The Seat Price

This is a report on a wargame. A wargame is a structured argument played under rules: each player has a role, a goal and a menu of moves, and dice decide whether a move works. Four AI players took four roles in the enterprise software market. They played the first five of eight half-year turns, and I wrote the last three myself from the hundred re-runs described below. The window is 2026 to 2030.

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. So when does the price get reset, and who pays for the reset?

The usual answer is a funeral: the seat is dead and usage pricing wins. Usage pricing means charging for how much work the software does, in messages, actions or credits, instead of for how many people can log in. Two prices matter here. The list price is the number the vendor publishes. The price actually paid is what one customer hands over after discounts, credits and private side agreements. That second number is confidential, and this report calls the gap between them the secret.

What decided the game was whether the secret ever became public, and who held the renewal contracts on that day. It was never seat pricing against usage pricing. Every large player had a reason to keep the secret, and they did. Where a reset does arrive, my reading is that it looks like hybrid pricing: the seat kept for the people who still log in, the agents' work billed by the action. The 2029 H1 beat shows that no player tried it.

How this one was made, plainly. No human refereed this game while it ran, unlike our previous wargame, where a person set the odds. I call the four roles the AI players, to keep them apart from the AI agents this report is about. A turn is one half-year of game time: a player names a move, an AI referee sets a number the dice must beat, the dice decide. The transcript calls that referee the arbiter. A second, rule-based version of the same game then ran one hundred times. It applies the same rules to the same situation every time, so only the dice change from one run to the next. One complete play-through of the eight turns is a world, and the hundred together are the ensemble. I set the question, read every transcript and wrote this account.

Four players. Atlas, a very large software company whose income is mostly per-user licences. Forge, a smaller AI-native rival that charges per job completed rather than per user. Meridian, a large corporate buyer, run by a finance chief paying for licences nobody logs into. Pillar, a global consulting firm paid to install and connect enterprise software. From here on I also call Atlas the incumbent, Forge the challenger and Meridian the buyer.

A beat is one dated section of the timeline below. A number with a bracketed marker, like 3, is a real-world fact checked against the evidence pack at the back. Every other number is the game's own record or a figure I chose, and the sentence says which.

What follows is written as a memory, from 2030 looking back. None of it has happened.

The half-year dates on these headings are the game's own calendar, set by me. A rebadge means putting an AI label on the product you already sell without changing the price. Atlas opened with one, and it worked, because everything Atlas gave away it gave away in private.

Two large customers had run agent trials, meaning 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 they could use that finding to argue for a discount. In exchange it handed over concessions that never appear on an invoice: a better support tier, a delay on the bill for using the software beyond the contract, and a renamed "agent-inclusive" package. Nothing was announced. The referee set a low bar and Atlas cleared it four turns running, rolling 23, 22, 22 and 13 against 11, 10, 12 and 13, before missing the fifth by one.

The real market is running the same harvest, harder. A vendor harvesting a seat model lifts the published number and gives the real price away underneath it, one renewal at a time, so every private discount is measured from a higher start. Salesforce, whose business is per-person licences, raised its list price by an average of 6 per cent across its Enterprise and Unlimited editions, effective August 2025, while shipping agents into every product 3. That rise is the harvest move made in public, by the vendor with the most seats to harvest. Salesforce sells its agents two ways at once: at US$2 per conversation, an option still on its price list 1, and through prepaid credits that work out at roughly ten US cents an action 2. The best-known per-seat AI product is Microsoft 365 Copilot, at US$30 per user per month 4. Microsoft also sells agents at a penny a credit 5. Both real vendors run both models at once and let each renewal decide which applies, in private.

Atlas knew what it was building. Listing its assumptions on the opening move, the player wrote: "Confidentiality of the side-letter terms is not perfectly enforceable" (Atlas, turn 0). A side letter is a private agreement attached to a contract, changing the terms without changing the price list. Atlas went ahead, because the renewal was this quarter and any leak would be some other quarter.

The referee called the move "orthogonal to the actual battlefield" (arbiter, turn 0), by which it meant public, provable pricing based on results, and let it through regardless. Atlas had built a machine for harvesting an old price model, meaning taking as much revenue from it as possible before it ends. It ran longer than anyone expected.

