Why AI Pricing Tools Can Make Brands More Predictable Than Profitable

What have AI pricing tools been shown to earn?

Nine authors from PepsiCo described in June 2026 how the company uses AI to plan its promotions and its pricing. In most markets about 85 percent of the promotion tool's recommendations were accepted. What the promotions earned is not reported, because "specific financial figures are commercially sensitive and cannot be disclosed".¹

In their place the paper cites an industry benchmark of "2.0-5.0% improvements in both revenue and profit", taken from a 2021 BCG article. That article says AI-based tools "can boost EBITDA by 2 to 5 percentage points", with no study or client base for the range.² A gain measured in points of margin is the far larger claim. EBITDA, earnings before interest, tax, depreciation and amortization, is a common measure of operating profit. At a 15 percent margin, two more points on the same sales raise EBITDA by 13 percent (2 divided by 15). A 2 percent rise in profit is less than a sixth of that.

A 2017 study in the journal Operations Research used one grocery retailer's data to estimate that an optimized promotion schedule could raise the retailer's profit by about 3 percent, or 5 percent if some business rules were relaxed. It is a model estimate; no store test is reported.³

Public evidence of what these tools add to a manufacturer's profit is thin, so each manufacturer has to measure the gain itself, with some retail accounts or regions planned with the tool and some without.

What are companies using AI pricing for?

In a 2025 survey by Simon-Kucher, a pricing consultancy, 72 percent said they used AI in pricing. The most common use, named by 41 percent, was gathering market intelligence: analyzing competitors' prices and customers' sensitivity to price.⁴

Rivals in consumer goods feed their tools much the same inputs: manufacturers in a category often buy their market data from the same few providers, so they share one sales history and its gaps in coverage. Their models make similar assumptions. PepsiCo's pricing model rules out any estimate in which a 1 percent price rise cuts sales by more than 2.5 percent, and assumes that competitors' prices move in step with the price of a matching PepsiCo product.¹ A rival with a similar tool will probably make the same reasonable choices, so both tools are likely to be wrong in the same direction when the market changes.

Any lasting advantage lies in what a rival does not have: a manufacturer's own shipment and account data, its own tests of list prices, packs and promotion mechanics, and the judgment of whoever decides when to overrule the model. Those tests show what shoppers are willing to pay for the brand, which is what its list price should rest on. Competitors' prices are one input, and costs set only the floor.

A tool that rivals can buy, fed with data they can also buy, is unlikely to do more than keep a manufacturer level with them.

What does it cost a manufacturer to be predictable?

Software that reads competitors' history reads the manufacturer's own history too. The Academy's lesson on calendar optimization tells planners to pull two to three years of each rival's promotions from market data, because a rival's "next year's calendar will likely resemble their last". Tools that learn from the same seasonal pattern also point rivals at the same best weeks, so two manufacturers can clash without either one reading the other.

An invented snack manufacturer normally sells one retailer 10,000 units a week, or 20,000 over two weeks. It receives 2.50 euros a unit after its everyday trade terms, and each unit costs 1.50 euros to make. That leaves 1.00 euro of gross profit a unit (2.50 less 1.50). For a two-week promotion it pays the retailer 0.40 euros on every unit sold and a fee of 4,000 euros for a leaflet feature. The shelf price during the promotion, and so how much of the 0.40 euros reaches shoppers, is the retailer's decision; no figure below is calculated from it.

With no rival on promotion, the manufacturer sells 2.4 times normal, or 48,000 units (2.4 times 20,000). The uplift is 28,000 (48,000 less 20,000). Shoppers who stocked up then buy 2,000 fewer units over the next two weeks, so net incremental volume is 26,000 units, worth 26,000 euros of gross profit at 1.00 euro a unit. The investment is 23,200 euros: 19,200 of funding (0.40 times 48,000) and the 4,000 euro fee. The promotion adds 2,800 euros to the manufacturer's gross profit (26,000 less 23,200).

Suppose a rival promotes in the same weeks and cuts the uplift by a quarter, to 21,000 units. Net incremental volume falls to 19,000 units (21,000 less the dip), worth 19,000 euros. The investment falls only to 20,400 euros: 16,400 of funding (0.40 times the 41,000 units sold, 20,000 normal plus 21,000 uplift) and the 4,000 euro fee. Each unit of uplift lost to the clash removes 1.00 euro of gross profit and saves only 0.40 euros of funding. The manufacturer ends 1,400 euros worse off than with no promotion (19,000 less 20,400).

If the rival instead promotes in the two weeks after and takes 3,000 units of the manufacturer's sales, the dip grows to 5,000 units. Net incremental volume falls to 23,000 (28,000 less 5,000). Against the same 23,200 euros of investment, the manufacturer ends 200 euros worse off.

Over a year of eight such promotions, a tool that plans each one as if no rival will respond forecasts 22,400 euros of extra gross profit (8 times 2,800). If a rival responds to four of them, two in the same weeks and two just after, the year brings in 8,000 euros. The four unanswered events earn 11,200 (4 times 2,800); the two clashes lose 2,800 (2 times 1,400) and the two answered just after lose 400 (2 times 200). The forecast was 2.8 times what the year earned (22,400 divided by 8,000).

To break even in a clash, the event must earn 26,000 euros of gross profit after funding, enough to cover the 4,000 euro fee, the 2,000 euro dip and the 20,000 euros the two weeks would earn without a promotion. At 0.60 euros a unit (1.00 less 0.40) that takes 43,333 units (26,000 divided by 0.60), or 23,333 more than normal. A clash keeps only three-quarters of the uplift, so with no rival the event would need an uplift of 31,111 units (23,333 divided by 0.75). That is 51,111 units in all, or 2.56 times normal (51,111 divided by 20,000).

