What Has to Change Before AI Agents Can Buy Software

TL;DR Agent-run evaluation is already happening. Agent-completed purchase arrives later than its advocates say, and four conditions explain the gap. Disclosure: an agent computes a defensible cost from published terms, and the terms that decide what a deal costs are the ones vendors have always held back. The buying side: procurement functions need the mandate, the training and the authority to act on what an agent returns, and the tooling is arriving ahead of all three. Variance: net prices for the same software differ widely from account to account, so publishing a computable price exposes the variance itself, and a most favored nation commitment turns a readable number into a claimable one. Ordinary inertia does the rest. The work still pays before the channel matures, because the properties that make a price computable are the same ones that let a rep hold it in a live negotiation.

Software reading a pricing page, projecting a year of usage, and ranking a field of vendors is work that happens today. It happens at scale, and it finishes before anyone at the vendor knows an evaluation opened. Agentic buyer pricing covers what that software retrieves off a published surface and what only the decision layer behind it can answer.

The evaluation gate has moved. The transaction gate has not, and the distance between them is where a defensible forecast lives. Four conditions have to change before AI agents can buy software:

1. Disclosure. Publishing the rule that turns a list price into a net price, which vendors have always held back.
2. The buying side. Procurement functions holding the mandate, the training and the authority to act on what an agent returns.
3. Variance. Net-price differences across the installed base that a published, computable price would expose, including to customers holding a most favored nation commitment.
4. Inertia. The ordinary kind, which nobody schedules and no forcing event has arrived to break.

The first condition is disclosure, and it goes deeper than a price list

Disclosure is the requirement that a vendor publish the rule turning a list price into a net price, and that rule is the part of pricing the industry has never made public. An agent computes a defensible annual cost, and a list price alone will not produce one. It takes the rule itself: how a commitment is credited, which reductions reach which buyers, what a longer term earns, where a volume threshold sits and what happens on either side of it.

Those are precisely the terms vendors have always held back. Published price lists have been normal in parts of this market for decades. A published account of how a net price is arrived at has rarely appeared anywhere, in any software category, at any size. How much pricing detail should appear on your website takes the page itself. This condition sits one layer under the page, and it asks a vendor to publish the shape of a calculation the industry has treated as private.

A vendor cannot publish a calculation it does not perform

The second reason disclosure stalls is the harder one, and it changes the sequencing of everything else.

Where discounting has been a sequence of decisions rather than a discipline, the net price on a signed deal is a residue. It records the rep’s authority, the week of the quarter, the buyer’s patience and whatever the deal desk approved that afternoon. Pricebook deviation is the measured form of that gap, the distance between what the pricebook specifies for a configuration and what closed deals record. A vendor carrying systematic deviation has no rule available to publish, because its pricebook is not what governs its prices.

That puts architecture ahead of disclosure in the order of operations. Margin-calibrated discounting is the architecture-side answer: a pricing surface that produces a scheduled net price at every commitment, so the number a buyer computes and the number a rep lands are the same number. Until a vendor has that, the disclosure question is premature, because there is nothing coherent to disclose.

The test to run against your own surface: if a buyer computed an annual cost for a named configuration from what you publish today, how far from your last five closed deals at that configuration would the answer land?

The second condition is the buying side, and the tooling is arriving first

An agent-run evaluation needs a buyer on the other end who is permitted to act on it.

Procurement functions are staffed, chartered and audited around human diligence, so an evaluation returned by software raises exactly the questions those functions exist to answer: who verified it, what happens when it is wrong, and whose signature stands behind a recommendation nobody produced by hand. A procurement lead who forwards an agent’s shortlist carries accountability for it that no policy has defined yet.

So the sequence inside the buying organization runs: tooling, then practice, then training, then a mandate to rely on what the tool returns, then the authority to transact on it. Tooling is the fast half. The remaining four are the slow half, they move at the speed of audit committees and job descriptions, and no vendor can accelerate them from outside. This is the condition a vendor has the least purchase on, and it is the strongest single reason to place this channel later rather than sooner.

