Microsoft, Google, OpenAI and Anthropic all publish documentation telling you not to make pricing decisions with their tools. That is a better starting point than any vendor pitch.
The famous claim that a 1% price improvement lifts operating profit by 11.1% is real, but it comes from a 1992 sensitivity calculation on aggregate accounting averages that assumes zero volume response by construction. The same article contains a far more useful finding that almost nobody quotes: the gap between invoice price and what companies actually banked ran from 16.7% to 28.9%, and identical products sold across price bands as wide as 500%. Rebuilding that picture from your own invoices, credit notes and rebates is the single best use of AI in pricing, because it needs no market knowledge, no forecasting and carries no legal exposure. What you must not do is let one chat window both find competitor prices and do arithmetic on them, or let anything a competitor told you privately anywhere near the process.
Somebody in your last strategy meeting probably said that a 1% price rise is worth more than a 1% cost cut. They were quoting Michael Marn and Robert Rosiello, writing in Harvard Business Review in 1992. Their numbers: a 1% improvement in price lifts operating profit by 11.1%, against 7.8% for variable cost, 3.3% for volume and 2.3% for fixed cost.[1]
From Marn and Rosiello, Harvard Business Review, 1992. The exhibit footnote states it is based on the average economics of 2,463 companies in Compustat aggregate, and the price figure assumes no loss of volume.[1]
Read the footnote though. That exhibit is “based on average economics of 2,463 companies in Compustat aggregate”, and the price row explicitly assumes no loss of volume.[1] It’s a static sensitivity calculation on 1992 accounting averages, not a study of firms that raised prices and saw what happened. Quoting it as proof that raising prices works is a category error, and it’s exactly the sort of error an AI will help you make faster.
The line from the same paper that deserves to be on the slide instead: a 1% price decrease destroys 11.1% of operating profit dollars, and a 5% discount on a transaction wipes out 60% of the operating profit on that transaction.[1] Most businesses I work with are not underpricing by design. They’re leaking.
Worth knowing too: prices move less often than people assume. Analysis of US consumer price microdata found a median duration of 4.6 months for identical items including sales, stretching to eight to eleven months once you strip sales out.[2] A survey of more than 11,000 euro-area firms found the vast majority review prices a maximum of three times a year, and that the median firm changes its price once a year.[3]
Hold that thought, because it reframes what AI is for here. Those firms were already reviewing more often than they were acting. Adding a faster review to a business that doesn’t act on reviews changes nothing at all.
The actual thesis of that 1992 paper has almost nothing to do with raising list prices. It’s about the money that vanishes between the invoice and the bank.
Marn and Rosiello measured the drop from invoice price to what they called pocket price across client cases: 16.7% at a consumer packaged goods company, 17.7% at a commodity chemical company, 18.6% at a computer company, 20.3% at a footwear company, 21.9% at an automobile manufacturer, and 28.9% at a lighting products supplier.[1] The spread between the highest and lowest price actually received for one identical product ran to 60% at a lighting fixtures manufacturer, 200% at a specialty chemicals company, and 500% at a fastener supplier.[1]
Their explanation of why nobody notices is the useful bit. The leakage sits in payment terms, co-op advertising, volume rebates, freight and settlement discounts, items that get buried in interest expense accounts and collected on a companywide basis rather than per transaction.[1]
A few practical notes from doing this with people. Ask it to show the calculation for two or three rows in full so you can check the logic before you trust the other four thousand. Ask what it had to assume when a field was blank, because it will have assumed something and it won’t volunteer it. And ask which deductions it couldn’t allocate to a transaction, since that residual is usually where the interesting money is hiding.
If spreadsheets are the bottleneck rather than the pricing thinking, we’ve covered the mechanics separately in how to use AI to analyse a spreadsheet and using ChatGPT for Excel.
This is where I see the most damage, and the cause is structural rather than a bad prompt.
