--- title: "How to Use AI for Pricing Analysis" description: "Test a price with AI without trusting stated willingness to pay: the sourcing rule for competitor prices, and the billing replay that decides if you can ship." canonical_url: "https://builderscamp.com/guides/tools/ai-for-pricing-analysis" date_published: "2026-09-18" date_modified: "2026-09-18" author: "Andre Albuquerque, Mário Araújo" publisher: "Builders Camp" guide_class: "tools" --- # How to use AI for pricing analysis without trusting what people say they will pay **TL;DR:** The highest-value pricing analysis AI can run is not a survey, it is the billing replay: recompute every historical invoice under the proposed price and count how many customers move more than 20 percent. Stated willingness to pay is systematically generous because saying a number costs nothing, so anchor research on the last comparable purchase a buyer actually made. Every competitor price in your file carries their own published page and the date you read it, or it is recorded as not public. ## Which pricing question are you actually asking? Three questions get called pricing and they need different evidence. Setting a price for something new, where you have no history and are reasoning from comparable purchases. Moving the boundary between existing plans, where you have complete history and almost no need for a survey. And raising prices on an existing base, where the analysis is about who leaves, not about what the product is worth. The second and third are where most teams actually are, and they are the two where AI is most useful, because both are answerable from data you already own. The first is the one everyone reaches for research about, and it is the one where research is weakest. ## Stated willingness to pay is systematically generous Saying a number in a survey costs the respondent nothing. Paying it costs them money, requires a budget line, and in a business purchase requires convincing someone else. Those are three different gates, and a survey clears none of them, which is why a stated number and a signed contract diverge in a predictable direction. The correction is to move the question from the hypothetical to the historical. What is the last comparable thing you bought, what did it cost, which budget did it come from, and who approved it? Those four answers are checkable, they are about a decision the person actually made, and they tell you the ceiling and the approval path in one go. A model is useful here for coding a hundred such answers into budget bands and approval patterns, which is a real job. It is not useful for producing the answers. ## What Van Westendorp and conjoint actually give you The Van Westendorp price sensitivity meter asks four questions about the point at which a price becomes too cheap to trust, cheap, expensive, and too expensive, and intersects the resulting curves to produce an acceptable range. That range is a useful guard rail and it is not a recommendation: it says nothing about how many people buy at each point, and it assumes respondents understand a product they may never have used. Conjoint analysis goes further by forcing trade-offs between bundles rather than asking about price in isolation, which is closer to how a real decision feels, at the cost of a much larger sample and a design that is easy to get wrong. Both are genuine methods with known limits. A model can help draft the instrument, check that your attribute levels are balanced, and code the open-text answers. It cannot make a 40-person sample behave like a 400-person one. ## The rule for every competitor price in your file Each competitor price carries two things: a link to that company's own published pricing page, and the date you read it. PostHog, for example, publishes its full price list openly, which makes a quotable row possible; plenty of companies publish nothing, and in that case the honest entry is "not public." Never fill that cell with an estimate. A range written in grey italics with a question mark becomes a number in the next deck and a fact in the one after, and nobody will be able to say where it came from. The collection discipline is the same one described in [how to use AI for competitive analysis](https://builderscamp.com/guides/tools/ai-for-competitive-analysis): a quote, a URL, a date, or the row is deleted. ## The billing replay is the analysis that actually decides things Take the proposed price, recompute every invoice you issued last year under it, and produce one table: how many customers pay less, how many pay the same, how many pay more, and the distribution of the increase. That table is where a pricing change lives or dies, and almost nobody builds it before the decision. A packaging change that looks elegant in a slide can mean a bill that more than doubles for a slice of the existing base, and that slice is often your longest-tenured accounts, because they accumulated usage under the old metric. Knowing which accounts those are before you ship converts the problem from a pricing debate into a migration plan with grandfathering rules and a communications sequence. This is also the piece of pricing work where a model is unambiguously good. It is deterministic arithmetic over your own billing export, every step is checkable, and the answer does not depend on anyone's belief about demand. ## Three checks before you trust the replay - **Reconcile the recomputed total against last year's actual revenue.** If replaying the old price against old invoices does not reproduce the real number, the model of your pricing is wrong and the new-price output is meaningless. - **Separate list price from realised price.** Discounts, annual commitments and legacy terms mean the price on the page is not the price on the invoice, and replays built on list price overstate the increase. - **Check the customers who fall off the edges.