The manufacturer's promotions sell 2.4 times normal, so they lose money whenever a rival answers them. The plan still adds 8,000 euros, and the case is for redesigning these events rather than cutting them. A feature fee 1,400 euros lower, 2,600 instead of 4,000, would remove the 1,400 euro loss and let each event break even in a clash, provided the smaller feature still sold 2.4 times normal.

Being harder to predict has a cost too, and the expected-value test in the Academy's war gaming lesson shows when it is worth paying. Suppose that moving an event three weeks away from its best week leaves it selling only 2.2 times normal. It then adds 400 euros instead of 2,800 (44,000 units at 0.60 euros is 26,400, less the 4,000 euro fee, the 20,000 euros of two ordinary weeks and the 2,000 euro dip). If a rival's answers split evenly between clashes and the weeks after, an answered event loses 800 euros on average (the average of the 1,400 and 200 euro losses). That is 3,600 euros below the 2,800 an unanswered one earns (2,800 plus 800). Giving up 2,400 euros (2,800 less 400) pays only if a rival would otherwise answer more than two in three events (2,400 divided by 3,600). The test also assumes that the move stops the rival answering at all.

A forecast that ignores rivals' responses can overstate what a promotion plan earns, in this example nearly three times over.

Where does the extra margin in the research come from?

Economists who find algorithms raising margins across a market trace the gain to weaker competition. A study of Germany's retail fuel market, where many stations adopted pricing software, found that in local markets with two or three competing stations, margins rose only when all of them had adopted it. The study does not show how the software weakened competition. Its authors raise two possible routes: one vendor acting as a hub between the stations it served, or each station's own software reacting to the other stations' prices.⁵

Regulators now treat the first route, a shared tool or pooled data, as a legal risk. The European Commission's guidelines of 2023 say that feeding pricing software with public data is legal, but that pooling commercially sensitive information in a tool to which competitors have access "could amount to horizontal collusion", that is, collusion between competitors.⁶

Two American appeals courts have reached different results on different facts. In August 2025 one upheld the dismissal of a case against Las Vegas hotels that had each bought the same pricing software, which did not pass any hotel's confidential data to the others. In July 2026 another revived a case against Atlantic City casino hotels whose software allegedly pooled competitors' non-public data. Because AI tools adapt to new data, the court noted, "collusive conduct may occur at later points in time when other competitors begin to share their non-public commercial data with the software".⁷

The second route needs no shared data. In a model calibrated to five large online sellers of allergy medicines, two economists found that a firm whose software reprices faster than its rivals can make a believable promise to beat whatever price they set. Rivals that know this compete less hard: in the model, average prices rose by 5.2 percent and profits by 9.6 percent.⁸ The advantage goes to whoever reprices fastest, and in grocery that is usually the retailer. A manufacturer's list price takes weeks to change, because each retailer has to accept it, while a retailer can change its shelf prices the same afternoon.

A manufacturer that keeps its data out of shared tools, and whose list price moves more slowly than its retailers' shelf prices, gains from neither route. The strongest case for its tools is that they plan better: fewer errors, calendars that stay within each retailer's rules, and faster work, with some of PepsiCo's tasks falling from weeks to minutes.¹ For many mid-sized manufacturers, judging promotions with no analysis at all is still a larger risk than being predictable. A steady rhythm of promotions also has value to a retailer and a factory that plan their year around it.

The evidence supports the case for better planning more than it supports any promise of higher margins.

Which decisions should stay with people?

Promotion planning, PepsiCo's authors write, "operates in a data-rich environment with discrete, reversible decisions", while strategic pricing "faces severe data limitations due to infrequent price changes".¹ List price and the response to a rival's move are harder to reverse and rest on less data. What the tool is told to maximize, and the profit floors it must respect, express the company's strategy. These belong with named people, as do the negotiation with each retailer and the options the key account manager puts to the retailer.

In the Atlantic City case, where the software allegedly pooled rivals' data, the appeals court said that an alleged "90 percent compliance rate certainly supports an inference of an agreement" between the hotels.⁷ For a tool a manufacturer runs on its own data, a high acceptance rate raises a management question: whether people are still checking the advice or simply accepting it. Research on automation bias, the habit of deferring to a machine's advice, finds it in experts and teams as well as in novices and individuals, and finds that training does not prevent it.⁹ The useful records are a company's own: which recommendations were overridden, what each override earned, and which accepted recommendations lost money.

RealPage, an American maker of rent-pricing software, agreed in a settlement with the Department of Justice in November 2025 to stop its software using competitors' non-public data and to use non-public data in training only once it is at least 12 months old.¹⁰ The settlement's terms suggest four questions to put to any pricing, promotion or benchmarking vendor in writing. Does the manufacturer's non-public data feed recommendations its competitors receive? Does competitors' non-public data feed the recommendations it receives? Does any feature limit price cuts or align prices across users? How old is any non-public data used in training?

The EU guidelines also warn that a retailer can act as a hub that relays information between manufacturers, so a manufacturer should not ask a retailer, a vendor or a data provider which weeks a rival has booked, and should give a retailer its own future prices and promotion plans only for that retailer's own planning.⁶ This article is not legal advice. The rules differ between the EU, American states and other countries, so vendor contracts and data-sharing terms need competition counsel in each place where a manufacturer sells.

Manufacturers that test each event against a rival's likely response, and that know what their vendors do with their data, will have a clearer measure of what their tools earn than an acceptance rate or a vendor's benchmark.

References