Price optimization: how AI sets prices from cost, demand and quote history
Price optimization is the work of choosing the price for each product, customer and order that best serves a goal, such as gross profit, given what the item costs and how buyers respond to price. Price optimization software does it with statistics on your own invoices and quotes, and AI handles the parts that used to need an analyst: rebuilding what each sale really earned, finding similar customers and orders, and explaining each recommendation. This guide covers how pricing optimization works for a manufacturer, a distributor or a food producer, the data it needs, how to start with one product family, and the legal and practical limits.
What price optimization means
Every pricing method answers the same question from a different starting point. Cost-plus pricing starts from what an item costs you. Competitor-based pricing starts from what others charge, which pricing intelligence covers. Price optimization starts from how your own buyers have responded to price in the past, and uses cost and the market as limits: a floor no price may go under, and a level above which you stop winning orders.
For a business that sells physical goods, prices are set in a few places at once: the list price and the price list, customer-specific prices and contracts, and the discount a rep gives on each quote. Price optimization can cover all three, starting with whichever one your data describes best.
Small price changes have a large effect on profit. In a 2003 analysis of the average income statement of an S&P 1500 company, McKinsey found that “a price rise of 1 percent, if volumes remained stable, would generate an 8 percent increase in operating profits.” The same article estimated that volume would have to rise 18.7% to make up for a 5% price cut. A smaller company’s numbers can differ, and the mechanism is the same: a price change goes straight to profit, while a volume change brings its costs with it.
How prices are set by hand
A manual process often looks like this. List prices go up once a year by a flat percentage. Sales reps discount from list within limits someone set years ago, and large customers have special prices kept in a spreadsheet or as overrides in the ERP. Quotes go out by email, and when one is lost, nobody records the reason or the price that won.
The price on the list is also rarely the price you keep. McKinsey’s pocket price waterfall follows a sale from list price down to the pocket price, which is what remains after every discount and allowance. In the article’s example, a lighting supplier that sold through distributors gave on-invoice discounts that left its average invoice price 32.8% below list. Off-invoice items, including prompt-payment discounts, the cost of carrying receivables, co-operative advertising allowances, volume rebates and freight, took another 16.3 percentage points, so the average pocket price came to about half of list. Some bulbs sold at a pocket price under 30% of list and others at 90% or more.
Averages hide a spread that wide. Seeing it means rebuilding every transaction from list price to pocket price, which is the first job for software in price optimization.
What AI changes
A pocket price for every invoice line
Price optimization starts by working out what each sale actually earned. Software reads invoice lines, credit notes, rebate agreements and freight bills, and allocates each discount and allowance to the line it belongs to. Rebate terms and freight charges often sit in PDFs and emails, which language models can read into structured data for someone to check. The result is a pocket price and a pocket margin for every line, and the rest of the analysis is built on those two numbers.
Here is one invoice line for a distributor, with made-up numbers:
| Step | Per unit |
|---|---|
| List price | $100.00 |
| On-invoice discount of 12% | −$12.00 |
| Invoice price | $88.00 |
| Prompt-payment discount of 2% | −$1.76 |
| Year-end volume rebate | −$3.00 |
| Freight paid for the customer | −$4.50 |
| Pocket price | $78.74 |
With a unit cost of $65.00, the invoice shows a margin of 26.1%, and the pocket margin is 17.4%. A margin report built from invoices alone would overstate this customer by almost nine points.
Similar customers and orders
Next, the software groups the transactions that should be priced alike, using attributes such as customer size and type, region, channel, order size and product family. Within each group, the spread of pocket prices shows where the same item sells at very different prices to similar customers. That spread is a good first place to look, because it can be narrowed without changing the list price at all.
How buyers respond to price
Then it estimates how volume responds to price. Economists call this price elasticity, which OpenStax defines as “the percentage change in the quantity demanded of a good or service divided by the percentage change in the price.” For catalogue items that sell every week, elasticity can be estimated from past changes in price and volume, once promotions, seasons and stockouts are accounted for.
For quoted work, the useful measure is the win rate. With a history of quotes and their outcomes, a model estimates the chance of winning at each price for a given kind of customer, product and order size, and multiplies it by the margin at that price. In a made-up example, a quote with $80 of cost has a 60% chance of winning at $96 and a 45% chance at $100. The expected margin is $9.60 at $96 and $9.00 at $100, so for that kind of customer the lower price earns more on average.
Price guidance people can act on
The output is guidance at the point of sale: a floor, a target and a stretch price for each item and customer group, shown in the ERP or quoting screen with the reason behind it. A rep who goes below the floor needs an approval, and the pricing manager sees which exceptions are granted most often. Guardrails keep the guidance within limits you set: no price below cost plus a minimum markup, no increase above a set percentage per quarter for any one customer, and contract prices left alone until the contract allows a change. Human in the loop covers designing that approval step.
