The End of the Seat: Why SaaS Has to Move to Tokenised Pricing

At the end of April 2026, half a sentence in Microsoft’s earnings call redescribed the pricing logic of an entire industry: Satya Nadella called every business in the group that had so far been billed per user a “per-user and usage business” [1]. For fifteen years, SaaS had been regarded as the best business model the software industry ever produced: predictable subscriptions, high net revenue retention, valuation multiples on recurring revenue. The seat, the licence per human user, was the currency of the whole sector. Nadella’s sentence says that this currency is being re-measured.

A few weeks later it became practice. GitHub Copilot moved all of its plans to token-based billing on 1 June [2]. On 15 June, Salesforce paid around $3.6bn for, of all companies, the vendor that had made outcome pricing popular [4]. And IDC expects that by 2028 around 70% of software vendors will abandon pure per-seat pricing [10]. What this means for business models, margins and valuations is the question of this dossier. In several places the answer differs from the doom narrative, and I have corrected the first edition precisely in those places.

Part 1: The seat is not disappearing, it is losing its monopoly

Per-seat pricing rests on a silent assumption: that the value of software is proportional to the number of people who operate it. The assumption was never entirely right, but for decades it was good enough, because work was, after all, done by people.

AI agents attack this proportionality from two sides.

From the demand side: if an employee with AI assistance produces a multiple of their former output, most of the value created stays with the customer. If the customer even cuts staff thanks to this productivity, the seat count falls and with it the vendor’s revenue, although its software does more than ever before.

From the cost side: classic SaaS had marginal costs close to zero. AI features have real inference costs per request. GitHub’s head of product summed up the problem at the Copilot transition: a short chat question and a multi-hour autonomous coding session had so far cost the user the same [2].

“The seat punishes the vendor for the productivity of its own product. That is the economic design flaw that AI agents turn from the exception into the rule.”

What the data actually shows

At this point I have to correct the common narrative, and with it the first edition of this dossier. Seat erosion is not yet measurable. What is happening is unbundling.

Microsoft most recently reported the strongest Copilot seat growth since the product launched: more than twenty million paid licences, up 250% year on year [1]. The assumed trigger is also still missing. In corporate surveys, AI has so far caused no meaningful headcount reductions but has shifted the composition of the workforce [5]. Cuts are certainly expected, a third of companies anticipate smaller teams [6]; they have not been measured.

Nadella himself therefore does not describe the seat as a legacy model but redefines it: as an entitlement to a consumption allowance that is conveniently attached to a user or an agent, with overage beyond that threshold [1]. His sentence contains an and, not an instead of. GitHub has not abolished subscription prices; it has laid a usage layer on top of them [2].

For modelling, the difference is decisive. Whoever models the thesis as “seats die, revenue falls away” models an effect that is still empirically outstanding. Whoever models it as “revenue per customer gains a second, more volatile component” captures the reality that can be documented. I model the second reading.

The industry’s most-quoted number is more than a year old

Hardly any discussion of pricing gets by without it: seat-based pricing supposedly fell within twelve months from 21% to 15%, hybrid rose from 27% to 41%. The figure comes from Kyle Poyar’s monetisation survey, collected, however, in spring 2025 [7]. It describes the movement from mid-2024 to mid-2025, yet has been quoted ever since as if it were current.

The follow-up survey a year later is the finding that counts: at the headline level, hardly anything has moved. Hybrid stands at 37% and remains the most popular model, and the figures resemble those of the previous year [8]. Within the sample the shift continues, but the market structure is more stable than the narrative.

Two details sharpen the picture. Of all companies, the largest vendors are sticking with the seat, while early-stage firms lean towards flat rates. And a considerable part of the pressure to switch comes from the capital side: in the respondents’ assessment, investors today clearly prefer hybrid, outcome and usage models, while flat fee and seat find hardly any advocates [8]. Tokenisation is thus not only a response to customer behaviour but also to capital-market expectations.

Part 2: Three target models, and none has won

“Tokenisation” is the umbrella term. Behind it stand three models of fundamentally different economic quality.

