tools, services or staff time?"
- "If this were billed per [task / report / resolved ticket], how would you estimate what you'd use in a month?"
- "Would a bill that changes month to month be a problem for you? What would make it acceptable?" (probe: caps, prepaid credits, an annual commitment)
- "Which would be easier to get approved: a fixed add-on per seat or a usage-based charge? Why?" (single choice, then probe)
- "At what monthly amount would this need someone else to sign off?"
- Close with the four Van Westendorp questions (too cheap to trust, a bargain, getting expensive, too expensive), asked about the feature as you described it.
Avoid "How much would you pay for this?" and "Would you pay $X?" Both produce numbers people do not stand behind. Also avoid naming your planned price. It anchors every answer that follows.
How many interviews
For a single buyer segment, 15 to 20 interviews usually show the same value units and objections repeating. If you sell to two segments that budget differently (for example, small teams paying by card and larger companies going through procurement), plan 12 to 15 per segment. That is 25 to 30 conversations, which is a lot of calendar time for one PM and very little for an AI moderator running them in parallel.
Mistakes that ruin pricing research
- Interviewing only fans. Your power users will tolerate almost any price. Include people on cheaper plans and people who evaluated you and left.
- Treating "it depends on value" as an answer. Ask what value they would measure, and how.
- Ignoring cost to serve. Interviews tell you what people will accept. Your inference cost tells you what you can afford. A usage price below your cost per task is a pricing error, however popular it is.
- Pricing once. Re-run a short version after three months of real usage, when buyers can talk about actual bills.
How to run it in Koji
You can do this on any plan with text interviews. Voice interviews need the Interviews or Enterprise plan.
- Start from the Willingness-to-Pay Research template. It is built on the rules above. It anchors to current spend, never asks directly what someone would pay, maps who approves the purchase, and closes with a verbal Van Westendorp exercise. Rename it for your AI feature and change the problem statement to describe the job the feature does.
- Fill in the brief. Set the decision ("launch as a per-seat add-on, usage-based credits or a hybrid") and your current hypothesis. If you have a one-page description of the feature or a pricing draft for internal use, attach it in the Studio chat so the interviewer understands the product before it asks anything. Keep the price itself out of the interview guide.
- Add structured questions where you need numbers. Koji supports open-ended, scale, single choice, multiple choice, ranking and yes/no questions, each with its own follow-up depth. Use a single-choice question for "seat add-on or usage-based", a ranking for the value units you are considering, and a scale for how much bill variability they can tolerate. Each one becomes a chart in the report, and the interviewer still asks why. See the structured questions guide.
- Screen for the people who hold budget. A screening question such as "Who approves software purchases in your team?" lets you turn away people who cannot speak to cost. Follow the screening questions guide.
- Preview before you invite. Take the interview yourself on the Preview tab, by text or voice. Previews are free and never appear in the report. Listen for anything that hints at your price.
- Invite customers and prospects. Send the link to customers, or import a list with personal invite links on the Interviews and Enterprise plans. To hear from buyers who have never used your product, panel recruitment lets you build an audience and see the cost on any plan; launching a recruitment needs a paid plan.
- Read the report. It opens with what you learned, then the reasons behind it, then what it means for your decision. Click any finding to read the interviews behind it. The responses grid puts every structured answer in a table you can export as CSV, which is handy for plotting Van Westendorp answers in a spreadsheet.
What it costs
A text interview is 1 credit and a voice interview is 3, and credits cost €1 on paid plans. Free includes a one-time grant of 10 credits. Insights is €29 a month with 29 credits, and Interviews is €79 a month with 79 credits, so a 30-interview text study fits inside the Interviews plan. Interviews that score under 3 out of 5 on quality do not use credits and are kept out of the report, so a respondent who clicks through without saying much costs you nothing. Plans change, so check the pricing page before you budget.
What you walk into the pricing review with
A good AI pricing study gives a PM four things they can defend:
- The value unit buyers already use. For example: "Support leads think in resolved tickets, not seats. Eleven of 14 described the job that way."
- A tolerance range for variable bills. For example: "Usage pricing was acceptable to most, but only with a monthly cap or prepaid credits."
- An approval threshold. The monthly amount where the purchase moves from the team lead's card to procurement, per segment.
- A price corridor. The Van Westendorp answers give a range where the feature is neither suspiciously cheap nor too expensive. Combine it with your cost per task to set the floor.
That turns the pricing meeting from opinions about competitors into a decision about evidence. For the general method, the guides to pricing research interviews, the Van Westendorp price sensitivity meter and the Gabor-Granger method go deeper. If you plan to raise prices on an existing plan to fund the AI work, read the price increase research guide too.
Why Koji for AI feature pricing research
| What a PM weighs | Pricing survey | Pricing consultant | Calls you run yourself | Koji |
|---|---|---|---|---|
| Finds the value unit | Only if you guessed it in the options | Yes | Yes | Yes, with follow-up on every answer |
| Time to results | Days | Weeks to months | Weeks of scheduling | Days |
| Reaches 25 to 30 buyers | Yes | Usually fewer | Rarely | Yes, in parallel |
| Avoids anchoring on your price | Depends on design | Yes | Hard when you built it | Guardrails keep your price out |
| Numbers and reasons together | Numbers only | Both | Reasons only | Charts plus the conversation behind them |
| Cost | Tool cost | Five figures or more | Your time | From €29 a month |
Koji is not a conjoint tool. If you need a statistically modelled price for a large consumer market, run a conjoint or Gabor-Granger survey as well. Interviews tell you which models and units to put in that survey, so it measures the right thing. Our roundup of the best pricing research tools compares the options.
Start before the price is public
Create a study from the Willingness-to-Pay Research template, swap in your AI feature, add the budget-holder screening question and preview it. Once your first five interviews are in, read them and adjust the brief, then let the rest run. You can start free with 10 credits, or see Koji for product managers for the other studies PMs run. For related reading, see the AI feature adoption playbook, pricing research without a consultant and pricing survey questions.
