Author
TL;DR In an agentic pricing platform, AI agents do the analysis and prepare recommendations. Deterministic arithmetic, not an agent, computes every number that reaches a quote, invoice or forecast. A person reviews and approves every change to licensing, packaging or pricing. Three questions put to any vendor show which work its agents do and which changes they can make alone.
An agentic pricing platform, defined
An agentic pricing platform is software in which AI agents analyze deals, trends and competitor moves and prepare pricing recommendations. Judge a platform by the work its agents touch, not by how many it lists.
Pricing work divides into three kinds: analysis, calculation and decision. Analysis reads deals, trends and competitors for what has changed. Calculation produces the numbers that reach a quote or a forecast. Decision settles what the company charges, for what, and to whom.
Each kind tolerates error differently. A missed pattern can be caught next week. A wrong number in a signed contract stays wrong for the term. A wrong decision on the value metric carries into every contract written after it. That difference is why each kind belongs to a different actor: agents for analysis, computation for calculation, people for decisions.
Pricing an agent you sell is a separate question, covered in agentic AI pricing strategy.
What agents do well in pricing work
Agents suit the analysis because the volume is beyond a team and an occasional miss is recoverable. A pricing team of two cannot read every closed deal, every refused quote and every competitor announcement each week. Agents can.
What does an agent see that a quarterly review misses?
It sees three kinds of signal. The first is patterns across deals: which editions sell together, where discounts cluster, which Customer Groups stall at the same stage. The second is trends as they form, such as a discount level that was rare last quarter and is common this one. The third is competitor moves: a new edition, a changed value metric, a price that held when the market expected a cut.
A quarterly review meets these after they have compounded. Agents read the same record each day and raise these signals while a response still costs little.
Tracking the impact of a pricing change by cohort
Most pricing work stops when the change ships. The company raises a price, moves a capability between editions or introduces a new value metric, and then turns to the next problem.
Across the software companies in our pattern library this is the common case. Six months after a change, the account of what it did is qualitative: a sales leader’s impression and a few renewals someone remembers. Nobody measured it, because nothing was in place to measure it with.
Without tracking each change’s effect by cohort, a company repeats changes without learning which ones yield which results. Agents make the tracking practical. They follow the accounts a change touched and report what those accounts did next. Over several changes the company builds a record of its own pricing decisions and their effects.
What agents should not do today: the calculations
In a well-built agentic pricing platform today, no agent calculates a number that reaches a customer. Deterministic arithmetic computes every number that reaches a quote, an invoice or a board forecast, and it returns the same answer each time it runs.
The reason is the error rate of today’s language models. A model that is right most of the time can read a thousand deals. It cannot be trusted to price one.
The dependable design puts a deterministic engine under the agents. The engine holds the computations as pipelines that return the same answer each time. The agents decide which pipelines to run and in what order, and they read the results. Engineers call that engine the harness.
When the whole job goes to an agent with no engine beneath it, the gap shows in three ways. Ask the same question twice and the answer can differ. A number cannot be traced back to the formula that produced it. The definition of a discount or a cohort is worked out again on each run, so last quarter’s result cannot be compared with this quarter’s.
That line describes today’s models. As they become more reliable it may move, on evidence. A person’s approval of each change is a separate line, and we do not expect it to move.
The same split separates AI-driven pricing from AI-augmented pricing: the machine supports the judgment, and a person sets the price.
If an agent can’t calculate the price, can your pricing rules calculate it alone?
Deterministic arithmetic only works when your licensing, packaging, and pricing rules are explicit enough to compute without judgment. Score your architecture and see which rules still live in a rep’s head instead of a formula.
A recommendation and a change are different events
An agentic pricing platform can recommend a licensing, packaging or pricing change. The recommendation becomes a change only when a person reviews it and approves it.
That approval is a judgment the person makes with their own knowledge, and the agent’s recommendation is one input to it. Licensing, packaging and the value metric are structural decisions. They change rarely, and each one carries through every contract signed afterwards. A recommendation can be well supported and still wrong for reasons no record contains: a product direction not yet announced, a customer relationship, a commitment made in a negotiation.
Who approves a change when the vendor runs pricing for you?
A person inside the software company approves every change, even when an outside firm runs the pricing platform. Some companies run a pricing platform themselves, with outside experts beside them. Others have a firm run it on their behalf. The operating work can move to the outside team, and the approval stays with someone inside the company who answers for the result.
Three questions for any agentic pricing platform
What does each agent do, and what does its recommendation rest on?