Forge's plan was to publish a price comparison and get one flagship customer, meaning a large and well-known buyer, to put its name to it. Forge published five times. No customer ever signed, and that failure is the spine of the whole game.

The move: publish a sheet setting Forge's outcome pricing against what Meridian spends on per-seat licences, with the cost of unused licences as the difference. Outcome pricing means charging only when the software finishes the job. It is real and cheap to point at: one live agent charges 99 US cents only when it resolves a support ticket 15.

The arithmetic was never the hard part. Publishing needs a large customer willing to be named, and every force at the table pushed against that: the risk to the buyer's own job if the deal goes badly, Atlas's private counter-offer arriving within days, and the habit of the buying department, called procurement, of using leverage quietly. Forge's sheet was right every time, and nobody signed it.

Across Forge's five attempts the difficulty went 16, 16, 18, 19 and 19, out of a possible 20, against Forge's own stated odds of 38 to 52 per cent. The dice came in at 9, 11, 8, 11 and 10. On the fourth attempt the referee wrote its rule out plainly: "the repetition compounds an already-elevated baseline rather than starting fresh" (arbiter, turn 3). In plain words, a failed move starts from a harder bar. Forge had the diagnosis a turn later: "the brief itself lands, the logo commitment is the harder ask" (Forge, turn 4). In plain words, the pitch itself convinced people, and getting a company to attach its name in public did not.

The pool of real challengers is thinner than the noise suggests. Of the thousands of vendors claiming to sell AI agents in 2025, Gartner estimated only about 130 were real. It named the rebranding of chatbots and screen-automation software "agent washing" 11. In the game, every fake challenger that failed in public made the real one's pitch harder to believe.

An audit here means counting licences to find the ones nobody uses. Meridian ran that count, found the waste, then kept the finding to itself.

The game has built-in events, meaning things we scripted to happen regardless of what the players chose. This one put unused licences at 40 per cent of those paid for, measured as no login in 90 days. One industry index says reality is worse. Zylo, which tracks software spending for its customers, found the average organisation in its 2026 index used 54 per cent of its software licences, so 46 per cent sat unused, wasting about US$19.8 million a year 9. The evidence pack at the back does not say which companies that average covers. This report assumes they are Zylo's own customers, companies with a licence bill large enough to pay someone to track it, so the dollar figure is a large-company average. Carry the 46 per cent to a smaller company and leave the dollar figure behind.

Meridian banked the savings privately. It forced true-ups, meaning resets of a contract to the licences actually used, with money moving to match. It switched off dormant licences and ran a hard renewal with an agent vendor's quote on the table. What it never did was publish the arithmetic. Publishing the finding would give every other buyer a number to point at, and that helps the market. Keeping it private makes it bargaining power at this renewal, for this finance chief's own bonus. That last clause is a modelling assumption: we gave the buyer a personal reason to prefer the private number, because buying teams are usually paid on this year's savings. Buyer and vendor are opponents on paper, and on the thing that mattered they wanted the same outcome.

By the third repeat the referee had stopped rewarding it, ruling that the audit "risks reading as MERIDIAN treading water instead of converting prior audit findings into the vendor renegotiation it keeps deferring" (arbiter, turn 2). In plain words, Meridian kept finding the same waste and never used it to force a better contract. On the fourth attempt the audit failed badly enough to cost Meridian strategic position, which is the single score the game keeps for each player and moves up or down after every turn. An audit that never becomes an invoice line is a press release to yourself.

The fifth attempt is the one to copy. Meridian stopped auditing and forced a true-up on its largest per-seat contract, and it landed. Same evidence, same quarter. This time the number had an invoice attached.

A pilot is the same thing as the agent trials above: a small test of new software on real work before a full rollout. The trough is the slump after the hype, when those tests fail and budgets get cut. This beat is the counterweight to everything above: the repricing argument assumes agents work, and in 2025 most did not.

MIT's Project NANDA found 95 per cent of organisations seeing no business return on generative AI, against US$30 to 40 billion of enterprise spending 12. Gartner forecast that over 40 per cent of agentic AI projects would be cancelled by the end of 2027, blaming rising costs, unclear business value or weak risk controls 10.