Slow stock and perishable goods
Price optimization also decides when to cut a price. A food producer with short shelf lives, or a distributor holding stock that has stopped moving, can set markdowns from the quantity on hand, the days left to sell it and how buyers responded to past discounts. Inventory optimization covers the stock side, and SKU rationalization covers dropping the items that no price can rescue.
Three ways to set a price
| Method | Starts from | Strength | Weak point |
|---|---|---|---|
| Cost-plus pricing | Your unit cost and a markup | Fast, and straightforward to explain | Ignores what buyers would pay |
| Competitor-based pricing | What others charge for the same item | Keeps you in line with the market | Needs matched products, and follows competitors’ mistakes |
| Price optimization | How your buyers respond to price, within cost and market limits | Sets different prices where buyers respond differently | Needs enough clean history, and people who trust the guidance |
Find out what your sales history can support
Tell Derik which ERP holds your invoices and quotes, and where you think prices leak. He will tell you what your data can support today.
Start a conversationWhat data price optimization needs
- Invoice lines. List price, discounts, net price, quantity, customer, ship-to, date and salesperson, for two to three years.
- Off-invoice items. Rebates, payment discounts, freight paid on the customer’s behalf, and credit notes, linked to the invoices they belong to.
- Cost at the time of sale. The unit cost when each sale was made, so that margin is measured correctly.
- Quotes and their outcomes. Every quote with its price and whether it was won or lost, and the reason when you know it.
- Customer attributes. Size, type, region and channel, which the grouping uses.
- Competitor prices. Where you have them, matched to your items.
- Stockouts. The periods when an item was unavailable, which otherwise look like weak demand.
Volume matters. Items and customer groups with few sales give noisy estimates, so start with product families that sell every week, and keep one-off custom work on cost-plus pricing with a reviewed markup. If your ERP has no export or API, AI for ERP covers the ways to read it.
Price optimization software
Price optimization software comes in several forms, and the right one depends on how you sell. B2B pricing platforms group customers and products and send price guidance to quoting and ERP screens. Pricing modules in ERP and CPQ (configure, price, quote) systems apply price rules and approvals inside the tools reps already use. Repricing tools for online retail adjust prices against marketplaces and competitors as often as those prices change. A custom model built on your own data fits a business whose data or sales process does not match any package.
Ask any vendor to explain one recommendation from start to finish: which transactions it used, which customers it compared, and why the price moved. Ask how guardrails are set, where your sales data is processed and stored, and how the product is charged, for example per user, by revenue band or by quote, and check the vendor’s own pricing page. Private AI for business lists the data questions to put in writing.
How to start with one product family
- Choose a family that sells often, with enough customers and transactions to compare.
- Build the pocket price for each line from invoices, rebates and freight, and check the totals against the income statement.
- Look at the spread. List the customers who pay far below similar customers for the same items.
- Set floors and targets, then test them on last year: what would margin and volume have been?
- Show the guidance to reps for a quarter, with approvals below the floor, and compare win rates and margins with the quarter before. AI evals explains how to test on past results.
Margin analysis covers the reports that tell you which family to start with, and cost of AI explains how each part of a build is charged.
Risks, limits and the law
Estimates of price response are noisy. Promotions, seasons, stockouts and a competitor’s move can all look like a reaction to your price, and a model that misses them recommends the wrong change. Customers also notice inconsistent prices, so explainable guidance and a person who can override it matter more than a precise number.
Algorithmic pricing also draws regulators’ attention. In its 2025 discussion paper on algorithmic pricing and competition, Canada’s Competition Bureau says that “Using common pricing algorithms or pooling data among competitors may raise issues under the Competition Act,” because these practices “may facilitate coordinated behavior, such as price-fixing.” Building price optimization on your own sales data, with competitor prices taken only from public sources, avoids pooling data with competitors. Agreeing on prices with a competitor is an offence under section 45 of the Competition Act.
If you sell goods to customers in the United States who compete with each other, the US Robinson-Patman Act can also apply. The Federal Trade Commission explains that charging competing buyers different prices for the same commodity may violate the Act, and also that “Price discriminations are generally lawful, particularly if they reflect the different costs of dealing with different buyers or are the result of a seller’s attempts to meet a competitor’s offering.” Ask counsel before you set different prices for customers who compete with each other.
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems on a company’s own data, for businesses that make, move or sell physical goods. For pricing, that means reading invoices, rebates, freight and quotes from your ERP and documents, rebuilding the pocket price of every sale, and drafting price guidance and exceptions for a named person to approve. Every answer shows where it came from.
ThriveAI’s systems read your ERP as it is, including older versions installed on your own server, and every connection only reads data. They are designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.