Agent-based pricing: Definition and Examples
March 16, 2026
Agent-based pricing entrusts a software agent with the task of achieving a price target by acting on its own, within limits set by the retailer. It differs from a “co-pilot,” which makes recommendations without taking action, and from rules, which execute actions without reasoning.
Autonomy is granted by product family and by decision type, not across all tools as a whole. Safeguards make all the difference: minimum thresholds, constraints, limits on the number of changes, logs, and undo functionality.
Agent-based pricing is often portrayed as simply a matter of a more powerful algorithm. The real issue is simpler—and more political: who sets the price, and within what limits.
This article defines agent-based pricing, distinguishes it from rules and the co-pilot, explains what a real-world experiment revealed, and outlines what we recommend delegating—and what we do not.

Agent-based pricing: definition
Agentic pricing refers to a system in which a software agent pursues an objective set by the retailer, independently chooses the actions needed to achieve it, and learns from their effects, all within predefined limits.
Its actions range from simple recommendations to the publication of a price, including simulations, alerts, and rollbacks.
This is neither automatic repricing —which mechanically matches a competitor's price—nor a tool that drafts responses. The difference comes down to one simple question: who sets the price.
Our take, after more than ten years of pricing projects: the key issue isn't the sophistication of the algorithm—it's delegation. The agent is just one of the building blocks of AI applied to pricing, and not even the most common one.
Rules, co-pilot, agent: Who approves the price?
Rule-based pricing software applies formulas that are set once and for all. AI-powered optimization software estimates the effect of a price on sales before suggesting it. The agent goes one step further: it takes action.
| Rules | AI Co-Pilot | Pricing Specialist | |
|---|---|---|---|
| Who sets the price? | The rule written in advance | Humans, as recommended | The agent, within its limits |
| Who Enforces It | The system, mechanically | The human, after validation | The officer, without waiting |
| Role of the Team | Write and maintain the rules | Confirm or correct | Set goals and limits; conduct audits |
| Main Risk | Rigidity, stacked rulers | Slow processing, bulk validation | Drift Not Detected in Time |
The assistant is often portrayed as a stepping stone to becoming a sales representative. In practice, however, it tends to depend on the product line: a single store may have some aisles staffed by assistants and others where the sales representative works alone.
Checking thousands of recommendations line by line isn't quality control—it's data entry in disguise. This is often what drives agents to focus on low-stakes items.
What a real-world experiment has shown
In 2025, Anthropic entrusted an AI agent named Claudius with managing a small store in its offices for about a month—including inventory, suppliers, and pricing.
The published report is informative. The employee sold items below their purchase price, distributed discount codes, and gave away products.
He also ignored an offer of $100 for a bundle that costs about $15, and announced the end of a discount before reinstating it a few days later (Anthropic, Project Vend, June 27, 2025).
The lesson isn't that the agent is useless. It's that a target without a margin floor, without a discount rule, and without human oversight produces exactly these kinds of errors. It's the safeguards that make the difference, not the model.
Levels of Autonomy: What to Delegate, What to Keep
Three stages are often described: validated suggestion, execution below thresholds, and post-execution verification. The framework is accurate, but it can be misleading if applied to an entire block catalog.
Our belief is that autonomy should be granted by product family and by type of decision, never for an entire tool. A price increase, a loss leader, or a product launch do not carry the same risk as a minor adjustment to a shelf-stock item.
In 2026, the areas we expect to remain under our direct oversight include: KVI, the depth of promotions, and new product launches. The areas we are more likely to delegate include: minor adjustments to non-flagship products, end-of-season markdowns for items below minimum stock levels, and anomaly alerts.
Even with competitor prices updated in real time, this retailer keeps decision-making at the corporate level.
Competitive prices that help you make decisions, not just sit on a shelf.
Competitors' prices are imported directly into the tool and matched to the retailer's products by cloning and linking SKUs. The price indices are up to date, and each price alignment is simulated before being approved by the team.
- Issue
- Track national competitors across a very extensive product catalog, store by store.
- What Has Changed
- Price indices, cloning and chaining matching, and centrally validated recommendations.
platform availability in 2025, with the platform having been in production since 2022.
Read more:Matching, Cloning, and Chaining
Pricing Department · National home improvement chain. Case study written by Booper based on an actual client project; the client has been anonymized.
The speed of data collection, therefore, does not dictate the speed of decision-making. To make a case-by-case determination, consider where to draw the line between AI that makes decisions and AI that executes them.
The Safeguards That Really Matter
The standard practices are well known: a minimum price based on cost, a price fluctuation range, an index linked to benchmark products, and price compliance rules. These prevent agents from selling at a loss or undermining the price image.
The most underrated safeguard lies elsewhere: the limit on the number of price changes per product and per period. A price that changes too often erodes customer trust, even when each change is profitable on its own.
Next come the log of every decision and the ability to undo a step with a single gesture. Without them, no one can justify a price, and the agent is taken offline at the first visible mistake.
Ultimately, an agent is at a loss without reliable data. If it compares a six-pack to a single unit, it’s making a decision based on inaccurate information—hence the importance of ensuring reliable product matching and price elasticity calculated by AI using its own historical data.
This is also the lesson we’ve learned from our first agent-based process in production, where the AI makes suggestions, the program monitors, and the user decides. We’ll share what we’ve learned by entrusting the chaining of product lines to an LLM.
Sellers' Agents vs. Buyers' Agents
The change doesn't just affect retailers who set their own prices. In the B2B sector, purchasing agents are already comparing offers in real time and spotting inconsistencies in pricing tables within seconds.
According to a Gartner forecast, B2B purchases will be handled by AI-powered procurement agents within three years.