The most telling figure in this chapter is not a market share but an average: companies now combine 1.7 pricing models at the same time, up from 1.5 half a year earlier [6]. The subscription is barely losing ground; usage and outcome are stepping in beside it.

Intentions to change also run in both directions. Around a third of vendors want to change their pricing model within a year, driven partly by customer demand for usage and outcome models and partly by the opposite wish for more predictability [5].

The menu as strategy or as admission

The market is not moving in one direction, it is spreading out. Almost every third vendor now lets customers choose between several pricing models [8]. Salesforce is the extreme case with four parallel variants: per conversation, per executed action via Flex Credits, as an add-on per user and as an unlimited enterprise licence at a fixed price [8][13].

Microsoft is taking the opposite route and announced E7, a predominantly seat-based suite. Asked by an analyst why it was doubling down on the seat, Nadella answered simply with the need for predictability in budgeting and procurement [1].

Bessemer argues against the menu, and the objection carries weight: nine different pricing approaches across the contract base become an operational nightmare at scale; better one model that works with ten and with a thousand customers [9]. For me as an investor the question therefore reads: is a pricing menu a bet on several metrics, or the admission that none of them fits?

Credits are a lifeline, not an end state

A fourth layer has slipped in between the three models: credits. GitHub, Salesforce, Workday and Microsoft now bill in their own credit currencies, and the wave is only beginning. Among companies above $50m ARR, roughly every second one plans to introduce credits within a year [8].

Credits hide the cost mechanics, allow exchange rates and can be readjusted at any time. For the vendor that is convenient. For me it is a warning sign: revenue whose billing unit the vendor defines itself is harder to compare and easier to attack than the revenue figure suggests. Poyar therefore calls credits a lifeline, not the end state; at the latest when customers have to manage dozens of credit schemes side by side, the advantage flips [8].

Not every usage metric is equally volatile

A detail that becomes central in the valuation section: a fee per token fluctuates with every usage decision. A fee per physical asset, that is per monitored door, per installed radio module, per connected vehicle, is formally usage-based as well, but behaves like an annuity. The asset does not decide anew every day about its consumption.

This distinction between behaviour-based and installed-base usage metrics later separates the winners from the losers.

Part 3: The inference curve is not a one-way street

The common narrative runs: AI features have variable costs, so revenues have to become variable. That is true, but it is half the story, and the second half has the larger consequences for valuation models.

The price for constant quality is collapsing. The frontier price is not.

a16z measured the effect and christened it “LLMflation”: for a constant level of capability, the inference price falls by roughly a factor of ten per year. What cost around $60 per million tokens with GPT-3 in 2021 was available three years later for six cents [11].

The decisive addition is in the same text and gets lost almost everywhere: the price of the best models at any given time does not fall. OpenAI’s top model at the time cost as much per output token as GPT-3 did at launch [11]. Translated, this means: whoever delivers constant quality gets dramatically cheaper every year. Whoever wants to stay at the frontier pays a constant price. Cost relief is a function of the willingness to stay behind the front.

Consumption grows faster than the price falls

Salesforce reported for a single quarter an increase in processed tokens of more than half compared with the previous quarter [3]. The customer’s bill therefore rises while the unit price falls.

The finding that follows contradicts SaaS intuition head-on: the share of inference costs in total spending increases the more mature a product becomes, while the share of personnel costs falls [5][12]. In classic software the cost block dilutes with scale. Here it thickens. Whoever assumes automatic margin expansion with volume is assuming the wrong curve.

Fifty per cent margin is not a transition but the target

Three methodologically independent surveys arrive at the same corridor. ICONIQ measures an expected AI gross margin of a good 50% for the current year [6]. Bessemer names a range of 50 to 60% against 80 to 90% for classic SaaS [9]. And Poyar arrives at the same result from the target side: the median targeted margin is around 50%, and only a small minority is even aiming for SaaS-typical values [8].

