Skip to content
A Long Game Project lab
Gilded graded 1763 engraving of a riot in the pit of Covent Garden Theatre, marked with tea-stain rings

The Quiet Discount

Business software is sold one licence per person, and AI agents now do those people’s work. When does the price reset, and who pays?

Report 002 · Dr Dan Epstein · version 0.10 · 12 October 2026
A Strategy Soup Scenario What is an exercised scenario?

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.

Our answer, from a simulation played once and then re-run a hundred times by fixed rules: the runs track no prices, so they cannot say when the price resets. What they count is whether the secret became public while the incumbent vendor, the large software company in the simulation, still had renewals to defend. Every large player had a reason to keep it.

In 62 of the hundred re-runs the secret was kept to mid-2030, when the runs end. In the other 38 it came out. In 16 a customer published what it really paid. In 26 a buyer publicly called out the incumbent vendor’s rename: putting an AI label on the product it already sold, instead of cutting the price. Four had both. We count the call-out because it tells other buyers the list price is not the real price, though it shows no price.

Who pays? In the re-runs, the incumbent vendor loses a little and the corporate buyer gains. Each player carries a score for its strength in the market, on a scale we chose. The incumbent vendor starts at 100, zero means out of the market, and its worst single move costs it six points. The incumbent vendor’s median score, the middle of its hundred results, fell from 100 to 95, while the corporate buyer’s rose from 70 to 79. The incumbent vendor ended lower where the secret ended: 92 against 99. The simulation has one buyer among its players, the corporate buyer. The customer who publishes and the buyer who calls out the rename are events outside the four players, so the runs cannot show which buyers gain least.

The claim, 2026 to 2030Fig. 1

In 62 of 100 re-runs, the secret was kept to mid-2030: no customer published what it paid, and no buyer called out the rename.

Out of 100 re-runs: secret still kept1009384757469656262 kept38 outH2H1H2H1H2H1H2H120262027202820292030
Secret out: a customer published what it paid, or a buyer called out the incumbent vendor’s renamed product in public.

How the simulation works

The simulation has four players, each with a role, a goal and a menu of moves.

The incumbent vendorA very large software company whose income is mostly per-user licences. It defends the list price.
The AI-native challengerA smaller rival that charges per job completed rather than per user. It needs a public price comparison.
The corporate buyerA large company, run by a finance chief paying for licences nobody logs into.
The integratorA global consulting firm paid to install and connect enterprise software, whoever wins.

The live run, in which each player chose its own moves and gave its reasons, lasted five turns, late 2026 to the end of 2028. We then built a rule-based version with fixed rules, so runs with the same chances always give the same result. We ran it a hundred times with different chances over all eight turns to 2030, and one full play-through is a run. The story’s last three half-years, 2029 to mid-2030, are written from those hundred runs. No player in the live run played them. That story follows the less common branch, where a customer publishes its real price, and places it in 2029. The real facts in this report stop in August 2026. Everything dated later is the simulation’s story, and some of its early dates have now passed unchecked.

The turnHalf a year. A player names a move. Chance decides whether it works, with odds set by the simulation’s rules.
RepeatsA repeated move is harder to land. A rule we built into the live run. The hundred re-runs do not use it.
The scoreOne score per player, moved after every turn: up when a move lands, down when one fails. The incumbent vendor starts at 100, the corporate buyer at 70, the integrator at 60, the AI-native challenger at 30, on a scale we chose.
Built-in eventsEvents the simulation puts in whatever the players do, such as an audit finding 40 per cent of licences unused.

Salesforce, Microsoft, Klarna, and the research firms Gartner and MIT are real, and each carries a source in the deep mode. The four players exist only in the simulation. This version is dated 12 October 2026.

OF THE AI-NATIVE CHALLENGER’S FIVE ATTEMPTS TO GET A CUSTOMER TO SIGN ITS PRICE COMPARISON, ATTEMPTS THAT LANDED

0 of 5

OF THE HUNDRED RE-RUNS, RUNS WHERE THE SECRET WAS KEPT TO MID-2030

62 of 100

OF THE HUNDRED RE-RUNS, RUNS WHERE THE INCUMBENT VENDOR WAS DRIVEN OUT OF THE MARKET

0 of 100

THE INCUMBENT VENDOR’S SCORE AT THE END OF THE MEDIAN RUN, FROM A START OF 100

95

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

The incumbent vendor 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. The incumbent vendor renewed both at full list price before either could argue for a discount.

What the incumbent vendor 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. The incumbent vendor knew the risk, and noted among its opening assumptions that the side agreements’ confidentiality “is not perfectly enforceable”. It went ahead, “knowingly trading a near-term revenue/logo win for continued non-resolution of the public pricing question”, as its assumptions put it. The move worked four turns running before it missed.

