The End of the Seat: Why SaaS Must Shift to Tokenization

The consensus of the past fifteen years was: SaaS is the best business model the software industry has ever produced. Predictable per-user subscriptions, net revenue retention above 110%, valuation multiples on recurring revenue. The seat — the license per human user — was the currency of the entire industry.

That currency is being devalued right now. On June 1, 2026, GitHub Copilot — one of the most widely used AI products in the world — switched all its plans to token-based billing: the plan price is now merely a monthly allowance consumed through usage. Workday reported over $400 million of AI-related ARR and now measures consumption through a flex-credit model that tracks API and agent usage across the entire platform. Zendesk and Intercom bill AI support per resolved case; HubSpot cut its price per resolved conversation to $0.50 in April 2026. These are no longer experiments — this is the new norm under construction. IDC forecasts that by 2028 roughly 70% of software vendors will abandon pure per-seat pricing; Gartner expects that by 2030 over 40% of enterprise SaaS spend will fall on usage, agent, or outcome models.

This dossier breaks the shift down into its components: why the seat is breaking economically, which target models are winning, why the transition is so hard operationally — and what all of it means for capital allocation and valuation.

Part 1: Why the Seat Is Breaking

Per-seat pricing rests on a silent assumption: the value of software is proportional to the number of humans operating it. That assumption was never entirely correct — but for decades it was good enough, because work was, after all, done by humans. More work meant more employees meant more seats.

AI agents destroy precisely this proportionality, and from both sides at once:

From the demand side: If an employee with AI assistance handles three times the workload of one without, per-seat pricing leaves two thirds of the value created on the customer's table — the vendor does not participate. Worse still: if the customer reduces headcount thanks to AI productivity, the seat count falls — and with it the software vendor's revenue, even though its software is delivering more than ever. The seat punishes the vendor for the productivity of its own product.

From the cost side: Classic SaaS had marginal costs near zero — an additional user cost the vendor almost nothing, and the seat price was nearly pure margin. AI features, by contrast, carry real, variable inference costs per request. A user who queries the AI assistant ten times a day and one who runs autonomous agents for hours generate completely different costs — but pay the same seat price. The flat-rate model turns into an uncovered customer option on unlimited compute consumption. GitHub's Copilot switch is the direct answer: a short chat question and a multi-hour autonomous coding session will no longer cost the same.

And from the conceptual side: When AI agents themselves become "users" — who, then, is the seat? An agent that works around the clock, scales in parallel and can be cloned in seconds explodes any person-based licensing logic. The industry needs a new unit of account, and it has found it in the most granular size available: the token, the API call, the resolved task.

The data shows the shift is already underway, not impending: per Bessemer tracking, the share of pure per-seat models among SaaS companies fell from 21% to 15% within twelve months, while hybrid models jumped from 27% to 41%. Bloomberg estimates see the share of classic subscription models falling from 60% to 30% over the next decade.

Part 2: The Three Target Models — and Their Very Different Quality

"Tokenization" is the umbrella term, but behind it stand three different models with fundamentally different economic quality:

The hybrid dominance is no accident but risk management: the base subscription delivers the predictable revenue floor that CFOs and investors demand, while the consumption component captures the AI upside. Most enterprise renewals in 2025–2026 land exactly here. Pure token models remain the domain of API-first products, pure outcome models that of narrowly definable tasks (support tickets, lead qualification).

One detail deserves special attention because it almost always gets lost in the discussion: not every consumption-adjacent metric is equally volatile. A fee per token fluctuates with every usage decision the customer makes. A fee per physical asset — per monitored door, per installed radio module, per connected vehicle — is formally also "usage-based" but behaves like an annuity, because the asset does not decide anew about its consumption every day. This distinction between behavior-based and installed-base consumption metrics becomes central in the valuation section.

Part 3: Why the Transition Is So Hard — the Five Construction Sites

The pricing shift sounds like a price-list change. In reality it is a rebuild of the entire finance and go-to-market operating system. Five construction sites every transitioning company must pass through:

Site 1: Loss of forecastability. Under flat-rate subscriptions, the next twelve months of revenue are essentially signed. Under consumption models, they are estimated. Even minimum-commitment contracts and committed-spend pools only partially solve the problem, because the timing of drawdown remains unknown. Revenue forecasting transforms from a quarterly routine into daily operational intelligence. One consolation from practice: at portfolio level, volatility smooths out — across fifty customers, consumption fluctuations offset each other considerably. The problem is primarily a single-customer forecasting problem, not an inevitable aggregate-volatility problem.

Site 2: Revenue recognition. Subscription revenue is recognized ratably over the contract term; consumption revenue arises when consumption occurs (ASC 606 / IFRS 15). Tiered pricing, free allowances, rollover credits and ramp phases make transaction-price determination complex. Attempting this with legacy billing systems and manual processes produces billing errors and compliance risks — investment in metering and billing infrastructure is not optional but a prerequisite.