I find the last observation remarkable. Vendors are not treating the lower margin as a transitional phenomenon; they are planning for it. Internal costs were by far the most important factor in pricing in Poyar’s survey, more important than the competitive comparison and more important than the productivity gain realised at the customer [8]. The industry is pricing its costs, not its value. Bessemer calls exactly that a mistake; it is evidently still common practice.

Where the margin is actually earned

Three levers move inference costs, and they are technical in nature, not a matter of negotiation.

Routing has long been an architectural standard: companies use a good three model providers in parallel on average, and cost is the second most important selection criterion after reliability [5]. Microsoft describes the same architecture from the vendor’s perspective: decouple the orchestration layer from the model, with several models and automatic routing behind it [1].

How much there is to gain is shown by the cohort analysis: between vendors with their own model and product development and those that merely pass through a third-party model lie around eight percentage points of gross margin [5][12]. At the same time, model ownership barely counts as a differentiator; respondents overwhelmingly name the application layer as their point of difference, hardly anyone their own model development [5].

From this follows a formula that is rarely separated cleanly in the debate: owning models is a margin lever, not a differentiation lever.

And one observation that weighs more for portfolio allocation than any average: the layer with the strongest margins is not the application but the infrastructure and platform layer [6]. In this gold rush, too, the shovel sellers earn better than the prospectors.

Microsoft shows both sides

On the cost side, the effect shows up in the segment figures: group gross margin was below the prior year, driven by AI infrastructure and growing product usage; for the Microsoft Cloud a further decline was guided, with explicit mention of increased GitHub Copilot usage [1]. A company of this size writes the inference costs of its own Copilot into its margin guidance. The thesis cannot be documented more clearly.

On the other side, the same company is investing massively in reducing precisely these costs: its own models explicitly for COGS reduction, an in-house accelerator with a noticeably better token-per-dollar ratio, significantly improved inference throughput [1]. And CFO Amy Hood contradicts the doom narrative directly: measured at the comparable point in the cycle, margins in the AI business were better than in the cloud transition back then, and have stayed that way [1]. That is an interested statement, but a testable one.

Part 4: Five construction sites of the transition

The pricing shift sounds like a change to the price list. In reality it is a rebuild of the finance and sales operating system.

First: the leading indicator gets lost. Under flat-rate subscriptions, the revenue of the next twelve months is essentially signed; under usage models it is estimated. The less obvious part concerns the metrics themselves. Amy Hood explained that the licensing logic stays but a meter is added, and that the usage part does not flow through bookings in the same way but is simply billed [1]. For me this means: bookings, RPO and backlog lose explanatory power the further the usage share grows.

Second: revenue recognition gets complicated. Subscription revenue is recognised rateably, usage revenue arises with consumption. Tiered prices, free allowances, rollover credits and ramp phases make determining the transaction price demanding. Metering and billing infrastructure is thus not an option but a prerequisite.

Third: sales compensation no longer fits. The classic model pays for the close. Under usage models there are two moments: landing the customer and the subsequent growth in consumption. Whoever wants to reflect both rebuilds the sales organisation and upgrades customer success. Friction in the transition phase is the rule.

Fourth: a few users carry the costs. By Poyar’s observation, the top ten per cent of power users drive over 70% of token consumption [8]. A vendor with a healthy average margin can be clearly negative on this tenth at identical revenue. How real this is, Bessemer’s example shows: Microsoft initially lost around $20 per user per month on GitHub Copilot [9]. As long as power users are a fringe phenomenon, the blended price hides this. Once agentic usage becomes normal, the loss-making case moves to the middle of the distribution.

Fifth: the migration has a J-curve. Light users pay less immediately, power users push back, and the new systems cost before they deliver. GitHub provided the textbook example in summer 2026: after the first full billing month, heavy agent users reported bills at a multiple of their old flat fee, while pure autocomplete users paid practically unchanged amounts. At the beginning of July, GitHub introduced hard spending caps [2].

Analytically more valuable is the second piece of evidence, because it comes from a segment report: Microsoft explicitly attributed the muted bookings growth at Dynamics 365 to weaker renewals; customers were rebalancing their spending between per-seat and “seats plus consumption” [1]. The J-curve is thus not a start-up phenomenon.