** The largest and smallest accounts usually break a new metric in opposite directions, and both ends need a named rule rather than an average. The first check is the one to run before anything else, because it is the only one that validates the whole exercise rather than one assumption inside it. ## What AI should not be deciding Elasticity from observational data is the main trap. Your historical prices correlate with your historical sales motion, your segment mix and your product maturity, so a model fitted to that history will confidently attribute to price what belongs to who you were selling to. A price test on comparable segments answers the question; a regression on the past does not. The second is anything with a legal dimension. Aligning prices with a competitor, and the treatment of discounts and published prices in a given market, are questions for counsel rather than for an analysis tool, and no research technique changes that. Keep the file factual and route the rest. ## What a price change is really measuring A pricing exercise that never touches the value metric is a rounding exercise. The deeper question is what you charge for at all, because the metric decides who grows with you and who is punished for adopting the product more. [Pricing strategy](https://builderscamp.com/guides/glossary/pricing-strategy), [monetization model](https://builderscamp.com/guides/glossary/monetization-model) and [product packaging](https://builderscamp.com/guides/glossary/product-packaging) carry the vocabulary, and [customer lifetime value](https://builderscamp.com/guides/glossary/customer-lifetime-value) is the number a metric change moves first. ## Replay last year before you argue about the number Before the next pricing meeting, run last year's invoices through the proposed model and bring the distribution rather than the price. The argument changes shape when everyone can see, by name, which accounts the decision actually lands on. [See the Business for Product Managers bootcamp](https://builderscamp.com/bootcamps/business-for-product-managers?utm_source=guide&utm_medium=organic&utm_campaign=ai-for-pricing-analysis) Business for Product Managers runs one week with 2 live sessions and 8 microlessons, covering revenue, costs, unit economics and the profit and loss thinking a pricing decision sits inside. Product Strategy covers the packaging choice the price expresses, across two weeks, and Growth for Product Managers covers monetization alongside acquisition and retention for teams who own the revenue number directly. ## Frequently asked questions ### Can AI tell me what to charge? No, and the places it sounds most confident are the places to trust it least. What it can do is arithmetic you would otherwise skip: replaying a proposed price against every historical invoice, coding hundreds of pricing objections from sales calls, and modelling how a change in the billing metric redistributes cost across your existing base. ### Why is stated willingness to pay unreliable? Because saying a number costs the respondent nothing, while paying it costs them money. A survey answer about a hypothetical purchase is a statement about how the person wants to be seen, filtered through a product they have imagined rather than used. Anchor the conversation on the last comparable thing they actually bought and what they paid. ### What does the Van Westendorp method actually produce? An acceptable price range, derived from four questions about when a price becomes too cheap, cheap, expensive and too expensive. It does not produce a revenue-maximising price, it assumes the respondent understands the product well enough to judge it, and it says nothing about volume at each point. Treat the output as a sanity band, not a recommendation. ### How should a competitor's price be recorded? With the competitor's own published page as the source and the date you read it. Some companies publish a full price list openly and some publish nothing; where nothing is published, the honest row is not public. An estimate in that cell will be quoted back as a fact within two meetings. ### What is the billing replay? Recomputing every historical invoice under the proposed price and counting how many customers move, and by how much. It is the single most decision-relevant analysis in a pricing change and it needs no survey, no model of demand and no assumptions: the data is your own billing history. ### Can AI estimate price elasticity from our data? Rarely in a way you should act on. Observational pricing data is confounded by everything that changed alongside price, including who you were selling to at the time. Elasticity claims need either a deliberate price test on comparable segments or a natural experiment you can defend, and most teams have neither. ### Which Builders Camp bootcamp covers pricing? Business for Product Managers covers the business intuition underneath pricing decisions, revenue, costs, unit economics and profit and loss thinking, across one week with 2 live sessions and 8 microlessons. Product Strategy covers the packaging choice a price expresses, across two weeks with 4 live sessions and 8 microlessons. ## Sources - [Wikipedia: Van Westendorp's Price Sensitivity Meter](https://en.wikipedia.org/wiki/Van_Westendorp%27s_Price_Sensitivity_Meter) - [Wikipedia: Conjoint analysis](https://en.wikipedia.org/wiki/Conjoint_analysis) - [PostHog: public pricing page](https://posthog.com/pricing) - [Builders Camp: Business for Product Managers bootcamp](https://builderscamp.com/bootcamps/business-for-product-managers) - [Builders Camp: Product Strategy bootcamp](https://builderscamp.com/bootcamps/product-strategy) ## How this guide was made Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.