The question

Most business software is sold per seat. A seat is a licence for one named person, billed monthly or yearly, so the bill tracks head count. AI agents, meaning software that carries out a multi-step task on its own, cut the cost of the work those people did. The bill and the work have come apart.

The usual forecast says the seat is finished and usage pricing wins. Usage pricing means charging for the work the software does, not for how many people can log in. Two prices matter here. The list price is the vendor’s own public sticker price. The price actually paid is what one customer hands over after discounts, credits and private side agreements.

That second number is confidential, and we call the gap between the two the secret. So the question: when does the price get reset, and who pays for the reset?

Our answer, after one live game and a hundred re-runs: seat against usage was never the deciding question. What decided the game was whether the secret ever became public, and who held the renewal contracts that day. Every large player had a reason to keep the secret, and in most runs they did.

How the simulation works

Four roles, each played by an AI following a written brief, with an AI referee that set the odds. We call the four roles the AI players, to keep them apart from the AI agents the report is about. No person refereed the game while it ran.

The AI players played five turns, late 2026 to the end of 2028. We then built a rule-based version with fixed rules and no AI, so the same dice always give the same result. We ran it a hundred times with different dice over all eight turns to 2030, and one full play-through is a run. The author wrote 2029 and 2030 from those hundred runs. No AI player played them.

Salesforce, Microsoft, Klarna, and the research firms Gartner and MIT are real, and each carries a source in the deep mode. Atlas, Forge, Meridian and Pillar exist only in the game. Nothing here is a forecast.

The incumbent kept its list price and gave everything away in private

Atlas opened by renaming the product it already sold as an “agent-inclusive” package, at the same price. Two large customers had run agent trials, tests of whether an AI agent could do a real job before they paid for it. The trials showed one workflow needing half as many licences. Atlas renewed both at full list price before either could argue for a discount.

What Atlas gave away never appeared on an invoice. It was a better support tier, a delay on the bill for use beyond the contract, and the renamed package. Atlas knew the risk, and wrote on its opening move that the side agreements’ confidentiality “is not perfectly enforceable”. Atlas went ahead because the renewal was this quarter and any leak would be some other quarter. The referee set a low bar, and Atlas cleared it four turns running before it missed.

The real market does the same thing, harder. Salesforce, whose business is per-person licences, raised its list price by an average of 6 per cent in August 2025. It did so while shipping agents into every product, and it sells those agents by consumption at about ten US cents an action. Microsoft holds its Copilot assistant at US$30 per user per month and sells agent credits at a penny each. Both run both price models at once, decided renewal by renewal, in private.

The challenger’s price comparison was right and nobody would sign it

Forge’s move was to publish a sheet comparing its outcome pricing with what Meridian spends on per-seat licences. Outcome pricing means charging only when the software finishes the job, and the cost of unused licences was the difference on the sheet. Then get one flagship customer, a large and well-known buyer, to sign the sheet in public. Outcome pricing is real: Intercom’s support agent charges 99 US cents only when it resolves a ticket. Forge published five times, and no customer ever signed.

The failures were not bad luck alone. The referee set the bar at 16, 16, 18, 19 and 19 out of 20 across the five attempts. Forge had put its own odds at 38 to 52 per cent, and the dice came in at 9, 11, 8, 11 and 10. Each time the same move came back, the referee raised the bar. On the fourth attempt it wrote that the repetition “compounds an already-elevated baseline rather than starting fresh”, meaning a move the market has already declined starts from a harder position.

Publishing needs a large customer willing to give its name, and three things push against that. The buyer’s own job is at risk if the deal goes badly. Atlas sends a private counter-offer within days. And procurement, the department that buys software, prefers to use its bargaining power quietly. Forge saw it by the last turn: the pitch convinced people, and getting a company to attach its name did not.

The buyer kept its audit private, and failed trials bought the vendor time

Meridian counted its licences and found the ones nobody used. A built-in event set that at 40 per cent of licences paid for, defined as no login in 90 days. The measured real-world figure is worse. Zylo, which tracks software spending, finds the average organisation leaves 46 per cent of its licences unused. That wastes about US$19.8 million a year on those licences alone.

