When AI Agents Buy for You, the Seller’s Agent Defines Your Needs

← Back · ACT II · Chapter 14: Who Controls Value Engagement?

Few people carry more standing in pricing than Tom Nagle. He wrote The Strategy and Tactics of Pricing, he built the economic value estimation framework that most serious value modeling still descends from, and he spent a career on a single question: how do you align what you charge with what you actually create for the customer.

In his foreword to a new book on pricing for the agent economy, he sets down where he thinks this goes. If the book’s vision holds, he writes, value models will become a critical machine generated and machine readable artifact that connects agent-to-agent commerce.

He is almost certainly right. That is the problem.

The Artifact Already Exists

This is not a forecast waiting on adoption. There is now a published open specification for computable value models and pricing models, at version 1.0.0, released under Apache 2.0 for the schemas and Creative Commons for the prose, explicitly designed for autonomous agent consumption and declared compatible with Google’s A2A protocol and Anthropic’s MCP.

It defines a value model with reference value and value drivers across revenue enhancement, cost reduction, risk reduction, operating capital reduction, capital investment reduction, and option expansion. It carries confidence intervals. It computes value surplus as total economic value minus price. It ships a pricing model schema alongside it covering flat, per-unit, tiered, volume, and metered pricing, with machine-executable equations and deterministic invoice output.

The engineering is careful and the schemas are well built, and what they describe is a vendor catalog. The spec says so in its own words: complete vendor pricing catalog, deal configuration, and deterministic invoice output. Its governance page is equally direct about the authorship: the specification was initiated by a value-modeling software vendor, that vendor holds every decision until ten production implementations exist, and the license carries permanent attribution to it as originator. Anyone can use the schema for free, which is a separate question from who decided what it measures.

The Date Is Already Set

Gartner has predicted that by 2028, ninety percent of B2B buying will be intermediated by AI agents, routing something on the order of fifteen trillion dollars in spend through agent exchanges. Treat that as what it is, an analyst forecast rather than a measurement. The direction does not depend on the number being exact.

The specification is candid about what that means for anyone without a conforming artifact. Agents do not accept PDFs. They evaluate on structured, machine-readable data, and they auto-reject what they cannot parse. A value model in a spreadsheet is invisible. Invisible means disqualified.

That warning is aimed at sellers. It applies with more force to buyers, and no one has aimed it there.

What Does The Buyer’s Agent Carry?

Picture the transaction the protocol is built for. The seller’s agent arrives holding a validated, computable, confidence-interval-bearing model of the buyer’s economics. Its value drivers are named, its equations will run on the buyer’s own numbers, and it has already worked out the surplus.

The buyer’s agent arrives holding a budget, an approval workflow, and a preferred-vendor list.

There is no buyer-side machine-readable declaration of the outcome. None exists as a category. The buyer has no schema that says what this organization counts as an outcome, what baseline stands, what evidence is admissible, and what result would constitute failure. So the only structured model of the buyer’s value in the room is the one the seller brought.

Both agents will then compute against it, agree, and settle. Very quickly.

What Automation Removes Is The Friction

This asymmetry is old. Buyers have always evaluated vendors using the vendor’s own model of the value. What has protected buyers, badly and inconsistently, is friction. Slide decks get argued with. Business cases sit in committee. A skeptical finance director asks where a number came from and the meeting stops.

A conforming schema does not stop. It validates, and validation is not agreement. It confirms that the model is well formed, not that its assumptions are the buyer’s assumptions, that its baseline is the buyer’s baseline, or that its definition of an outcome is one the buyer would have chosen.

Remove the friction and the asymmetry does not merely persist. It executes at machine speed, deterministically, across every transaction, with an audit trail that makes it look rigorous.

The Counterpart Artifact

Nothing here is an argument against credit-based pricing, against machine-readable models, or against the specification. The engineering is sound and the direction is correct. A protocol for value is overdue.

The argument is that a protocol with one side declared is not a protocol. It is a broadcast.

The buyer needs its own artifact, and it needs it before the agents transact, because after they transact there is nothing to negotiate. What this organization counts as an outcome. What baseline it will measure against. What evidence it will accept as proof. What result it will call a failure and what it will do then. The buyer writes it, keeps it, and puts it in front of the seller’s model as the test that model has to pass rather than the test it gets to set. I have made the underlying argument before: the buyer sets the definition. The same problem reappears after the purchase, when the software begins computing its own return. There is a version of the same problem about the venue rather than the schema, where the seller convenes the conversation and keeps the record of what was said, and that one is worked through in the buyer’s system is the table.

The seller’s model is a proposal. It should be tested against a standard the buyer wrote. Without that, it is not a proposal at all, because there is nothing on the other side capable of refusing it.

Whoever supplies the only machine-readable definition of value in the transaction has already decided it.

Sources: Tom Nagle, foreword to Pricing for the Agent Economy by Steven Forth, Michael Mansard and Wolfgang Ulaga, quoted from the text as published by one of the book’s authors; Nagle and Holden, The Strategy and Tactics of Pricing, origin of the economic value estimation framework; thevalueproject.org v1.0.0, open specification for computable value and pricing models, schemas Apache 2.0 and prose CC BY 4.0, stated compatible with Google A2A and Anthropic MCP, read in full 11 August 2026; Gartner press release, 21 October 2025, top strategic predictions presented at IT Symposium/Xpo, forecasting that ninety percent of B2B buying will be AI agent intermediated by 2028 with over fifteen trillion dollars routed through agent exchanges. The Gartner figure is a forecast, not a measurement.

Keep reading: ACT II · Chapter 14: Who Controls Value Engagement?