Site 3: Sales compensation. The classic comp model pays for the close (booking). Under consumption models there are two moments of persuasion: landing the customer and the subsequent consumption growth. Compensation plans must reflect both — which rebuilds sales organizations, elevates customer success, and regularly creates friction and attrition during the transition.

Site 4: Margin risk from inference COGS. This is the most dangerous site. Classic SaaS had 75–85% gross margins because delivery costs were fixed. AI products carry variable inference costs that scale with usage — whoever sells AI features under a flat rate watches gross margin erode with every power user. The pricing architecture must therefore be built as a mirror image of the cost architecture: variable COGS demand variable revenues. That is the deeper reason the shift is inevitable — it is not a marketing fashion but margin defense.

Site 5: The migration J-curve. Whoever moves an installed base from per-seat to consumption almost always experiences a transition dip: light users immediately pay less, power users resist higher costs, and the new systems cost money before they deliver. The transition eats revenue and margin in the short term to gain value alignment in the long term — a classic investment problem that punishes quarterly thinking.

Part 4: What This Means for Capital Allocation

This is where the dossier becomes an investment tool. The pricing shift changes three levels of capital allocation: that of the software company, that of the valuation model, and that of the portfolio.

Level 1: Capital allocation inside the company

The transitioning company must redirect capital to new places: metering and billing infrastructure (build or buy — the existence of specialists like Metronome or m3ter shows an entire infrastructure layer emerging here), customer success capacity (which under consumption models drives the expansion revenue that sales used to sell), and FinOps capabilities for inference-cost attribution per customer and feature. At the same time, the relative value of classic sales capacity falls: when expansion scales automatically with usage, every dollar invested in quota-carrying sales is worth less than before. The capital allocation signature of successful transitioners: less CAC, more infrastructure and retention.

Level 2: Valuation — re-measuring revenue quality

For the investor, the most important consequence is uncomfortable: not every "recurring" revenue is worth the same anymore. The market has begun to re-tier revenue quality:

The line of tension: usage models systematically achieve higher net revenue retention because expansion grows with the customer's value creation without sales intervention — and NRR has become the dominant valuation driver in 2026 (a 10-point NRR improvement is associated with a 20–30% valuation premium). At the same time, uncommitted consumption revenue trades at a discount to contractually fixed revenue, because in a recession it can shrink without a cancellation — the customer does not have to churn, only to consume less. For the DCF model that means concretely: consumption-based cash flows deserve a higher discount rate or more conservative growth assumptions than contractually committed ones — revenue beta rises.

Level 3: Portfolio — winners and losers of the shift

Structurally vulnerable: Horizontal per-seat software with high natural churn whose value proposition hangs on human operation — collaboration tools, generic productivity software, tools for entry-level knowledge work (data entry, simple case processing). Here the double lever applies: AI agents reduce the seats and force pricing concessions. The Eagle Point observation hits the core: when an agent replaces the junior clerk, the software for the junior clerk becomes obsolete too.

Structurally protected: Deep vertical software with non-seat value metrics, compliance anchoring, and proprietary data assets. The Constellation Software camp has been arguing exactly this for months: AI primarily threatens horizontal software with high churn rates, while for mission-critical VMS (emergency dispatch, hospital billing, utilities) it acts more as a feature accelerator — customers there are not buying code but regulatory reliability and process trust. Independent analyses estimate 40–50% of the CSU portfolio as the "fortress" quadrant (government, utilities, compliance-heavy healthcare). The market nonetheless prices the group like generic software — CSU's valuation correction (−26% in 2025, further declines in 2026 against ~15% revenue and ~20% EPS growth) is the bet that the market is punishing too indiscriminately here.

Already at the destination: Business models whose unit of account was never the seat but the physical asset. The fee per installed radio module, per monitored door, per connected device is formally consumption pricing — but with annuity stability, because the installed base does not decide anew about its consumption every day, and no AI agent replaces a physical fire-alarm radio module. These models (in my own coverage universe: NAPCO's device-bound RSR and the per-door pricing of MVP Access) have anticipated the end state of the pricing shift without ever having to pass through the migration J-curve. They carry no inference COGS, no seat erosion, and no transition costs.

Part 5: The Investor's Checklist

The dossier yields a concrete analytical grid. For every software position (and every candidate), five questions need answering:

First: What is the unit of account — and does AI replace it? Seats for routine knowledge work are the highest risk class. Transactions, assets, devices, beds, vehicles are robust metrics. Tokens are volatile, but at least value-aligned.

Second: How high is the committed share of revenue? Contractually fixed minimum commitments, multi-year deals and installed-base fees deserve the recurring multiple; spot-like consumption does not.

Third: Does the company carry inference COGS — and are they reflected in pricing? AI features under a flat rate are a ticking margin bomb. The question for every management team: how is inference-cost attribution per customer measured?

Fourth: Where does the company stand in the migration J-curve? Before the transition (risk underestimated), in the middle (revenue dip, but progress measurable), or through it (the NRR profile should have structurally improved)?

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