Ask for the list of agents and the job each one has. Ask what an agent calls when it needs a number: a fixed computation, or its own arithmetic. Then ask what each recommendation rests on.
Start with the competitive intelligence. One kind is collected from secondary sources: pricing pages, review sites and anything else an agent can gather from the public web. Another comes from live conversations with competitors’ customers, who know what they were quoted, what they were refused and why they stayed. Ask which kind the platform runs on.
Then ask how the firm identifies itself when it gathers intelligence. A firm that approaches a competitor under a false pretext hands you evidence you cannot stand behind, which is the case for ethical competitive intelligence. Getting a competitor’s confidential pricing by misrepresentation can also break trade secret law, and the company that uses what was gathered can share the liability.
Ask the same of willingness to pay. Willingness-to-pay surveys and simulated buyers report what a buyer says about a hypothetical price. Interviews with live customers explore value, return on the purchase and how the license should be structured, with people who have already bought.
Then ask what the analysis is compared against. It needs a pattern library: a record of how software pricing decisions played out elsewhere. If the agent reads your deals without one, it starts with no software-specific patterns, and the model’s general training fills the gap with what every other company is already doing. The AI Pricing Observatory is the dated record of those moves and what followed them.
A pattern library also holds how customers and salespeople behaved when a price, an edition or a license term changed. Without one, that record is built by hand, one client at a time. When a competitor changes its value metric, a platform with a pattern library can show how similar changes played out before. Without one, it can only report that the change happened.
An agent with no pattern library gives you no more than a general-purpose AI assistant your team already pays for. Last, ask whether the firm behind the platform prices only software, or prices software among many industries.
Which changes reach a customer without a person approving them?
The answer a buyer wants is none. If the answer is “routine ones” or “changes inside set bounds,” ask who set the bounds and when they were last reviewed. Ask what the platform did the last time a change turned out wrong.
Ask who on the vendor’s side reads a recommendation before it reaches you. Ask whether those reviewers have run pricing, product or finance inside a software company, or have only advised such companies. Someone who has owned a renewal number reads a recommendation for what it will do to the sales team and the next renewal.
After a change ships, how do we learn what it did?
Ask to see the tracking for a past change: which accounts it touched, what they did next, and how the result compared with the forecast.
Then ask what outcomes stand behind the platform: results the firm has carried through with software companies, which you can check. When the AI sits on a record of outcomes, it is helping your company reach one. Without that record, what it enables is discovery, and in our experience discovery alone seldom ends in an outcome.
How LevelSetter separates agent analysis from human pricing decisions
LevelSetter is a pricing decision platform: the decision layer where a B2B software company’s licensing, packaging and pricing decisions are made and validated against its own deals, above the systems that execute them. It is an agentic pricing platform built for B2B software: people make the judgment, and agents extend what those people can see.
Under the agents sits a deterministic engine that does the computing. The agents do the analysis: patterns, trends, competitor moves and the impact of each change by cohort. The platform can recommend a change to licensing, packaging or pricing, and a person reviews and approves every one. A company runs LevelSetter with our experts, or we run it for them, and it can switch between the two as its capacity changes.
Human judgment leads. Our experts bring their own judgment and the pattern library to each question, and the agents extend what they can see. Sometimes an agent finds a pattern no person would have found. Other times a person does more because the agent did the reading.
Nobody here rubber-stamps an agent’s recommendation, because today an agent cannot see everything a person can. Blending the two works the way blending qualitative and quantitative insight does. We are not building toward a world where the machine changes pricing.
Our own evidence comes from closed deals and from live conversations with customers and with competitors’ customers. Those conversations are how we do competitive intelligence and how we research willingness to pay. We run no surveys, and we approach no competitor under another name.
Those conversations also feed LevelSetter qualitative insight that no deal record holds. The experts who hold them are trained to listen for how a buyer says something, which in our experience tells more than the words a transcript keeps. The pattern library holds interviews with every function in a software company, and with its customers, its competitors’ customers and its partners, through each change in how software is sold. None of it was ever published, so no model was trained on it.
The analysis is read against a pattern library built from software and nothing else, because software is the only business we price. The experts who review each recommendation have run parts of software companies from the inside. The platform sits on the outcomes of that work.
This is how Continuous Monetization runs in practice: the pricing decision layer stays current, and each decision stays deliberate.
If you are weighing an agentic pricing platform, or a vendor’s claim to be one, talk to a pricing expert.