A language model cannot know today’s prices unless it goes and looks. Anthropic publishes reliable knowledge cutoffs for its models, and every major vendor does the same.[9] Anything after that date simply isn’t in there. Browsing is a separate capability that has to be switched on or invoked. Google’s Sheets documentation is explicit that to get answers from the web you must include a phrase like “Use Google Search” in your prompt.[13]
Then the second half of the problem. OpenAI’s documentation states plainly that the Python environment used for data analysis “cannot make external web requests or API calls”, and that if analysis depends on external data you must upload it or connect a source first.[10]
So the sandbox that does the maths can’t browse, and the browsing turn isn’t doing careful maths. Ask one chat window to both find eleven competitors’ prices and build you a positioning grid, and you get a plausible-looking table where some numbers were fetched this morning and others were remembered from 2024, presented in exactly the same font with exactly the same confidence.
One more habit worth building. Before you rely on any pricing prompt, run it on a case where you already know the answer, then check whether it got it right. Our guide on testing AI prompts before you trust the output covers doing that properly, and for the broader research workflow there’s using AI for competitor research.
AI has made survey design nearly free, which sounds like good news for pricing and mostly isn’t, because the hard part was never writing the questions.
People overstate what they’d pay. A meta-analysis of 77 studies across 47 papers, covering more than 24,000 observations of hypothetical willingness to pay and nearly 21,000 of real willingness to pay, put the average hypothetical bias at 21%.[17] The same work found something counterintuitive: in the authors’ modelling, indirect measurement methods overstated real willingness to pay more than direct ones did, at roughly 19% against 9%, rising to about 40% against 28% for specialty goods.[17] That runs against what most people assume, which is that clever indirect questioning gets you closer to the truth.
You’ll also hear that people overstate by a factor of two or three. That one’s worth correcting, because it’s a mean being dragged around by a handful of outliers. A meta-analysis of 28 stated-preference studies found a median ratio of hypothetical to actual value of 1.35, with half the calibration factors between 0.85 and 1.50 and 70% below 2. The top ten observations averaged 10.3 against 1.54 for the other 73.[18]
So: yes, adjust downward, but 35% is a better default than 200%, and the honest answer is that it varies enormously by what you’re selling.
It genuinely helps with drafting the Van Westendorp price sensitivity questions consistently, tabulating open-text responses about value, and pulling the language customers use about cost out of your support tickets and sales call notes. That last one is underrated, because the phrasing customers use when they flinch at a price is more useful than the number they give you on a survey.
It also introduces a bias most people never see. Ask AI to design a price sensitivity study and it will quietly make design choices that move your answer: how many price points to show, where to set the top and bottom of the range, whether to ask about your product alone or against alternatives. None of those arrive labelled as judgement calls. They arrive as a finished study design.
So make it show its working. Ask what it chose and why, ask what a different choice would have done to the result, and ask which of its decisions a pricing researcher would argue with. You’ll get a much more honest picture, and occasionally you’ll catch a design that was going to hand you the answer you wanted.
If you’re building the wider research picture rather than just the price question, using AI for market research covers the surrounding work.
Ask any chatbot for pricing psychology tactics and you’ll get charm pricing, anchoring and the decoy effect, delivered as settled science. The evidence is considerably shakier than that, and this makes a useful live demonstration of how these tools fail.
Take the $9 price ending. Anderson and Simester ran three field experiments with national mail-order clothing catalogues, randomised by postcode, the largest covering 270,000 customers across 308 items. The pilot on four dresses showed a demand increase of roughly 40%. Study 1 came in around 35%. Study 2 was about 15%. Study 3, the biggest, was 7%.[14] The effect is real and it shrinks sharply as the sample grows. It also concentrated on new items: in Study 2 the estimated increase was 22% for new items against 10% for established ones, and the established-item coefficient wasn’t statistically significant.[14]
Now the decoy effect, the one every pricing deck uses. Frederick, Lee and Baskin ran 38 studies. They found significant attraction effects in four of five cases using highly abstract stimuli, two of five in mixed cases, and in the remaining 27 studies where at least one attribute could be directly experienced, they found “no instances of a significant attraction effect”. Their verdict: “The boundary conditions for the effect appear so restrictive that one should question its practical validity.”[15]
The bit I find genuinely disarming is that the original authors agree. Huber, Payne and Puto, who introduced the effect in 1982, wrote in 2014: “We suspect that the asymmetric dominance effect occurs rarely in the marketplace today”, and noted that their original article “was designed as a demonstration study” and that they “did not set out to suggest a tool for marketing practice.” Huber also disclosed testing it on a commercial dataset of 586 respondents across nearly 4,000 qualifying choice sets and finding no consistent increase in the target’s share.[16]
| The claim you will be given | What the research actually shows |
|---|---|
| “Prices ending in 9 lift demand around 40%” | 40% was the four-dress pilot. The 270,000-customer study found 7%, and the effect on established items was not statistically significant[14] |
| “The decoy effect reliably shifts choice” | Significant in 6 of 38 studies, and in none of the 27 where people could see or taste the product[15] |
| “It is a proven pricing tool” | The original 1982 authors say it “occurs rarely in the marketplace today” and was never meant as a practice tool[16] |
| “Customers overstate willingness to pay by 2 to 3 times” | Median ratio across 28 studies is 1.35; the mean is dragged up by about ten outliers[18] |
Each row compares the version an AI tool will typically produce with the primary research it is drawn from. The pattern is consistent: models return the most-repeated version of a claim, not the best-evidenced one.
