Every retail leadership team in 2026 is being pitched “AI pricing” every other week. The pitches are getting more confident; the public and regulatory tolerance for what’s underneath them is getting lower. Most buyers don’t yet have a clear framework for separating the honest tools from the others. Here is one — built around the questions a retailer should be asking before signing anything.
The buyer’s problem in 2026
The decision facing most retail leadership teams this year is not whether to bring AI into pricing. That decision has been made by the market — every credible technology stack now includes machine learning somewhere in the price-setting flow. The decision is which AI, for what scope, and with how much human oversight.
That decision is being made in a much harder environment than it was 24 months ago.
In December 2025, Consumer Reports — working with Groundwork Collaborative and More Perfect Union — published a landmark investigation into what is now widely called “surveillance pricing.” The investigation documented machine-learning systems that abandoned the traditional “one price for all” model in favour of pricing varied by individual customer. The same basket at the same store could differ by roughly seven percent depending on who was buying it, potentially adding more than US$1,000 a year to some households’ grocery bills.
Within days, Instacart halted retailer AI pricing experiments on its platform, ending the use of Eversight, a tool it had acquired for US$59 million in 2022. The Federal Trade Commission issued a civil investigative demand. State legislatures had already been moving: as of 1 January 2026, several states require retailers to disclose when prices are derived from algorithms that use personal data, and other states have banned the use of shared pricing algorithms across competitors as a restraint of trade. In March 2026, the chair of the House Oversight Committee sent an investigative letter to Uber over surveillance-pricing practices. The Department of Transportation is examining airline algorithmic pricing. Walmart, separately, obtained two patents covering AI-driven price recommendation tied to consumer-specific data including individual ID, and was widely condemned for the disclosure.
This is the environment a retailer is walking into when a vendor knocks on the door with an “AI pricing” pitch. It is no longer the 2018 question of “is the technology ready.” It is now the much harder question of: do you want to be in the news for the right reasons, or the wrong ones — and how do you tell the difference before you sign.
What honest AI pricing actually does for a retailer
It would be easy, reading the 2025-26 headlines, to conclude that AI in pricing should be quarantined. That would be a mistake. There is genuine, valuable work AI does in pricing that has nothing to do with surveillance and nothing to do with the bleeding edge.
The most reliable wins are unglamorous. Cleaning and matching data across point-of-sale, inventory, and competitor systems is a job AI does at a scale humans cannot. Surfacing pricing candidates — which SKUs deserve a human review this week and which can stay where they are — is another. Building elasticity models that hold more factors in view simultaneously than a human pricing team ever can: seasonality, stock position, cross-elasticity, competitor moves, newness, market trends, advertising activity, promotional history.
With proper training, calibration, and monitoring, AI can do more. It can discover real promotion opportunities from real-time data — promotions that genuinely move volume, rather than the predictable weekly cycle most retailers run on autopilot. It can identify unwanted inventory and clear it on schedule, before it becomes a write-off. And in fresh and perishable categories, it can meaningfully reduce food waste — possibly the single most important sustainability lever in retail in 2026, and one that improves margin and reduces environmental impact at the same time.
None of those wins require pricing different customers differently. None require surveillance. All of them improve the retailer’s operations and the customer’s experience at the same time. They are also exactly the kind of work that does not make headlines, which is part of why they get under-invested in by leadership teams more interested in transformation stories.
The line between honest pricing and surveillance pricing
The most damaging misconception about AI pricing in 2026 is not a technical one. It is a moral one. A significant share of the public — increasingly amplified by the press, by regulators, and by politicians on both sides of the aisle — believes AI in retail pricing exists to trick customers. To surveil purchase histories and identify each shopper’s pain point. To surge-price daily essentials. To extract maximum willingness-to-pay from each individual at each moment.
The conflation is doing real damage. The same algorithms can power waste reduction or surge pricing. The same vendor can sell either. The difference is intent, scope, and governance — not the technology underneath. But the label “surveillance pricing” has stuck to the entire category, and honest retailers trying to adopt AI for genuinely useful work are being tarred with it.
The buyer’s first job is to know which side of the line each AI capability falls on, and to be clear with the team, the board, and the customer base about which they are doing.
Honest AI pricing matches supply and demand more accurately. Reduces waste. Surfaces real promotions. Routes inventory away from write-offs. Keeps the same price visible to every customer in a given moment. Operates on the assortment, not on the individual.
Surveillance pricing collects individual purchase histories, payment patterns, location data, and behavioural signals to identify each customer’s price ceiling. Charges different shoppers different prices for the same item at the same moment. Operates on the individual, not on the assortment.
Same mathematics, different politics. Different ethics. Different regulatory exposure. The retailers who adopt the first and reject the second will compound an advantage in 2026 and beyond. The retailers who do not draw the line publicly will find the line drawn for them, eventually, with a blunter pen.
The cockpit, not the empty plane
Two pilots fly a complex commercial airliner with technology doing most of the work. The cockpit is full of automation. The autopilot handles the cruise. The flight management system handles navigation. The autoland system can put the aircraft on the runway in zero visibility. But two trained humans sit at the front, watching the systems, ready to intervene, accountable for the outcome.
There is no reason AI cannot eventually run more of retail pricing the same way. The question is the sequence, not the destination.