The most famous case of an agent replacing staff ran its arc in fifteen months. Klarna announced its assistant was doing the work of 700 full-time customer service staff 6 and put a US$40 million profit improvement on it 7. Then it reversed, resumed hiring people, and its chief executive conceded that when cost is the main test, "what you end up having is lower quality" 8.

In the game this is the burned pilot: a failed trial that resets trust inside the buyer and hands Atlas another year. Atlas's renewal staff never had to argue that agents will never work, only that this renewal is not the place to bet the workflow. Through 2028, in most worlds, that argument cleared the bar. Every burned pilot re-arms the rebadge.

The fog is the confusion about which pricing model will win. A systems integrator is a consulting firm paid to install and connect enterprise software. Pillar is one, and it spent the whole game staying neutral and selling that confusion, because an unsettled market keeps paying for advice.

A vendor-neutral scope of work here, a shared cost baseline there, positioned as the layer under whichever pricing model won. Pillar ran that move five times, and the referee priced each repeat higher, exactly as it did Forge's. The difficulty went 8, 8, 10, 12 and 14. The first two hedges missed, the middle two landed, and the fifth is where the bill arrived.

The fifth hedge was the same offer, made in public for the first time. Pillar said the broken layer was per-seat licensing and its own implementation work was sound, and again offered to build the repricing framework for whichever player won. It failed. The referee ruled that Pillar was offering to cooperate with Atlas while calling Atlas's model structurally flawed, "a real tension the proponent did not surface or address" (arbiter, turn 4). Five turns of neutrality had bought nothing Pillar could spend on the one turn it wanted to be heard.

Selling the transition is real revenue, because somebody has to wire the agents in. But the confusion ends on the day the pricing question is settled, and an integrator that sells nothing else has nothing left to sell.

Total software spending goes up in this beat while per-seat licences lose ground. Both are true at once, and the standard forecast misses that.

If the seat were dying the way the funeral notices claim, the software line in company accounts would shrink. It grew. Worldwide software spending grew 11.9 per cent in 2025 to US$1.24 trillion, and was forecast to grow another 15.2 per cent in 2026 13. An analyst at Gartner, a research firm that tracks the technology industry, said why. AI features built into the software companies already own are pushing its cost up 14.

What is collapsing is the unit the vendor charges for. In one 2026 survey of 230 business software and AI companies, three in four had changed their pricing or packaging inside a year. In the same sample, hybrid pricing, meaning seats and usage sold to the same customer at once, went from 25 to 37 per cent adoption in twelve months, and AI-credit schemes grew 126 per cent year on year 16. The seat is being replaced by a menu of seats, credits, actions and outcomes. That menu makes the price one buyer pays impossible to compare with the price another pays, and a market that cannot compare prices cannot reset them.

My reading is that hybrid pricing is the reset this report keeps looking for. A seat-plus-usage contract keeps the licence for the people who still log in and bills the agents' work by the action. The vendor reprices the work one contract at a time and never cuts the list price, so a hybrid contract belongs to Ending A below and never forces Ending C. The game record holds no such move. The rule-based version gives Atlas four moves, and the only usage-shaped one is the pure pivot the forced fork tests. In the live run no player proposed a seat-plus-usage contract. Pillar named hybrid once, in turn 4, as one of three models its neutral framework would serve. So the players never tried the reset the survey says the market is choosing, and the seams list that gap.

By this point in most worlds, Atlas's list price had not moved in six turns, while the price its customers actually paid had fallen throughout. The dashboard beside this beat tracks those two lines pulling apart, and the figure under Ending A draws them against each other.

A reference price is a price one customer paid that other customers can point at, and it exists only once it is public. In this beat one customer publishes what it really pays and what switching would cost. That is the end of the secret, and in the worlds where it ended, a customer ended it.

The pressure builds by accumulation. Every private side letter is a document somebody can leak, and every audit is a number some board member has seen. One built-in event is an analyst story naming seat collapse as the risk of the year, which lowers the career risk of publishing first. The customer who breaks first is a large, well-known buyer already on its way out of the relationship.

The rule-based version has a named event for this move, and across the hundred worlds it fired in 16. A second event, a buyer calling the rebadge out in public, fired in 26. I cannot tell how far the two overlap, so something goes public in at least 26 worlds of the hundred and at most 42. In the rest the secret holds to 2030. This branch is not the base case. I have written the timeline as though it happens because it is the branch that changes what you should do on Monday.