The silent twin: launched but not yet billed

Workday shows the other variant of the same problem. The company has introduced Flex Credits but has published no public price list. At the same time, broad grace windows are running: overages on application APIs remain unbilled until early 2027, and individual agents are free for now [14].

Economically this means: the model is in place, the revenue effect is deferred, the cost side is already running. The relevant question is therefore not whether a vendor has switched, but whether it is already billing.

Part 5: What this means for capital allocation

Inside the company: capital moves into the cost architecture

Whoever switches has to direct capital to new places: metering and billing infrastructure, customer success capacity, FinOps capabilities for attributing inference costs per customer. Newly added is the inference cost architecture itself: routing, caching, batch processing and, where applicable, in-house models. That is where the eight percentage points of gross margin mentioned above lie.

At the same time, the relative value of classic sales capacity falls. If expansion scales automatically with usage, every dollar invested in quota-carrying sales is worth less than before. The signature of successful switchers: lower acquisition costs, more infrastructure, more retention.

In valuation: not all recurring revenue is worth the same

The line of tension runs right across the table. Usage models systematically achieve higher net revenue retention because expansion grows along without sales effort. At the same time, uncommitted usage revenue trades at a discount to contractually fixed revenue, because it can shrink in a recession without any customer cancelling. The customer does not have to churn, it only has to consume less.

For my DCF model, this implies no blanket increase in the discount rate but a clean price-volume decomposition of the revenue forecast. Token-based revenue is exposed to the deflation of its own unit: a seat tends to get more expensive every year, a token cheaper. Whoever ties revenue to a unit whose market price falls by an order of magnitude per year needs volume growth of the same order of magnitude just to hold revenue flat.

Whoever defines the unit defines the revenue

In outcome models, the vendor determines what gets billed. And it changes that determination. In May 2026, Zendesk switched to a scheme that charges only for verified resolutions; assisted escalations have been free since. HubSpot halved its price per case in April and tied billing to the condition that no human intervenes within 72 hours [15].

Both vendors have materially changed their billable unit within twelve months. Whoever capitalises outcome revenue at a subscription multiple is capitalising a definition, not a contract.

The same flexibility, incidentally, also works in the vendors’ favour. When Fin launched on a per-resolution basis, the resolution rate stood at around 35%; at that level the vendor lost money on every ticket in weak months. Cheaper inference and better context engineering have lifted the rate to over 70% [4]. The same price tag today means a completely different margin. Whoever values outcome models is primarily valuing the resolution rate.

The finding that weighs most heavily on the sector outlook

If AI revenue were incremental revenue, tokenisation could be read as a pure growth story. The data says otherwise: seven out of ten vendors report that their customers’ AI spending comes from the existing technology or software budget; among established SaaS vendors retrofitting AI, the share is even higher [8]. Only AI-native vendors tap other pots to any meaningful degree, above all services budgets.

That is the most uncomfortable number in the whole dossier. It says that the impressive AI revenue of individual vendors comes to a considerable extent from the budgets of other software vendors: a redistribution, not an expansion. Whoever holds a software position should therefore ask not only whether the company monetises AI, but whether it is a net winner or a net payer in this redistribution.

The re-rating is already behind us

The software sector completed its revaluation in the winter of 2025/26, and it did so along precisely the lines this dossier describes. Horizontal, user-oriented software compressed the most; vertical, infrastructure-adjacent and AI-native software held its multiples considerably better [16].

I deliberately do not cite specific median multiples here: the published values vary considerably depending on how the peer group is defined. The direction is robust, the level is not.

Part 6: Winners, losers and the quiet victors

Structurally vulnerable is horizontal per-seat software with high natural churn whose value proposition depends on human operation: collaboration tools, generic productivity software, tools for entry-level knowledge work. Here the double lever applies, because AI agents prospectively reduce seats and force price concessions.

