AI Pricing Strategy: Test What Customers Will Pay

intermediate8 min read · updated 2026-06-20

AI Pricing Strategy: Test What Customers Will Pay

Most founders pick a price the way they pick a font: a few minutes of gut feel, then never touch it again. That is a mistake everywhere, and it is a bigger mistake with AI products, where your cost-to-serve is a moving target and your buyer's sense of "what this is worth" hasn't settled yet. Companies spend roughly 6 hours a year on pricing on average, while a 1% improvement in pricing can lift profit by about 11% — a bigger lever than equivalent gains in acquisition or retention.

This entry is a practical playbook for pricing an AI product: how to find willingness to pay, which model to charge on, and how to protect your margin when the cost floor keeps shifting.

Why AI pricing is its own problem

Three things make AI pricing different from classic SaaS:

• Variable cost-to-serve. Every inference call costs you tokens. A power user can be 50x more expensive than a casual one, so a flat monthly seat price quietly turns your best customers into your worst margins.
• A falling cost floor. Foundation model inference prices have been dropping fast — roughly 75% a year by some accounts. A price that looks tight today may have 60% gross margin baked in six months from now. Don't anchor your price to today's cost.
• Unsettled willingness to pay. Buyers don't yet have a reference price for "an AI that does X." That is a gift — you get to set the anchor — but only if you actually test instead of guessing.

The market has responded by moving away from pure flat subscriptions. Hybrid models — a predictable base fee plus a usage or outcome component — are now common, with around 31% of AI vendors using a hybrid approach. And pricing here is genuinely experimental: roughly 92% of AI companies that launched with usage-based pricing later changed it at least once. Expect to iterate. Build for it.

Step 1: Stop guessing — measure willingness to pay

The single highest-leverage thing you can do is run a Van Westendorp Price Sensitivity Meter before you lock a price. It's a four-question survey you can run on 50-200 prospects for the cost of a gift-card incentive. Ask each respondent, about one clearly-scoped unit (e.g., "per seat per month"):

• Too cheap: At what price would the product be so inexpensive you'd question its quality?
• Cheap / bargain: At what price would it be a great buy for the money?
• Getting expensive: At what price would it start to feel expensive, but still worth considering?
• Too expensive: At what price would it be so expensive you wouldn't consider it?

Plot the four cumulative curves and read off the intersections:

• Optimal Price Point (OPP) — where "too cheap" crosses "too expensive." Your headline anchor.
• Indifference Price Point (IPP) — where "cheap" crosses "expensive." Often near "just right."
• Range of Acceptable Prices — from the Point of Marginal Cheapness to the Point of Marginal Expensiveness. Your tiering corridor lives here.

Two honest caveats: respondents lowball, and stated intent isn't behavior, so treat Van Westendorp as directional, not gospel. Triangulate it with a real-money test (Step 4) and, if you can, a conjoint or A/B test. See the framework detail at Umbrex's Van Westendorp guide.

Step 2: Pick the metering model that matches value

Choose what you charge on, not just how much:

• Per-seat subscription — simple, predictable, easy to forecast. Bad fit when value and cost both scale with usage rather than headcount. Best for collaboration-style AI tools.
• Usage / consumption-based — charge per call, per token, per document, per minute. Aligns price with both value delivered and your cost-to-serve. The downside is buyer anxiety: unpredictable bills kill enterprise budgeting.
• Hybrid (base + usage) — a committed platform fee for predictability plus metered overage. This is where most AI products are converging because it protects margin without scaring off the CFO.
• Outcome-based — charge per resolved ticket, per booked meeting, per qualified lead. The strongest value alignment and the hardest to instrument and attribute. Reserve it for cases where you can prove the outcome was yours.

Rule of thumb: the metering unit should be something the customer already counts as a win (a closed deal, a finished draft, a resolved case) — never something they count as a cost (compute, tokens).

Step 3: Build tiers around the value, not the cost

Use your Van Westendorp range to set three tiers. Anchor the middle tier near the OPP, make a deliberately overpriced top tier to anchor perception, and put a real (not crippled) entry tier at the bottom. Gate tiers on value metrics customers understand (seats, projects, monthly resolved tasks) rather than on raw infrastructure (GB, tokens) that means nothing to them. Make the jump between tiers feel justified by a capability they'll grow into.

Step 4: Test with real money, fast

Surveys tell you intent; checkout tells you truth. Run cheap, reversible experiments:

• Price A/B test — split new traffic across two price points on the same offer; measure conversion and revenue per visitor, not just conversion.
• Fake-door / pre-order — show the price, capture intent or a deposit, measure how many proceed before you've built the tier.
• Sales-assisted discovery — for higher-ticket B2B, let reps quote a range and log where deals stall and where they close without negotiation (a sign you're underpriced).
• Grandfather and migrate — when you raise prices, protect existing customers short-term but migrate them deliberately; clean tier alignment is where the LTV gains actually compound.

Step 5: Re-price on a schedule, not on a crisis

Because your cost floor falls and your perceived value rises as the product matures, a price set at launch is almost always wrong within a year. Put a quarterly pricing review on the calendar. Each review: pull cohort gross-margin-per-account, find the tier where customers cluster against the ceiling, and re-run a lightweight willingness-to-pay check on new prospects. Treat pricing as a product surface you ship — because the data shows nearly everyone in AI does end up changing it.

Apply it this week

• Run a Van Westendorp survey on 50+ qualified prospects before you publish a price.
• Switch your metering unit to a customer win, not a cost input.
• Stand up a two-price A/B test on new signups and let revenue-per-visitor decide.
• Don't anchor price to today's inference cost — it's going to drop.
• Calendar a quarterly re-price review. Pricing is the highest-ROI hour on your roadmap.