One published customer reprices nothing by itself. It turns every private number into a floor for the next negotiation. Before, Atlas's renewal conversations opened at the list price and discounted quietly down. Now they opened at the real price that customer had published, and argued upward. Same contracts, same product. What changed was who knew what.

The window closes on a settlement, not a collapse. Atlas's harvest ends because buyers gather enough information to end it. No rival beat them.

Atlas, in the middle world of the hundred, ends at 95 per cent of the strategic position it started with. The seats that survived are welded to work agents cannot do. The rest of Atlas's estate has been repriced through credits, bundles and true-ups, almost none of it through the list price. The repricing landed where the bargaining power sat at each renewal, unevenly and quietly.

Forge was eliminated in none of the hundred worlds. It ends small and important: the outside option that disciplined every renewal it never won. Meridian's licence bill is smaller, but not by as much as its audits had predicted. Pillar sold the transition and owns what it built before the question settled.

Against their 2026 H2 starting positions, the hundred worlds end with Atlas at 95 from 100, Meridian at 79 from 70, Pillar at 63 from 60, and Forge at 36 from 30. The starting numbers are a scale I set for each player's market weight in 2026, and nobody measured them. The buyer gained the most points, nine. The challenger gained the most in proportion, a fifth, and is still the smallest thing on the board. The incumbent lost five. Nobody was destroyed and nobody won. That is what a repricing looks like when it is negotiated one renewal at a time instead of announced. The seat did not die. The secret did.

One game is an anecdote, so we ran the rule-based version one hundred times: same roles, same menu of moves, different dice.

Atlas's final strategic position runs from 82 at the 5th percentile to 112 at the 95th, with a median of 95 against a start of 100. Read that as: five worlds in a hundred ended below 82, five ended above 112, and the middle world ended at 95. The worst single world landed on 72 and the best on 121. Not one world produced a collapse: no run drove Atlas out of the market inside the window.

Grouping the hundred worlds by how they unfolded gives two shapes, and both are versions of one sentence: Atlas holds. The grouping code found six clusters carrying only two labels: rebadge and hold list price (33, 20, 11 and 8 worlds), and multi-year lock-ins (17 and 11 worlds). Seventy-two worlds hold by the rebadge and twenty-eight by early lock-ins, a lock-in being a multi-year contract that stops the customer repricing until it ends.

In this game, Atlas's harvest is the most common path. The outcomes that should worry a vendor sit in the tail, where lock-ins fail early and the reference price goes public years early.

A forced fork is a change we made on purpose, in every world, to test one decision: should Atlas cut its own licence revenue early and launch usage pricing before Forge forces it?

So we made Atlas do exactly that, in all one hundred worlds, at the start of 2027, a year before the scores would have forced the question. We then compared each world against its own baseline, with the same dice in both.

The bold move loses more often than it wins. The baseline, meaning harvest first and pivot later under pressure, beat the early pivot in 57 of the 100 paired worlds. The early pivot won 39, and 4 tied. The median cost of pivoting early was one point of final position. The downside was fatter than the upside: the middle half of worlds ranged from seven points worse to two points better.

That is a claim about this scenario only. The game's odds for a usage pivot are built from the same pilot-failure rates the evidence lists 10, 12. If you believe your organisation beats those rates, the fork bends. Whether Atlas pivots early or late barely moves the game, next to the variable nobody frames as a decision: how long the real price stays private. That is the argument that decides the game.

You have read one path through the eight beats. It was the most common path, and most common is not settled. Three endings cover the range the hundred worlds produced, and no single world matched any exactly. Where I describe likelihood below, the number is the ensemble's own count of worlds, and a count of games says nothing about the odds in the real world.

The rebadge does its job for the whole window. Every renewal is defended one account at a time, with credits, bundles and side letters, and never a price anyone outside the deal can see. The price customers actually pay falls hard and the list price barely moves. No buyer can anchor on anyone else's outcome, so the repricing lands worst on whoever renewed last with the least information. This is the ensemble's most common shape, 72 of the hundred worlds.