Try it yourself. Ask for the evidence on the decoy effect, then hold the answer against six out of thirty-eight. It’s the clearest demonstration I know of why these tools need checking, and it lands harder than any abstract warning about hallucination because it’s a topic people already believe they understand.
Here’s the part that turns an interesting exercise into a serious one, and the thing to understand is that regulators are not policing whether you used software. They’re policing what went into it.
Read the US Department of Justice’s November 2025 proposed settlement with RealPage as a specification. It requires the company to stop having its software use competitors’ nonpublic, competitively sensitive information to determine rental prices at runtime, to limit model training to backward-looking nonpublic data aged at least 12 months, to stop determining geographic effects narrower than state level, to remove features that limited price decreases or aligned pricing between competing users, and to accept a court-appointed monitor.[6] Every clause is about inputs.
The UK’s Competition and Markets Authority says the same thing in one sentence, and it’s the sentence to remember: make sure any pricing guidance generated by a pricing solution you use is “not influenced in any way by competitively sensitive information from rivals, even if you do not receive this information directly.” They add that if you can reasonably expect a recommendation could be drawing on a competitor’s confidential information, you may still be breaking the law.[7]
Which means the practical test for a small business has nothing to do with algorithms at all. If you paste a competitor’s confidential rate card into a chatbot to help set your prices, the chatbot is irrelevant to the offence.
Three more things from the CMA worth knowing before anyone gets clever:
Separately, if you’re thinking about personalising prices to individuals, the US Federal Trade Commission’s staff study on surveillance pricing found that behaviours ranging from mouse movements on a page to items left in a cart can be tracked and used to tailor pricing, and that the intermediaries involved worked with at least 250 clients.[4] The vendors’ own claims to the FTC were revenue growth of 2% to 5% and margin increases of 1% to 4%, though those are marketing claims made in documents produced to the regulator, not verified outcomes, and the FTC says explicitly that its study makes no assessment of whether anyone acted illegally.[5]
Five steps. None of them ends with AI producing the number.
Invoices, credit notes, rebates, freight, settlement discounts. Per customer, per product, as a percentage of list. This is the step that pays for the whole exercise and it uses only your own data.
Output the price, the source URL and the date observed. Nothing else. Public prices only, and nothing a competitor has told you privately, ever.
Paste the gathered numbers in as data. Let the analysis tool compute on a file you can see, so that every figure has a visible origin.
This is the highest-return use of AI language ability in pricing. Given the leakage you found in step one, ask it to write the rules: who can approve what, at what volume, with what exchange of value. Discount policy is a writing problem, and writing problems are what these tools are for.
Every vendor tells you this in their own documentation. Microsoft: “Avoid using Copilot for decisions in sensitive areas such as finance, legal, or medical topics.”[11] Anthropic says Claude for Excel is not recommended for audit-critical calculations without verification.[12] Google says don’t rely on Gemini features as financial or other professional advice.[13] That is unusually strong material, because it comes from the vendors themselves rather than from a critic, and it sits in the same documents that sell you the feature.
For the pieces around this, we’ve covered building a budget forecast with AI and using ChatGPT for financial analysis.
