Where AI earns its place in pricing (and where it doesn't)

There is a lot of pressure to "put AI into pricing". Some of it is justified. Much of it skips the question that matters: what would AI do here that a clear rule, a clean data feed and a weekly review don't already do?

Our view is simple. AI is optional. It belongs where it clearly pays back, and the business should be able to switch it off without the pricing falling over. Google's engineering guide Rules of Machine Learning opens with the same idea: "Don't be afraid to launch a product without machine learning." It argues that a basic heuristic will often get you a good part of the way, and that machine learning needs data before it is worth using.

Where AI usually earns its place

The common thread is routine, high-volume work where a person still checks the output.

Rate checks. Collecting competitor and market rates, normalising them into a comparable unit and flagging the moves that matter is tedious, repetitive and well suited to automation. The output is an input to a pricing decision, not the decision itself. Only verified observations should feed any rule.

Exception and anomaly flags. A model that scans overrides, approvals and transactions for the unusual (a discount far outside a team's normal pattern, a price below floor, a sudden change in mix) lets analysts spend their time on the few cases that need judgement.

Classifying messy text. Free-text override reasons and lost-deal notes are full of useful information that nobody reads. Language models are good at sorting them into a short list of categories, so the weekly review can see that most "competitor match" claims came from one region, or that one product keeps losing on terms rather than price.

First-draft commentary. The weekly pricing pack needs a paragraph explaining what moved and why. AI can produce a first draft from the numbers, which an analyst then checks and edits. The analyst stays accountable for what is said.

Report refresh and reconciliation. Much analyst time goes on pulling data, rebuilding the same tables and chasing reconciliation differences. Automating that is often the quickest payback of all. Our team's reporting automation work freed 90–95% of analyst time. These figures come from roles our team held directly, not from client engagements.

Forecasting and elasticity, when the data supports it. Demand forecasting in high-volume businesses, and win-probability or elasticity models in quote-driven businesses, can add real value. They need enough clean history, with outcomes recorded, to be trustworthy, and they should prove themselves against simple rules in a controlled comparison before they take over anything.

Illustrative example. A pricing team spends most of Monday building the weekly pack by hand. Automating the data pull and drafting the commentary means the meeting starts with the five exceptions that need a decision, rather than with a debate about whose numbers are right. The pricing decisions themselves don't change hands.

What should stay with people

Some decisions set the business's risk, fairness and reputation. These should stay human, even where a model could technically make them.

• Guardrails. Floors, ceilings, change limits and competitor response bands. Models work inside these limits. They never set their own.
• Break glass. Suspending a guardrail during a market shock or a feed failure is a judgement call with a named owner.
• Prohibited inputs and protected segments. Which characteristics prices must never respond to, and which customers or products are outside dynamic pricing altogether.
• Strategic stances. Whether to respond to a competitor's price war, hold price through a downturn or reprice a legacy book.
• Large or unusual deals. Anything outside normal authority, where relationships and context matter.

Regulators are moving in the same direction. The EU AI Act takes a risk-based approach and treats some uses, such as credit scoring that can deny people a loan, as high-risk, with obligations that include appropriate human oversight. The UK Competition and Markets Authority's paper Algorithms: How they can reduce competition and harm consumers covers personalised pricing and algorithmic collusion. Neither means pricing models are off limits. Both mean someone has to own what the model does.

How to decide, case by case

For each candidate use, ask four questions:

1. What is the payback? Name the hours saved or margin protected, and compare it with what it costs to build and maintain.
2. What happens when it is wrong? If a mistake is caught in review, the risk is low. If it reaches customers' prices directly, the guardrails must catch it first.
3. Is the data good enough? If the inputs are incomplete or unreconciled, fix that first. Missing data is a finding, not a reason to add a model on top.
4. Who owns the outcome? Every model needs a named owner who can explain what it does and switch it off.

A risk framework helps make this routine. NIST's AI Risk Management Framework is voluntary and written for any organisation, and it is a useful structure for thinking through trustworthiness before a tool goes live.

Data handling matters too. Keep commercial data in your own systems wherever possible, and use AI tools on it only with the data owner's agreement and only where the tool doesn't train on it.

Figures come from roles our team held directly, not from client engagements. Examples marked illustrative are hypothetical; third-party figures are from the sources linked.