One part of this ending already exists in a weaker form. This report assumes that benchmarking co-operatives, member groups that pool what each member really paid and show the pool to members, already run for business software. If so, a price record, meaning a pool of real paid prices a buyer can see, exists today. Ending A assumes the record stays weak: members only, anonymised, months behind, never quoted in public. I checked no named co-operative, and the seams list both assumptions.

What it looks like: vendor revenue holds while the margin on each unit of work falls. Renewal cycles lengthen, discount approvals move up the chain, and the pricing page stops meaning anything. Buyers with good internal data do well at the expense of buyers without it.

Load-bearing assumption: the private discount is always worth more to the buyer holding it than the real price is to the market once it is public. That held because every player had a private reason to stay quiet. It fails the moment publishing starts to help the career of the person who publishes. A benchmarking co-operative that quoted its pool in public, a procurement co-op that bought as one, or a regulator asking for effective rates would each do that.

Atlas gets multi-year signatures in early, before the agent evidence hardens into a bargaining position, so the repricing lands after the term rather than during it. Twenty-eight of the hundred worlds resolve this way. It is the same harvest as Ending A. The incumbent has bought certainty instead of flexibility, and the buyer has sold a year of bargaining power for a discount it will regret.

What it looks like: a wave of three-year and five-year agreements signed in 2026 and 2027 close to list price, then a much uglier renewal round in 2029 and 2030, when the evidence arrives all at once instead of quarterly.

Load-bearing assumption: buyers sign multi-year deals at exactly the moment their own trial data says they should not. Meridian's audits kept failing to become invoice lines, so the game supports it. It is also the ending most exposed to one competent procurement team, and in the ensemble the lock-in push backfired outright in 14 of the hundred worlds.

A flagship customer publishes its switching arithmetic years ahead of the schedule above, or a buyer calls the rebadge out in public. Every renewal conversation after that opens from the real price that customer paid, instead of from the list price. Both events are minority outcomes: the published switch fired in 16 of the hundred worlds, the call-out in 26. This is the tail that should keep a vendor awake. Some customers move to the published price and most do not, and the price record decides which: a buyer that can point at a published price moves, and one that cannot stays put.

What it looks like: one document, then four more inside two quarters, then a market where the discount is the starting point rather than the concession. The list price moves last, and as a formality.

Load-bearing assumption: one flagship customer's reputational gain from publishing exceeds its relationship cost, while Atlas still has renewals to defend. That needs a specific buyer at a specific moment: mid-divestment, mid-crisis, or already out the door. Without one, Ending C never fires and Ending A runs to 2030.

Everything above is written as memory and none of it has happened. It is September 2026, and the renewal that decides which ending you are walking toward is on somebody's desk right now. It will be settled by email, in private, with no announcement.

Three things you can check today: Salesforce's list price is public and it went up 6 per cent while the agents shipped 3, one industry index puts unused licences at 46 per cent 9, and the agent failure rates are published and brutal 10, 12. Those three are the load-bearing walls of this account.

One indicator is narrower than the list price. Watch the share of deals in your market that open from a real price some customer has published. Today it is close to zero, which is why the rebadge works. The day it stops being close to zero, the timeline above compresses and every renewal conversation you have planned needs rewriting.

A seam is a place where this report could come apart. It was machine-played end to end and you should price that in.

What the machine did. Four AI players took the four roles in a five-turn live run, each with a private brief, win conditions and red lines. A rule-based version then ran the full eight turns one hundred times, which produced the ensemble and the forced fork. No human touched a move in either. Our method pages carry the work on whether such players behave like their written characters. The short version: a written character measurably changed how each AI played, and we make no claim that the games predict what will happen.

The finding rests on one referee rule. The mechanic doing the most work is repetition fatigue, and it surprised me. I expected the challenger to land one of five attempts, and it went nought for five. By the third repeat the referee had raised the difficulty on all four players and kept raising it. Turn that rule off and the challenger probably lands one of five attempts, the audit becomes an invoice, and this report says something else. I kept it because a pitch the market has already declined does age badly. But we built that rule into the game, and nobody has verified it as a fact about markets.