Structurally protected is deep vertical software with non-seat metrics, compliance anchoring and proprietary data sets. The Constellation camp has argued for months that the hard part of vertical software is not writing code but selling to a risk-averse hospital or a municipal transit operator — security reviews, legacy integration, training and trust built over years [17]. I find it remarkable that ICONIQ’s finding on differentiation supports exactly this argument, from an entirely different direction.

The market has nevertheless punished this group indiscriminately. Between summer 2025 and January 2026, Constellation went through the deepest drawdown in its corporate history, more than a halving from the high. The operating figures for the first quarter of 2026 did not confirm the fear: revenue grew around 20%, free cash flow available to shareholders jumped to a record, and management deployed more capital for acquisitions in a single quarter than ever before in an opening quarter [17]. That is the signature of a capital allocator treating a valuation correction as a buying opportunity.

Already at the destination are business models whose billing unit was never the seat but the physical asset. The fee per installed radio module or per monitored door is formally usage pricing, but with annuity stability. No AI agent replaces a physical fire-alarm radio module.

NAPCO is the cleanest case in my coverage universe. Recurring service revenues most recently grew a good 15% and now make up around half of group revenue, at a gross margin of over 90% [18]. Such a margin on device-bound usage revenue is exactly the profile the industry is only now trying to reach through tokenisation: a variable revenue form without variable inference costs. These models anticipated the end state without ever having to pass through the migration J-curve.

Part 7: The checklist

For every software position and for every candidate, I answer five questions.

1. What is the billing unit, and does AI replace it? Seats for routine knowledge work are the highest risk class. Transactions, assets, devices, beds and vehicles are robust metrics. Bessemer’s mapping helps with classification: copilots that assist people produce soft ROI and therefore weak pricing power; agents that take over a workflow entirely produce hard ROI and strong pricing power [9].

2. How high is the committed share of revenue? Minimum commitments, multi-year contracts and installed-base fees deserve the recurring multiple; spot-like usage does not. And beware of the metrics: a declining bookings trend can mean migration rather than weak demand.

3. Does the company bear inference costs, and is the cost architecture mature? The concrete question for management: how is traffic routed between models, what is the cache hit rate, what is batch-processed, and how are costs attributed per customer?

4. Where does the company stand on the migration J-curve? Before the switch, in the middle of it, or through it? And is a grace window still running? A launched usage model with suspended overage billing is an announcement, not a revenue source.

5. What remains of the moat when software creation deflates? Data, workflow anchoring, compliance, installed base and trust relationships survive. Pure feature superiority does not.

For outcome positions a sixth question is added: who defines the billable unit? Poyar’s CAMP framework names the four conditions on which such models stand or fall: consistency of results across customers, attribution vis-à-vis the customer, measurability in real time and predictability [7]. If one of them is missing, outcome revenue is a negotiating position, not an entitlement.

Conclusion

The tokenisation of software pricing is not a hype cycle but the answer to two structural breaks: AI agents decouple the value of software from headcount, and inference costs make delivery variable.

I consider three refinements of the common narrative decisive. The seat is not disappearing, it is losing its monopoly; seat erosion through headcount cuts is so far a forecast, not a measurement. The inference curve is not a one-way street: the price for constant quality falls rapidly, the frontier price does not, and consumption grows so fast that the cost share actually increases with product maturity. And the burden of proof is shifting: bookings and RPO lose power as leading indicators, while resolution rates, cache hit rates and the committed share of revenue gain in importance.

Update log

• 05.09.2026 — Editorial revision to house style, list of sources expanded; figures and assessment unchanged.
• 12.08.2026 — Full revision: new chapter on inference economics (Part 3), corrected pricing mechanics of the vendors named, primary-source-based rewrite of the cost arguments, numbered references.
• 11.06.2026 — First publication.