(Gartner, as reported by Digital Commerce 360, November 2025)
When dealing with a buyer, the base price becomes the only factor compared if no other information is available. Perceived value, packages, and service terms once again become the levers that set the process apart from a purely automated bidding war.
When dealing with an agent that trades at machine speed, knowing how far to lower the price without eroding the margin becomes crucial. That’s the difference between feeding market signals into an AI and calculating a profitable price.
From the consumer's perspective: the agent makes the choice before the visit
For consumers, the AI assistant is becoming a gateway to purchasing. The customer expresses an intention, the agent compares options on their behalf, and the customer can now make a purchase without leaving the conversation.
U.S. retail sales could be driven by AI agents by 2030 ($900 billion to $1 trillion), and $3,000 to $5,000 billion worldwide.
(McKinsey, “The Agentic Commerce Opportunity,” October 2025)
This trend is already evident: during the 2025 holiday season, traffic driven by generative AI tools to U.S. e-commerce sites increased by 693% year-over-year (Adobe Analytics, January 2026).
A purchasing agent selects two or three bids and discards the others without reviewing them. The agent eliminates bids for products that are unavailable, prices that are out of line with the market at the time of the request, or product listings that are inconsistent.
For the retailer, it is the price image that is changing in nature: the agent does not perceive it, but rather measures it— product by product. We detail how agent-based commerce is changing the price image.
Price, forecasting, and inventory come together in a single decision: a well-timed price cut that causes a stockout removes the product from the market at the very moment when it was most competitive.
The Real Risk: A Price War Between Algorithms
Independent agents responding to the same signals can converge on abnormal prices without any explicit coordination. The research refers to this as tacit algorithmic collusion.
Regulatory Oversight
Algorithmic pricing is already under scrutiny
The European Commission confirmed in 2025 that it would examine several cases involving algorithmic pricing. In September 2025, the Polish competition authority (UOKiK) confirmed that it was investigating the bank and the pharmacy.
In the United States, the agreement reached in late 2025 between the DOJ and RealPage limits what a single tool can recommend to multiple market participants (Wilson Sonsini, 2026).
The documented approaches all follow a similar pattern: agents making decisions based on their own data rather than on shared data, and price floors and ceilings set before deployment, not added afterward.
In BOOPER MPS, the AI makes recommendations and runs simulations, while business rules set the parameters: target prices, minimum prices, and maximum prices may vary from one retail location to another, with an alert triggered whenever a local price falls outside these limits.
Each scenario is simulated in terms of sales, margin, and inventory before being implemented. The team decides, product family by product family, which aspects to approve itself and which to delegate.
Details are available on the modular pricing platform.
To learn more, see how an AI pricing engine works and why generative AI gets prices wrong.
See also in this feature
- AI and Pricing: What Really Falls Under the Umbrella of Artificial Intelligence
- Automating Your Pricing Without Losing Control: The Complete Guide for 2026
- AI Pricing Recommendations: Explainable and Auditable for Retail
- General-Purpose AI vs. Dedicated Pricing Solution: Why Combine the Two
FAQ
This pricing strategy is managed by a software agent that pursues a specific objective (margin, volume, price image), selects its actions, and learns from their effects, within limits set by the retailer.
It can make recommendations, run simulations, publish a price, send alerts, or undo actions, depending on the level of autonomy granted to it.
The co-pilot makes recommendations, and the human approves each price. The agent implements its own decisions—subject to safeguards—and the human reviews them afterward.
The two roles often coexist within the same catalog: co-pilot for sensitive products, agent for low-stakes adjustments.
The first applies predefined, handwritten formulas: coefficients, thresholds, and alignments. The second estimates the effect of a price on sales and gross margin before proposing it, based on the retailer's data.
Rules remain useful in both cases, as safeguards.
It improves the margin when it detects adjustments that no team has time to make, and when it is constrained by thresholds and variation limits.
Without these limits, it may, on the contrary, sell off assets at bargain prices, as demonstrated by Anthropic's Project Vend experiment in 2025.
A minimum price based on cost, a price fluctuation range, an index linked to benchmark products, compliance with legal requirements, and a limit on the number of changes per period.
Add a decision log and an immediate undo feature.
For stable, low-risk customer segments, yes—but only under strict limits and with post-hoc monitoring. For loss leaders, promotions, and new product launches, we recommend retaining manual approval.
No. It shifts the nature of their work: less data entry, more decision-making. They set goals and limits, decide what to delegate, and review the agent's decisions.
By setting minimum and maximum thresholds before any deployment, refusing to automatically and fully align with every competitor’s move, and letting the agent make decisions based on your own data.
Sources: Anthropic, Project Vend (June 27, 2025); Gartner via Digital Commerce 360 (November 2025); McKinsey, “The Agentic Commerce Opportunity” (October 2025); Adobe Analytics (January 2026); Wilson Sonsini, “2026 Antitrust Year in Preview.” Updated: October 2026.
Well-defined business rules are often sufficient for a narrow, stable, and loosely interdependent product catalog. AI becomes cost-effective when decisions number in the thousands, products influence one another, and price sensitivity varies.
But you still need a usable sales history. In most cases, the right answer is a hybrid approach: the rules set the framework, and the model optimizes within it.
AI handles the volume at every stage of the data journey: collection, product matching, verification, forecasting, and simulation. The expert selects competitors, resolves ambiguous cases, provides context, formulates the strategy, and makes decisions.
At Booper, Booper Mind makes suggestions, MPS monitors, and the user decides. Knowing how to interpret elasticity and question a forecast is more important than knowing how to code.
Dynamic pricing is permitted in France. Personalized pricing based on automated decisions is allowed, but customers must be notified of it. Algorithm-based price-fixing remains prohibited, even in the absence of contact between competitors.
The strongest safeguard for a retailer: being able to say, for each price, which rule generated it, what data it was based on, and who approved it.
.avif)