Meridian then banked the finding privately, and never published it. It reset contracts to the licences actually used, switched off the idle ones, and ran a hard renewal with a rival quote on the table. Publishing would give every other buyer a number to point at, which helps the market. Keeping it private makes it bargaining power at this renewal, for this finance chief. We gave the buyer that motive on purpose, because buying teams are paid on this year’s savings.

The referee stopped rewarding the audit by its third repeat, and the fourth attempt cost Meridian points. The fifth turn is the one to copy. Meridian stopped auditing and forced a contract reset on its largest per-seat agreement, and it landed. Same evidence, same quarter, and this time the number had an invoice attached.

The trials bought Atlas time too. MIT’s Project NANDA found 95 per cent of organisations seeing no business return on generative AI, the broad family of AI that agents come from. Gartner forecast that over 40 per cent of agent projects would be cancelled by the end of 2027. Klarna, the payments company, said its AI assistant did the work of 700 support staff, then went back to hiring people. In the game every failed trial reset trust inside the buyer, and Atlas only had to argue that this renewal was not the moment to switch.

In a hundred re-runs the secret mostly held and nobody collapsed

One game is an anecdote, so the rule-based version ran a hundred times with different dice. Rank the hundred runs by Atlas’s final score and the fiftieth is the median run. Atlas ended the median run at 95 from a start of 100. Ninety runs ended between 82 and 112, with five below and five above. The worst single run landed on 72 and the best on 121, and not one run drove Atlas out of the market.

Grouping the runs by how they unfolded gives two families, and both are versions of one sentence: Atlas holds. Seventy-two runs hold by the rename and twenty-eight by early lock-ins. A lock-in is a multi-year contract that stops the customer repricing until it ends. The secret went public in a minority. A customer published the real price it paid in 16 runs, and a buyer called the rename out in public in 26.

Where the secret did end, a customer ended it: a large, well-known buyer already on its way out of the relationship. One published customer reprices nothing by itself. It turns every private number into a floor. Atlas’s renewal talks had opened at the list price and discounted down. Now they opened at the published price and argued upward. Figure 1 draws that chain.

The window closes on a settlement, not a collapse. The median run ends with Atlas at 95 from 100, Meridian at 79 from 70, Pillar at 63 from 60 and Forge at 36 from 30. The challenger gained the most in proportion and is still the smallest thing on the board. Nobody went under and nobody won.

We then tested one decision on purpose. In all hundred runs we made Atlas launch usage pricing at the start of 2027. In the unforced game Atlas holds its licence income first and switches later, under pressure, so we compared each forced run against the unforced run with the same dice. The early cut lost 57 of the 100 pairs, won 39 and tied 4. The median cost was one point of final score, and half the runs sat between seven points worse and two better.

What it means

The reusable idea is not about software. When both sides of a deal gain from keeping the real price private, the price does not reset when the technology arrives. It resets when one of them publishes. So watch for who has a reason to publish, and pick which of the four players is closest to your own job.

What this does not prove

One referee rule does most of the work: a move that already failed starts from a harder bar. Switch it off and the challenger probably lands one attempt in five, the audit becomes a contract reset sooner, and the report says something else. We kept it because a pitch the market has already declined does age badly. It is a rule we built into the game, not a verified fact about markets.

No human refereed, and the AI made one plain error we left in the record. From its first move Forge described its comparison as exposing “Meridian’s incumbent pricing”, as if the buyer were the vendor. The referee built on the slip for four turns. It did not change the outcome, because Forge failed every roll either way. It is the strongest single argument for a human referee.

“Atlas never collapsed” is partly a property of the board. Atlas starts at 100 and a player is out at zero. Its heaviest single loss is six points, and rival moves take two to four more. So reaching zero in eight half-years needs close to worst-case dice on almost every turn. Read it as: eight half-years is not long enough to kill this incumbent on this board.

Three more limits. The AI players played only five turns, and the author wrote the last three from the hundred runs. The scores, start values and dates are ours, and the dates are the least defensible thing here. The evidence also cuts against the game twice, and both cuts put more pressure on the secret than the game did. Salesforce raised its list price while Atlas held flat, and the real unused-licence rate, 46 per cent, is worse than the game’s 40.

What to watch

Four facts on the public record, and one indicator.

The indicator is close to zero today, which is why the rename works. The day it stops being close to zero, every renewal conversation you have planned needs rewriting.