Skip to content
Implementa.
AutomationPlaybook··7 min

Automating pricing with AI: before dynamic pricing, get your prices current

Automating pricing with AI almost never starts with dynamic pricing. In most companies margin doesn't leak because you missed the optimal price this hour: it leaks because your supplier raised prices in March and your price list caught up in June. What the machine can decide, what a person has to sign off, the two legal limits you don't automate, and where to start.

Senior AI Operations Implementer

AI Operations Pod

The email lands on a Tuesday: your main supplier is updating its price list, PDF attached, “effective the 1st.” It's 1,800 SKUs. Someone in purchasing forwards it to finance, finance parks it until month-end close, and your selling prices get updated six weeks later. For six weeks you've been selling at a new cost and an old price. Nobody made a bad pricing decision. Nobody made one at all.

The thesis in one line: automating pricing with AI isn't about charging the perfect price every hour, it's about not charging the wrong price for weeks. Dynamic pricing is the last layer, not the first. Two duller, more profitable layers come before it: your price list reflects cost the same day cost changes, and every price move happens inside a corridor a named person has signed off.

Automating pricing with AI: what it actually means (and what it doesn't)

“AI pricing” gets used for four very different things, with risks and returns that look nothing alike. Separate them before you ask anyone for a quote:

  • Passing through costs. When purchase cost changes, the selling price is recalculated with your margin rule and proposed for approval. It's the layer that recovers the most money in distribution and retail, and the one fewest companies do.
  • Rules with a corridor. Price adjusts for stock, shelf life, turnover or competitors' public prices, always between a floor and a ceiling set by someone with a name.
  • Demand-based dynamic pricing. Price moves with real-time demand. It makes sense with lots of transactions, lots of data and comparable products: airline seats, hotels, some ecommerce. In a B2B catalog with repeat customers, rarely.
  • Personalized pricing. A different price depending on the buyer's profile. It's the layer with the heaviest legal and reputational load, and in B2C it almost never pays off.

Notice where the AI actually is. In the first layer, the hard work isn't calculating margin —that's a formula— it's reading the supplier's price list: PDFs whose format keeps changing, part numbers that don't match yours, different units of sale. That's where a language model earns its keep, because it turns a messy document into a table your ERP understands. In the third layer, what does the work usually isn't an LLM: it's good old statistical elasticity models. If someone sells you a chatbot that sets prices, be suspicious.

Why the price nobody updates costs more than the price that isn't optimal

Price is the lever with the biggest effect on profit, and that's not our opinion. In “The power of pricing” (McKinsey Quarterly, 2003), Marn, Roegner and Zawada calculated that, on the average income statement of an S&P 1500 company, a 1% price increase at steady volume lifts operating profit by 8%. It's an illustrative average for US public companies, not a forecast for your business, but the mechanism holds anywhere: almost every point of price drops straight to the bottom line, because the cost is already paid.

The same arithmetic runs in reverse. If your cost goes up and your price doesn't, every order in those weeks ships with a thinner margin, and that margin doesn't come back. So the first question isn't “what's the optimal price?” It's “how many days pass between a cost change and a price change?” If the answer is measured in weeks, that's your project.

What the machine can decide and what a person has to sign off

Technology doesn't draw the line. Whoever answers for margin does. This is the split we use by default:

DecisionThe machineA person
Read the supplier price list and map SKUsYes, with review of the ones that don't matchResolves the exceptions
Recalculate price with the margin ruleYes, and proposesApproves the batch
Move price inside the corridorYes, logging every changeReviews the weekly trend
Set the floor, ceiling and margin ruleNoYes, per product family
Price for a customer with negotiated termsNoYes, always

The logic is the same as any human-in-the-loop automation: the machine handles volume and the person sets the rules and approves whatever falls outside them. What you never delegate is the policy: how much margin you want per family and how low you'll go to keep a customer.

The corridor: floor, ceiling and speed

A pricing algorithm with no limits is a risk nobody decided to take. Three parameters make it governable, and all three get written down before you switch anything on:

  1. Floor. The price you won't sell below, usually cost plus a minimum margin. If the model proposes something lower, it doesn't publish: it raises an alert.
  2. Ceiling. The price above which a repeat customer walks or gets angry. It protects the relationship, not the spreadsheet.
  3. Speed. How far a price can move in a day or a week. A price that changes three times in one morning generates customer calls, even if every change is defensible.

The first is antitrust, and the reference case is American. On November 24, 2025, the US Department of Justice announced a proposed settlement with RealPage, whose software helped landlords set rents. Among other things, it bars the software from using competitors' nonpublic data to set prices at runtime and limits model training to historical data at least twelve months old. It still needs court approval, but the lesson is clear: if your pricing tool feeds on your competitors' nonpublic data, the problem isn't technical. Competitors' public prices and your own data, yes; a shared pool of sensitive data with other sellers, no.

The second is consumer law, and it applies if you sell to consumers in the European Union. Directive (EU) 2019/2161 requires telling consumers when the price they see has been personalized through automated decision-making based on their profile; each member state wrote it into national law. Dynamic pricing that moves with demand or time of day isn't the same as personalized pricing, and that's the distinction you need to be able to explain if anyone asks. This is general information, not legal advice: check your specific case with counsel.

When NOT to automate pricing

Sometimes the right project is not doing it, and the list looks a lot like the one for when not to automate a process with AI:

  • Few SKUs and little cost movement. If you update forty prices twice a year, a well-built spreadsheet wins.
  • Price negotiated per account. In B2B with contracted rates, price is a commercial decision, not a calculation. There you automate the document, not the number, as we argue in automating proposals without losing margin.
  • Cost isn't reliable. If the cost in your ERP leaves out freight, rebates or shrinkage, the algorithm optimizes against a fake number, just faster.
  • Nobody owns margin. Without a person who signs off floors and ceilings, any automation ends up as an argument between sales and finance.

Where to start: four weeks, one product family

  1. Week 1: measure the lag. For the last price changes from your top three suppliers, count the days between the notice and your new price. That number is your business case, and it's how you choose which process to automate first with data instead of gut feel.
  2. Week 2: write the policy. Margin rule, floor, ceiling and speed for a single family. Signed by whoever answers for margin.
  3. Week 3: connect. The supplier price list comes in, becomes a table, gets matched to your SKUs and generates a price proposal in your ERP or store. It's a problem of integrating AI with your systems, whether that's Odoo or Shopify.
  4. Week 4: approve in batches and measure. One person approves the weekly proposal. You measure days of lag before and after, and how many proposals got corrected by hand.

If lag drops from weeks to days and manual corrections fall over time, you extend to the next family. If not, the problem was your cost data, and you found out with one family instead of the whole catalog. Building this loop end to end, from the supplier's email to the published price, is exactly the kind of work we do in operations automation.

The line for your next meeting: “Before we talk dynamic pricing, how many days does it take us to raise a price when our cost goes up?”

Shall we get it shipping?

If this resonated, 30-minute conversation with no commitment. We tell you what fits, what doesn't and the approximate price.

See cases
Automating pricing with AI: before dynamic pricing, get your prices current · Implementa