The bleeding-edge approach skips the sequence. It hands the autopilot the entire flight on day one, with no certified pilots, no checklist, no manual override. That is the approach that produces the failure stories everyone remembers — and the seven-percent basket variation that triggered the 2025 Consumer Reports investigation. It is what “AI pricing” looks like when a retailer treats it as a replacement for human judgement rather than a tool the team learns to fly.
The leading-edge approach is the cockpit. Start with the parts of the work the algorithm is genuinely better at — data cleaning, candidate surfacing, anomaly detection, the multidimensional models a human team cannot hold in their heads. Add complexity one step at a time. Keep humans on the price decisions until each new layer has earned trust by track record. Build override mechanisms before you need them. Log everything.
After 21 years of working on retail pricing across three continents, the pattern is consistent. Every successful AI pricing program I have been part of, or watched succeed, has looked like the cockpit. Every failure has looked like the empty plane.
The first question to ask any vendor
When an AI pricing vendor walks into a retailer’s office with a deck in 2026, the first question to ask is not about model architecture, training data, or accuracy benchmarks. The first question is:
“Show me where humans stay in the loop — and on which decisions specifically.”
If the answer is detailed, with specific decision points named and specific override mechanisms documented, the conversation can continue. If the answer is hand-waving — “the system handles it,” “you do not need to worry about that,” “our customers trust the model” — the conversation should end. A vendor that cannot describe where humans stay in the loop is selling either an unfinished product or an irresponsible one.
The follow-ups matter too. Who is accountable when a price is wrong — your team, my team, or the algorithm? What happened the last time the model recommended something the customer’s pricing team rejected? Can you show me the model’s reasoning on a specific SKU in a specific week? What data does the system use — and, crucially, does it use individual customer data? The first question — humans, where, on what decisions — is the one that separates honest vendors from surveillance-pricing pitchmen, and an unwillingness to answer it cleanly is a louder signal than any feature on the slide deck.
The biggest risk to plan for
The biggest single risk in adopting AI pricing is the one that has not changed in 15 years: the algorithm prices something absurd before anyone notices.
The modern failures look smaller than the famous historical ones, but they compound faster. Instead of a single absurd price on a single SKU, the failure mode in 2026 is a slow drift across an assortment — a basket of essentials that becomes seven percent more expensive for some customers than others, week by week, until a regulator or a journalist starts looking. By the time anyone notices, the damage to brand trust, regulatory standing, and competitive position is substantial. Walking it back is harder than walking back a single bad price on a single SKU.
The other risks are real but secondary. The pricing team stops developing intuition because the algorithm does it for them. Competitors learn to game the system by setting fake reference prices. Internal trust breaks down after one bad price and the whole programme stalls. Each is a planning problem. Each is solved by the same things — daily review cadence, documented override mechanisms, comprehensive logging, and a culture of treating divergences between algorithm and human as research opportunities rather than embarrassments.
Plan for the failure modes before the system goes live. The retailers who have done this well have an answer to every one of those failure modes documented and rehearsed. The ones who have not are running a different kind of pilot.
The 90-day starting plan
For a retailer adopting AI pricing in 2026, here is a sequence that works:
- Pick one category. Ideally specialty or a single curated assortment — base margins are higher, the assortment is small enough to manage, and data structures tend to be clean enough to model from day one.
- Connect cost, volume, and observed competitor pricing in one place. The data does not have to be perfect. A loose connection is enough to start.
- Tag every SKU in scope with a role — profit driver, image item, can promote, improve visibility, hold. The roles are the strategic decision; the prices follow.
- Run the model. Log every recommendation. Log the human decision the pricing team would have made independently. Run this comparison for 90 days minimum.
- Where the algorithm and humans agree, you have a candidate for expanded autonomy. Where they diverge, treat each divergence as a research case — sometimes the algorithm is right; sometimes the human is. Either way, the learning is the asset.
- After 90 days, expand by track record. Add another category. Add another decision layer to the autonomy budget. Keep humans on the decisions you have not yet learned to trust.
- Re-optimise the entire scope at least every six months. Faster in categories where costs or competitive moves are more frequent.
- Build explainability into the model from the start — not as a marketing feature, but as an operational fact, available on request. In 2026, a retailer is building trust on three audiences simultaneously: internal teams, customers, and regulators. A retailer that has built the first two has usually built the third by accident. A retailer that has not built any of the three is in real trouble.
The honest version
AI is the most important tool to have come into retail pricing in 20 years. It is also the most consequential one to misuse. The retailers who treat it as a collaborator that earns autonomy step by step will compound the gain. The ones who treat it as a replacement that gets handed the keys on day one will produce the next batch of stories the rest of the industry has to spend years explaining away.
Two pilots in the cockpit, the technology doing most of the work, the humans accountable for the outcome. That is the model. Buy for that. Build for that. Be clear publicly about which side of the surveillance-pricing line you are on, and why. The regulators will eventually draw a line. The retailers who draw their own first will be the ones still in business when they do.
Kiran Gange is the founder and CEO of RapidPricer, author of The Expert Guide to Retail Pricing (Routledge), and host of the Retail Price Talk podcast. He has 21 years of experience in retail pricing across three continents.