Sources

1. FY2026 Q3 Earnings Call, transcript, Microsoft, 29.04.2026. Nadella quote “per-user and usage business”, Copilot seats (+250%), E7, gross-margin commentary, Dynamics 365 bookings, Amy Hood on the meter and bookings. → microsoft.com/investor
2. “GitHub Copilot is moving to usage-based billing”, GitHub Blog, 27.04.2026, plus the community FAQ on the transition dated 01.06.2026. Switch on 1 June, head-of-product quote, spending caps in July. → github.blog · Community FAQ
3. Q1 FY2027 results (8-K / IR release), Salesforce, 27.05.2026. Agentforce ARR, token growth quarter on quarter. → investor.salesforce.com
4. “Signs Definitive Agreement to Acquire Fin”, Salesforce, 15.06.2026. Purchase price of around $3.6bn, Fin’s resolution rate. → investor.salesforce.com
5. “State of AI: Bi-Annual Snapshot”, ICONIQ, January 2026 (survey December 2025, ~300 executives). Workforce effects, intentions to change pricing, model providers per company, cost share by product maturity, differentiation. → iconiq.com
6. “2026 State of AI: The Builder’s Economy”, ICONIQ, July 2026 (survey Q2 2026, ~305 executives). Expected AI gross margin, 1.7 pricing models per company, team-size expectations, margins by layer. → iconiq.com
7. “The state of B2B monetization in 2025”, Kyle Poyar, Growth Unhinged, 04.06.2025 (survey April–May 2025, 240+ companies). Original source of the share figures 21% → 15% (seat) and 27% → 41% (hybrid) and of the CAMP framework. → growthunhinged.com
8. “The state of B2B monetization in 2026”, Kyle Poyar, Growth Unhinged, 13.05.2026 (survey April–May 2026, 230+ companies). Hybrid share 37%, investor preferences, pricing menus, credits, target margins, power-user consumption, budget origin. → growthunhinged.com
9. “The AI pricing playbook for founders”, Bessemer Venture Partners, February 2026. Margin ranges, critique of the pricing menu, Copilot loss per user, copilot-versus-agent mapping. → bvp.com (PDF) · companion page
10. “Is SaaS Dead? Rethinking the Future of Software in the Age of AI”, IDC, 02.12.2025, based on IDC FutureScape (October 2025). Forecast: around 70% of vendors leave pure per-seat by 2028. → idc.com
11. “Welcome to LLMflation — LLM inference cost is going down fast”, a16z, 12.11.2024. Price decline by a factor of ten per year, GPT-3 price, constant frontier price. → a16z.com
12. “AI Unit Economics: Pricing & Margins”, Digital Applied, 30.06.2026. Own calculation on model routing, prompt caching and power-user economics based on published vendor prices. → digitalapplied.com
13. Agentforce pricing page, Salesforce, as of August 2026. Four parallel pricing variants. → salesforce.com
14. Overview of Workday Flex Credits, Commit Consulting, 01.07.2026. Secondary source; Workday publishes no price list. Grace window until early 2027. → commitconsulting.com
15. Pricing pages of Zendesk, HubSpot and Intercom Fin, as of August 2026. List prices, changeable by the vendor at any time; outcome definitions and the 72-hour rule. → zendesk.com · hubspot.com · intercom.com
16. Aggregated valuation data of public software peer groups, March to August 2026. Published median multiples vary considerably depending on how the peer group is defined; a uniform primary basis (such as SaaS Capital, the KeyBanc SaaS Survey or Bessemer’s “State of the Cloud”) is still outstanding. Market data without URL.
17. Q1 2026 results, Constellation Software Inc. (TSX: CSU), published 12.05.2026, plus coverage of the 2026 annual general meeting. Revenue growth, free cash flow available to shareholders, acquisition volume; the share-price levels cited are market data, not a company disclosure. → constellationsoftware.com · filings via SEDAR+
18. Q3 FY2026 results (8-K, Exhibit 99.1), NAPCO Security Technologies, 04.05.2026. Growth and share of recurring service revenues, gross margin. → sec.gov

The core statements of this dossier rest on sources 1 to 11: audited segment reports, transcripts, official price lists and surveys with disclosed methodology and sample size. Item 16 carries the weakest evidentiary weight and is therefore used in the text without a specific figure.