The real market goes further than the simulation. 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 keeps 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 AI-native challenger’s price comparison was right and nobody would sign it

The AI-native challenger’s move was to publish a sheet comparing its outcome pricing with what the corporate buyer 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. The AI-native challenger published five times, and no customer ever signed.

The AI-native challenger had put its own odds at 38 to 52 per cent, and all five attempts missed. A rule in the live run made each repeat of a move harder to land than the last, on the view that 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. The incumbent vendor sends a private counter-offer within days. And procurement, the department that buys software, prefers to use its bargaining power in private. The AI-native challenger saw it by the last turn: the pitch convinced people, and getting a company to attach its name did not.

The corporate buyer kept its audit private, and failed trials delayed the repricing

The corporate buyer 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. For that average organisation, Zylo puts the waste at about US$19.8 million a year. We assume the firms it tracks are large ones, big enough to pay for licence tracking.

The corporate buyer kept the finding private. It switched off some idle licences to bank quick savings. 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 corporate buyer that motive on purpose, because buying teams are paid on this year’s savings.

The third audit was harder to land because it repeated the first two, and the corporate buyer still landed it. The fourth audit failed badly. The fifth turn is the one to copy. The corporate buyer stopped auditing and forced a contract reset on its largest per-seat agreement, and it landed. The evidence was the same, and this time the number had an invoice attached.

Failed trials delayed the repricing 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 simulation every failed trial reset trust inside the buyer, and the incumbent vendor only had to argue that this renewal was not the moment to switch.

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

Rank the hundred runs by the incumbent vendor’s final score and the fiftieth is the median run. The incumbent vendor 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 the incumbent vendor out of the market.

When the secret came out, in 38 runs, it was usually early: 25 of the 38 by mid-2028.

We also sorted the hundred runs into groups of similar play, to see how the incumbent vendor defended its price. In 72 it leaned mainly on the rename, and in 28 mainly on early lock-ins, multi-year contracts that stop the customer repricing until they end. The secret survived about as often either way, in 44 of the 72 and 18 of the 28. So the incumbent vendor’s choice of defence did not decide whether the secret came out.

In our telling, the customer that ends the secret is a large, well-known buyer already on its way out of the relationship. One published customer reprices nothing by itself, but it gives every other buyer a real price to quote. The incumbent vendor’s renewal talks had opened at the list price and discounted down. Now they opened at the published price and argued upward.

The window closes on a settlement. The median run ends with the incumbent vendor at 95 from 100, the corporate buyer at 79 from 70, the integrator at 63 from 60 and the AI-native challenger at 36 from 30. The AI-native challenger gained the most in proportion and is still the smallest player. No player was driven out, and none took over the market.

We then tested one decision on purpose. In all hundred runs we made the incumbent vendor switch to usage pricing at the start of 2027, and compared each forced run with its baseline: the same run, with the same chances, and the incumbent vendor left to its usual moves. Left alone, the incumbent vendor switched in only 2 of the 100. The early switch lost 57 of the 100 pairs, won 39 and tied 4. The median cost was one point of final score, and half the pairs sat between seven points worse and two better.

The secret moved the incumbent vendor’s score more, by seven points. Part of that gap is built in, because a public call-out of the rename costs the incumbent vendor four points.

What it means

In this simulation both sides of the deal gained from keeping the secret. Our judgement, beyond what they measure, is that buyers cannot push the price down together until 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.

If you sell licencesEvery private concession adds a document someone can leak later. Ask which of your renewals survives starting from the first real price a customer publishes.
If you are the challengerThe arithmetic only gets you into the room. Your real product is the first public signature, so budget for the risk to whoever signs.
If you buy softwarePublishing your audit helps every buyer. Keeping it private helps you more at this quarter’s renewal. Decide on purpose, and check the failure rates before you move first.
If you sell implementationThe confusion about which price model wins is a real product with an expiry date. Build the thing that survives the day the question is settled.

What this does not prove

The repeats rule shaped the live run. It did not sink the AI-native challenger, because even its best try fell short of what its first attempt needed, before any repeat. It may have sunk the incumbent vendor’s fifth move and the integrator’s fifth hedge, since both would have landed at the odds of that player’s first attempt. The hundred re-runs have no such rule. We built it in because a pitch the market has already declined does age badly, but nobody has verified it as a fact about markets.

A player made one plain error, and we left it in the record. From its first move the AI-native challenger described its comparison as exposing “Meridian’s incumbent pricing”. In the transcript, Meridian is the corporate buyer, not the incumbent vendor. The live run built on the slip for four turns. It did not change the outcome, because the AI-native challenger failed every attempt either way. But the slip went uncorrected for four of the five turns.