Where the machine got it plainly wrong, and I left the wreckage in. The challenger misidentified who it was attacking. From its first move, Forge described its comparison sheet as exposing "Meridian's incumbent pricing" (Forge, turn 0). Meridian is the buyer. Atlas sells the seats. The referee did not catch the error and built on it. At turn 0 it ruled that the real risk was whether the deal survived "Meridian's counter-offer window", a sentence about a company with no counter-offer to make. The slip recurs through turn 3. It did not change the outcome, because Forge failed every roll either way, but it went uncorrected for most of a five-turn run. This is the strongest single argument for a human referee, which report 001 had and this one did not.

No player abandoned a losing move. Twenty moves were played, containing four distinct plays, one to a player. Atlas rebadged five times, Forge published five times, Pillar hedged five times, and Meridian audited four times before converting the audit into a true-up. The whole roll record, turns 0 to 4, each roll against the referee's bar. Atlas rolled 23, 22, 22, 13 and 13 against 11, 10, 12, 13 and 14: four wins, then a miss. Forge rolled 9, 11, 8, 11 and 10 against 16, 16, 18, 19 and 19: five misses. Meridian rolled 16, 13, 20, 8 and 17 against 8, 6, 11, 13 and 13: three wins, a bad miss, then the true-up landed. Pillar rolled 7, 5, 28, 13 and 11 against 8, 8, 10, 12 and 14: two misses, two wins, a miss. Real executives would have switched tactics sooner. Read the worlds first as evidence that an AI given a strategy will execute it past the point a person would have stopped.

"No world produced a collapse" is partly a property of the board. Atlas starts the rule-based game at 100 and a player is eliminated at zero. The heaviest single loss any Atlas move carries is six points, and the rival moves that pull position away from the leader take two to four more. Reaching zero inside eight half-years therefore needs close to worst-case dice on almost every turn, and the observed floor across the hundred worlds was 72. Read "Atlas never collapsed" as "eight half-years is not long enough to kill this incumbent on this board".

Where I overrode the machine. Beats six through eight are mine, written from the ensemble. The rule-based game does carry the flagship-publishes event, which fired in 16 of the hundred worlds, but the profile of who defects in 2029 H2 is my own synthesis. The dashboard series are curves I drew to match the worlds, and no run record holds them. The recap sheet at the foot reads three numbers off those curves, a quarter, forty per cent and four more customers in six months, so those are authored too.

The players were not calibrated, and the sample is too small to say by how much. Calibration means whether a player's stated odds matched how often it succeeded. Where a player said 70 per cent or better, it succeeded twice out of two. Where it said 40 to 69 per cent, it succeeded 8 times out of 16. Where it said under 40 per cent, it failed both times. That is 20 claims, nowhere near enough to call anything. But the least confident group never landed once, and every one of those claims was Forge's. A model that talks itself down before it rolls is not the same thing as a market that says no.

The confessed list: what this report assumes and did not check. Four assumptions carry weight and have no source. Benchmarking co-operatives pool what members really paid, and the pool stays members only, anonymised and lagged. Zylo's index covers companies large enough to pay for licence tracking, so its US$19.8 million a year is a large-company average. Hybrid seat-plus-usage pricing is the reset the real market is choosing, and it belongs to Ending A: neither game holds a hybrid move, so no dice ever tested that reading. The figure under Ending A is the dashboard's authored curve, since no run record holds a price series.

Where the evidence cuts against the scenario. Three honest collisions. First, the game lets Atlas hold its list price flat, while Salesforce raised its own by 6 per cent 3. The real vendor harvests harder than the game's, so the game understates how far list and paid can pull apart. Second, the game's unused-licence event is 40 per cent, milder than the 46 per cent in Zylo's index 9, which puts more pressure on the secret in reality than in the game. Third, if the pilot-failure rates 10, 12 fall faster than the referee's odds assumed, every beat here arrives early. The scenario's dates are the least defensible thing in it. Its mechanisms are the point.

Pick which of the four players is closest to your own job, then run the game against that job.

If you sell licences: the hundred worlds say your harvest lasts longer than the funeral notices claim, and that is the trap. Every private concession that buys a renewal adds a document, a number or a precedent to the pile that goes public later. Ask your next pricing meeting this. On the day the first flagship customer publishes its switching arithmetic, which of our renewal conversations survives starting from that number?

