AI and Machine Learning in Price Intelligence

Most pricing teams did not spend this year wondering whether a machine learning model could set a price. That question was settled a long time ago. The work has moved somewhere less glamorous: deciding which market signals are worth trusting, how quickly they need to arrive, and how much of the decision anyone is willing to hand to software.

The constraint has shifted too. Models are cheap, capable and widely available. Clean, current, correctly matched market data is none of those things. And the rules governing what a company may do with a price once it has been calculated tightened sharply over the past eighteen months, particularly in the United States. Both of those facts change what a sensible AI pricing programme looks like in 2026.

What machine learning improved

Four things got measurably better, and they are worth separating from the broader claim that AI transformed pricing.

Demand forecasting improved because models can hold far more variables at once than a planner working in a spreadsheet. Seasonality, competitor promotional cycles, stock cover and category-level elasticity can be weighted together rather than reviewed in sequence. Anomaly detection improved in a similar way: instead of alerting on every price change above a threshold, a model trained on a category's normal behaviour can tell a routine weekend promotion apart from a competitor structurally repositioning a range.

Product matching improved, which matters more than it sounds. Comparing your price to a competitor's price only works if the two listings are the same product, and retailers describe the same item in inconsistent ways across variants, bundles and pack sizes. Machine learning made large-scale matching viable where rules-based matching had stalled.

Finally, extraction itself became more resilient. Retailer sites change layout constantly, and models that adapt to structural changes keep data flowing where hard-coded selectors would break. Our overview of web scraping techniques covers the mechanics of that in more detail.

On the commercial return, the most durable benchmark remains McKinsey's: dynamic pricing tends to deliver roughly 2 to 5 percent sales growth and 5 to 10 percent margin improvement when it is rolled out against tested pilot categories rather than applied to a whole catalogue at once. The same body of analysis found that a 1 percent price improvement, with no loss of volume, lifts operating profit by an average of 8.7 percent, which is why pricing keeps attracting investment ahead of most other levers.

Bain's work points in a consistent direction. Surveying more than a thousand commercial executives, it found that companies in the top quartile for revenue growth deploy generative AI use cases in sales and marketing about twice as often as those in the bottom quartile, and that teams working from data-driven deal guidance win more deals than they lose at a rate 12 percentage points higher than everyone else. Worth noting alongside that: those same executives named using AI effectively and managing pricing through volatility as their two most pressing challenges. Adoption and confidence are not the same thing.

Where these programmes usually stall

In our experience the model is rarely the reason an AI pricing programme underdelivers. The data layer underneath it is.

Coverage is the first gap. A pricing model is only as informed as the retailers and marketplaces it can see, and the sources that matter most to a category are often the ones that defend themselves most aggressively against collection. Freshness is the second. Categories move at different speeds, and a daily refresh that suits homeware is close to useless in consumer electronics during a promotional period. Matching accuracy is the third, and it is the one that quietly poisons everything downstream, because a model fed mismatched products will produce confident recommendations built on comparisons that were never valid.

History is the fourth and the most overlooked. Detecting that a competitor dropped a price is straightforward. Establishing who moved first, by how much, and whether an automated price-matcher simply followed them requires a timestamped record going back weeks. Without that history, a pricing team can see the present clearly and still be unable to explain it. Our piece on separating real market signals from routine noise goes further into that distinction.

These four constraints compound, which is why they are worth looking at as a chain rather than a checklist.

Public price monitoring and personalised pricing are now separate legal categories

This is the change most 2026 pricing content misses, and it has real consequences for how a programme should be scoped.

Monitoring competitors' publicly advertised prices and repricing your own catalogue against market conditions is one activity. Setting an individual shopper's price using data about that individual is a different one, and US regulators have spent the past year drawing a firm line between them. New York's Algorithmic Pricing Disclosure Act now requires businesses to display a conspicuous notice when a price shown to a consumer was set by an algorithm using their personal data, with civil penalties reaching $1,000 per violation. California's AB 325 took effect on 1 January 2026 and prohibits the use or distribution of common pricing algorithms in anticompetitive agreements. In late January, California's Attorney General opened an investigative sweep into how retail, grocery and hotel businesses use consumer data to set individualised prices. By March, New York's Attorney General was publicly backing a package of bills that would ban personalised algorithmic pricing outright in many contexts. Skadden's summary of the New York law is a useful starting point for anyone assessing exposure.

Consumer sentiment runs the same way. Gartner found that 68 percent of US consumers feel taken advantage of when a brand uses dynamic pricing, while 80 percent said brands holding prices steady are more trustworthy and 42 percent would pay more for guaranteed price consistency. A pricing engine can be technically sound, legally compliant and still cost a brand repeat buyers if shoppers read it as opportunism aimed at them.

The practical implication is that competitor and channel monitoring built on public web data sits on the safer side of that line, and it is where most of the commercial value has always been anyway. Minimum advertised price work belongs there as well, though the enforcement rules differ significantly by region, which we cover in our comparison of MAP monitoring software.

Bounded autonomy is the working standard for pricing agents

Agentic pricing tools arrived this year, and the honest picture is narrower than the marketing. Mature teams let agents monitor markets, simulate scenarios and recommend changes, then execute automatically only inside bounded domains with human decision rights above defined materiality thresholds. Clearance markdown automation within finance-owned floors works in production today. Autonomous repricing of core list prices, contract pricing or key accounts remains rare, and treating it as imminent is getting ahead of the evidence.

The failure modes are instructive because almost all of them are governance problems wearing technical clothing: pricing actions taken when a model has lost market context, silent drift as an upstream data pipeline changes shape without anyone noticing, and legal exposure from collusion or surveillance-pricing rules. An agent with a margin floor, an audit trail and a human sign-off above a set threshold is a manageable system. The same agent without those three things is a liability regardless of how good the underlying model is.

What to ask before committing to a platform

Compliance features, integration options and support tiers appear on every vendor's page, so they discriminate poorly between options. More useful questions are specific to your own catalogue.

Ask for a coverage demonstration on the retailers that matter to your category, including the difficult ones, rather than an aggregate site count. Ask what matching accuracy looks like on your hardest SKUs, the multi-variant and bundled items, and ask to see the failures. Ask what the refresh cadence is per source and whether it can differ by category. Ask what happens operationally when a major retailer changes its page structure on a Friday afternoon, and who is responsible for noticing. Ask how far back the price history goes and whether it is admissible as evidence in a retailer conversation. Ask how data lands in the systems your team already uses.

Teams that would rather not own that operational surface at all tend to end up with managed data delivery, where coverage, monitoring and repair sit with a provider. Teams that want to run detection and enforcement workflows in-house on governed data tend to look at Aperture. Either way, the questions above are the ones that separate a programme that holds up from one that produces confident numbers nobody can defend.