“The incumbent vendor never collapsed” is partly a property of the board. It 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 chance draws on almost every turn. The result shows only that eight half-years is not long enough to drive this incumbent vendor out on this board.

Three more limits. The live run lasted only five turns, and the last three half-years are written from the hundred runs. No run tracks a price, so every price movement in our telling is ours. The scores, start values and dates are ours, and the dates are the least defensible thing here. The evidence also cuts against the simulation twice, Salesforce’s list-price rise and the real unused-licence rate of 46 per cent against the simulation’s 40, and both put more pressure on the secret.

What to watch

Four facts on the public record, and one indicator.

List pricesSalesforce’s list price is public and went up 6 per cent in August 2025. Watch whether any large per-seat vendor cuts its list price.
Unused licencesThe measured rate is 46 per cent, from Zylo’s 2026 index. Count yours.
Failure ratesMIT found 95 per cent of organisations with no return on generative AI. Gartner forecasts over 40 per cent of agent projects cancelled by the end of 2027.
Pricing churnThree in four software companies changed pricing or packaging inside a year, in a May 2026 survey of 230 companies by Growth Unhinged. The same survey has hybrid pricing, seats and usage sold together, up from 25 to 37 per cent in twelve months.
The one indicatorThe share of deals in your market that open from a real price some customer has published. We know of no public count of it.

Our judgement is that 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.

The record

Each caption says where its numbers come from.

The mechanism, our readingFig. 2

Our telling of how the price breaks once the real price is public.

How the price breaks, in the runs where it breaksThe list pricestays levelprivate discounts pullthe real price downThe gap stays asecretvendor and buyer bothwant it keptOne big customerpublishesalready on its wayout, with nothing leftto protectTalks open atthat real priceand argue upward fromthereThe list pricemoves lastas a formality
This chain is our reasoning. The runs count the events, not the price moves: a customer published its real price in 16 of 100, and the secret came out in 38.
The live run, five turnsFig. 3

The incumbent vendor’s move landed four turns in five. The AI-native challenger’s never landed.

Incumbent vendor: rename, discount in private4 of 5 landedAI-native challenger: publish the comparison0 of 5 landedIntegrator: stay neutral, sell advice2 of 5 landed
One dot per turn, filled where the move landed. The corporate buyer, not shown, audited four times, then switched to a contract reset on the fifth turn, and it landed.
The AI-native challenger’s five attemptsFig. 4

The AI-native challenger put its own odds at 38 to 52 per cent, and none of its five attempts landed.

Attempt 1 (2026 H2),missed38Attempt 2, missed41Attempt 3, missed38Attempt 4, missed41Attempt 5 (2028 H2),missed52per cent, the AI-native challenger’s own odds foreach attempt at the same move
The AI-native challenger’s own odds for each of its five attempts at the same move in the live run. Each repeat was harder to land than the last.
The real-world base ratesFig. 5

The repricing argument assumes agents work. A real survey found most organisations had no return yet, and a forecast expects many projects cancelled.

Agent projects to becancelled byend-2027, % ofprojects (forecast)40No return ongenerative AI, % oforganisations(measured)95each bar is a share of a different whole: compareeach with 100, not with the other
Each bar is a share of a different whole, so compare each with 100, not with the other.
The incumbent vendor across the hundred runsFig. 6

The incumbent vendor’s score drifted down slowly, and the middle 80 runs stayed in a narrow band.

20262028 H12030The incumbent vendor’s score, middle 80% of runsAt the end: middle 95, middle 80% from 87 to 109
The incumbent vendor’s score by half-year. Line: the median run. Shaded band: the middle 80 runs.
The hundred re-runsFig. 7

The secret came out in 38 runs, and a customer published its real price in only 16. In the other 62 the secret was kept to mid-2030.

A customer published its real price16 of 100 publishedA buyer called the rename out26 of 100 called outIncumbent vendor driven out of the market0 of 100 collapsed
One dot per run, all eight turns, filled where the event fired at least once. Four runs had both of the first two events, so 16 and 26 make 38 runs, not 42.
The score at the endFig. 8

No player was driven out or took over. The corporate buyer gained nine points and the incumbent vendor lost five.

dotted tick: where the player startedIncumbent vendor95-5Corporate buyer79+9Integrator63+3AI-native challenger36+6
Each bar is the player’s median score in 2030, the dotted tick where it started. The four start values are a scale we chose.
The forced decisionFig. 9

Switching early lost more pairs of runs with the same chances than it won, at a median cost of one point.

forced decision wins · 39baseline · 57tied 4 of 100median difference -1 points of final score
Baseline: the same run, the incumbent vendor left to its usual moves. Dark grey: switching early won. Cyan: the baseline won.

Want the evidence - the dated record, the sources, and our own list of weak points?

Read deep mode