The number arrives without a source. The most common failure isn’t a wildly wrong price, it’s a table where three cells came from live browsing and two came from the model’s memory of 2024, and nothing distinguishes them. If a figure in your pricing analysis has no URL and no date attached, treat it as unverified regardless of how confident the sentence around it sounded.
A faster review that nobody acts on. Remember the euro-area finding: 57% of firms reviewed prices at most three times a year, but 86% changed them less than quarterly.[3] Reviewing was never the constraint. If AI just adds another analysis nobody has the nerve to act on, you’ve bought a more detailed picture of the same problem.
Competitor data walks in sideways. Nobody sets out to fix prices. What happens is that a salesperson forwards a competitor’s confidential quote, it gets pasted into the pricing analysis because it’s the most useful data anyone has, and now the CMA’s test is engaged.[7] Make the rule explicit with your team before the temptation arrives, because it will arrive.
The psychology tactics get implemented at face value. Charm pricing on established items and decoy tiers are the two most common outputs of an AI pricing prompt, and both have much weaker evidence than the confident summary suggests.[14,15,16] Test them on your own customers rather than adopting them because a model presented them as established practice.
If you do only one thing from this article, do step one. Pull twelve months of invoices and find out what you actually banked against what you invoiced. Most businesses find something in the first afternoon, and it’s the only pricing question where AI does the heavy lifting and the answer belongs entirely to you.
No, and every major vendor says so in its own documentation. Microsoft advises avoiding Copilot for decisions in sensitive areas including finance. Anthropic says Claude for Excel is not recommended for audit-critical calculations without verification. Google says not to rely on Gemini features as financial or professional advice. AI is genuinely useful for structuring the decision: reconstructing what you actually got paid, drafting a discount policy, tabulating research. The final number needs a human who can re-derive it.
Because those are two separate capabilities that do not run together. A model only knows what was in its training data up to its knowledge cutoff, so current prices require browsing, which is a distinct feature that has to be invoked. Meanwhile OpenAI’s documentation states that the Python environment used for data analysis cannot make external web requests. Ask one window to do both and you get a table mixing freshly fetched prices with remembered ones, presented identically. Browse first and output prices with URLs and dates, then analyse that file separately.
Using AI is legal. What is not legal is pricing influenced by competitors’ confidential information, whatever tool is involved. The UK CMA states that pricing guidance must not be influenced in any way by competitively sensitive information from rivals, even indirectly, and that not knowing what an algorithm is doing is no excuse. The US DOJ’s RealPage settlement is entirely about data inputs: no competitor nonpublic data at runtime, training data aged at least 12 months, geography no narrower than state level. In one UK case involving two online sellers using repricing software, a managing director was disqualified for five years.
Rebuilding your pocket price waterfall. Export twelve months of invoices, credit notes, rebate accruals, freight and settlement discounts, then have AI reconstruct what you actually banked per customer and per product as a percentage of list. The original Harvard Business Review research that produced the famous profit-leverage statistic measured gaps between invoice price and money received ranging from 16.7% to 28.9%, with price bands on identical products as wide as 500%. It uses only your own data, needs no market knowledge and carries no legal exposure.
Test them rather than adopting them. The evidence is much weaker than the confident summary you will be given. The $9 price ending effect fell from roughly 40% in a four-item pilot to 7% in a 270,000-customer study, and was not statistically significant on established items. The decoy effect was significant in only 6 of 38 studies, and in none of the 27 where participants could see or taste the product. The researchers who introduced it in 1982 have since written that it occurs rarely in the marketplace and was never intended as a practice tool.
This guide is built on primary sources rather than pricing folklore: the original 1992 Harvard Business Review analysis behind the profit-leverage statistic, price-change frequency data from US consumer microdata and an ECB survey of more than 11,000 euro-area firms, enforcement and guidance documents from the US Department of Justice, the Federal Trade Commission and the UK Competition and Markets Authority, published vendor documentation from Microsoft, Google, OpenAI and Anthropic, and the peer-reviewed pricing psychology literature including the replication work that substantially weakened several popular claims. Where a widely repeated figure could not be traced to a primary source, it was left out.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.