If you are the challenger: the arithmetic only gets you into the room. Your real product is the first public signature. Pricing, guarantees, legal posture and choice of first customer all follow from what it takes for one flagship customer to survive publishing. Budget for the risk to the job of the person who signs.

If you buy software: your audit is worth more published than banked, to the market. It is worth more banked than published, to you, this quarter. That conflict between you and every other buyer is the silence the game kept reproducing, so decide your position in it rather than default into it. Before you push for repricing, reread the trough beat 8, 12. If you move first on a job your agents cannot actually do yet, you risk the public reversal Klarna had to make.

If you sell implementation: the confusion is a real product with a real expiry date. Sell the transition, but build the thing that survives the pricing question being settled, because on that day "neutral between models" stops being a position and becomes an absence.

The machine played this table and re-ran it a hundred times in an afternoon. What it cannot do is put your own leadership team in the four roles and make them defend their assumptions out loud. That is the Workshop, a paid half-day session: your decision, your people, and someone whose job is to argue with the comfortable answer until it holds or falls over in a room instead of in the market.

v0.6 - 9 September 2026. Argument fixes from a cold read. Benchmarking co-operatives named as the weak form of a price record. Hybrid pricing named as the reset path, and the record shows no player tried it. Salesforce's rise reframed as the harvest move. Ending C states the middle case. Zylo's figure given its population, as an assumption. One authored figure added under Ending A, list price against price paid. A confessed list of four assumptions in the seams. "X, not Y" sentences cut to two. Fifteen figures plus the dashboard. No number, marker or heading changed.

v0.5 - 8 September 2026. From a cold read. The four roles are "players" throughout, "turn" replaces "tick", the reference price is defined at the beat that carries it, and Forge's document is a price comparison. Klarna's reversal is dated fifteen months on and Salesforce's per-conversation price is still listed, as the sources say. The public-price range reads 26 to 42 worlds in a hundred. From the run record: Atlas's rolls, the twenty-move roll record, what the starting scores are, and Forge eliminated in no world. Zylo's 46 per cent is an index figure and the 230-company survey a survey result. The six clusters are named by count, and the recap sheet's three numbers are marked authored. Pillar's fifth move is described one way. Fourteen figures plus the dashboard. No number, marker or heading changed.

v0.4 - 6 September 2026. Plain-language rewrite of all three modes. Every number, date, source row and heading unchanged, one marker occurrence added in the contract as the worked example. Each beat now opens with a sentence saying what its title means, and every pricing and game term is defined where it first appears. Three naming clashes resolved: "seat" means a licence only and the four roles are players, "list price" is the vendor's sticker price and "the real price it paid" is what a customer makes public, "AI agents" is the software under study and "the AI players" are the four roles. The So what section now names the Decision War Room. Byline sign-off pending.

v0.3 - 5 September 2026. Editorial pass, prose only. No number, marker, date, heading, ending title or source row moved. Reveal-shaped sentences, self-announcing openers, stress italics and semicolons cut back, and nine closing lines flattened onto the record. Added: the reading-convention paragraph in the contract, and the line that the players played only the first five turns. One correction: the seams had claimed no player ever changed its mind while recording Meridian's switch to a true-up. Two prose-bar failures stand as the author's call: the aphorism budget and the uniform beat shape.

v0.2 - 2 September 2026. Figures re-set from the run record. Every number quoted from the hundred worlds was reproduced by re-running the game's own code. Nine new figures, thirteen at that date. Three endings added on the ensemble's counts. The "You are reading this in 2026" section added. The seams gained five entries from the transcript, including the challenger player mistaking the buyer for the vendor and the referee carrying the error. One correction: the integrator's first two hedges failed, where v0.1 had the first succeeding. Evidence pack unchanged at 16 claims.

v0.1 - 26 August 2026. First draft. Machine-played end to end: four AI players over a five-turn live run under an AI referee, then the rule-based version run one hundred times for the ensemble and the forced fork. No human refereed either. Sixteen-claim evidence pack, four collision figures, one fork. Soft spots flagged in the seams: the referee's repetition-fatigue rule, the authored beats six to eight, and the dates.

Every marker in the text resolves to a verified claim below. A claim marked * holds with a recorded caveat. Full provenance, verbatim quotes and retrieval dates live in the evidence pack (the-seat-price-v1).