It also sets the test for anyone reading a confident timeline: does the forecast describe what the software can do, or what a buyer is allowed to do with it?

The third condition has teeth: variance, and who can claim against it

Net prices for the same software at the same configuration differ widely from one account to the next. Peer-reviewed research on enterprise software sales documents the same pattern, with the timing of a rep’s compensation moving the discount on deals that are otherwise comparable. Some of that difference is architected and traces to structure a buyer can see. Customer Groups derive value differently, editions carry different capability, and the value metric scales the bill with value received. The remainder is variance nobody chose.

A computable published price prices the whole installed base, not only the next deal, which puts it in the same category of decision as moving existing customers to new pricing. Every existing customer can compute where their own agreement sits against it. Where the variance is architected, that comparison has an answer, because the structure produced it and the structure explains it. Where the variance is accidental, no answer exists, because none was ever designed.

Most favored nation commitments make the exposure concrete. A customer holding one has contracted for a defined relationship between their price and what other customers pay. A number an agent can read is a number that customer can measure their own agreement against, and they hold paper describing what follows. Whether a particular commitment reaches a particular published number depends on how it was scoped, what it covers, and what the published figure represents. That is diagnosis work against specific agreements and a specific installed base, and it belongs with a pricing expert reading the paper alongside the pricebook. If that describes your base, describe the situation and a pricing expert will reply.

The timing point is narrower and holds regardless: a vendor with these commitments in its base cannot treat publishing a computable price as a marketing decision. It reaches into signed contracts, and the reach has to be understood before the number goes up.

The fourth condition is inertia, which is ordinary and real

Inertia is the condition with no owner. Pricing pages belong to marketing. The pricebook belongs to finance. The discount rule belongs to sales leadership. The paper belongs to legal. A change that touches all four has no single owner and no forcing event, and nothing about agent buying supplies one yet, because the buyers who would apply that pressure are still working through the slow half of their own sequence.

What usually breaks that pattern is a competitor publishing first and winning evaluations the incumbent never sees. That has not happened at any scale in this market, so the cost of doing nothing stays invisible and the change stays unscheduled. A vendor waiting for the forcing event is waiting for a competitor to create it.

Inertia is a poor reason to dismiss the channel. It is an excellent reason to distrust any timeline that treats vendor readiness as a publishing exercise.

Why the work pays before the channel matures

Add the four together and this channel arrives later than its advocates say. The case for doing the work now survives that timing, on grounds that have nothing to do with agents.

A machine-readable surface is a defensible surface

A pricing surface a machine can read is a pricing surface a salesperson can defend, because the properties are identical: a value metric the buyer can count without asking, editions with boundaries that hold under pressure, and a rule that produces the same net price for the same commitment on Tuesday and on Friday. A rep who can answer “why is this the price” without escalating is working from the same artifact an agent would read.

The exposed gaps are costing deals now

A capability sitting behind the wrong edition boundary reads as a capability you do not offer, whether the reader is a person or a program. A cell that says contact sales records as a hole in a comparison either way. Ambiguity about what an active user means stalls a human evaluation as reliably as it stalls a machine one, and it stalls it silently, which is why it rarely reaches a loss report.

Machine-readable pricing is the architecture underneath all of this, and agentic AI pricing strategy is the separate question of pricing the agents you sell. The agent channel puts a deadline on work that already pays for itself. Build the pricing surface because your own sales motion needs it, and let the arrival of the channel decide when, and how much of it, you publish.

None of the four conditions argues for waiting. They argue for doing the architecture work on its own merits, so that when a buyer’s agent does arrive at your pricing page, what it finds is a surface you designed rather than a record of every deal you ever negotiated.

If software is already reading your pricing page and nobody has checked what it comes away with, describe your pricing situation and a pricing expert will reply.

FAQs