The End of the Per-Seat AI Pricing Model

Per-seat pricing is the dominant SaaS pricing model of the last twenty years and it is structurally incompatible with AI products. The reason is variance. Traditional SaaS usage variance across seats is roughly 2 to 4x; power users use the product more than light users, but the cost-to-serve a power user is roughly 1.2 to 1.5x the cost-to-serve a light user, so per-seat pricing absorbs the variance acceptably. AI product usage variance across seats is 10 to 50x, and the cost-to-serve scales linearly with usage because most call costs inference dollars. A seat consuming 50,000 tokens/month costs the vendor roughly $0.02 to serve. A seat consuming 500,000 tokens/month costs roughly $0.20. A seat consuming 5,000,000 tokens/month costs roughly $2.00. Per-seat pricing forces the vendor to either price for the heaviest user (and lose the light-user market) or price for the median (and lose money on the heavy 20 percent). Neither survives. The end state is consumption pricing, with seats as a wrapper feature for governance rather than a pricing primitive. This piece explains why per-seat fails for AI, what the transition looks like, and why most B2B SaaS vendors will move to consumption pricing by 2027.

This is a spoke under the AI project economics manifesto. The manifesto’s central claim; that evaluation cost replaces feature cost; implies a corresponding change on the pricing side: usage-based cost-to-serve replaces seat-based revenue capture. This piece works that implication.

Why per-seat worked for traditional SaaS

Per-seat pricing dominated SaaS for two decades because three properties of traditional software made it work:

1. Cost-to-serve was roughly constant per seat. Storage, compute, and bandwidth scaled gently with user count. Heavy users cost 1.2 to 1.5x light users; close enough that a flat per-seat price worked.

2. Usage variance was bounded. Even on consumer-leaning enterprise tools (Slack, Notion, Asana), the heaviest user generated maybe 4x the activity of the median. Variance was visible but tolerable.

3. Procurement liked seat counts. Enterprise procurement had built two decades of policy around seat-based licensing. Quarterly seat reviews, departmental allocation, headcount-based budgeting; seats integrated cleanly with HR and finance systems.

The model worked so well that it became the default. SaaS startups in 2010 to 2022 priced per-seat without thinking, because the alternative (usage-based) was harder to budget, harder to procure, and produced unpredictable revenue.

Why AI breaks the per-seat assumption

The two underlying assumptions; constant cost-to-serve and bounded variance; both break for AI products. The breakage is not subtle.

Cost-to-serve scales linearly with usage. Most AI call has a marginal inference cost. A seat that asks the AI to summarize 1000 documents costs the vendor 1000 inference calls. A seat that asks the AI to summarize 1 document costs 1 inference call. The 1000x usage difference produces a roughly 1000x cost-to-serve difference (with some non-linearity from caching and batching, but the linear baseline is real).

Usage variance is unbounded. Heavy AI users in 2026 are running agentic workflows that fire 50 to 500 calls per task and execute hundreds of tasks per day. Light users are running 5 to 20 chat sessions per week. The variance ratio between heavy and light is 100 to 1000x, far beyond what a flat per-seat price can absorb.

The variance is also structural. It’s not “some users are heavier”; it’s “some use cases are 100x more expensive than others.” A team using the AI for one-shot Q&A and a team using the AI for autonomous research agents are paying the same per-seat price for products with 100x cost-to-serve differences. The vendor either over-charges the Q&A team into churn or under-charges the agent team into bankruptcy.

The 10-50x usage variance problem

Real numbers from AI products we’ve audited in 2025 to 2026:

Product type	Median seat (tokens/month)	90th percentile seat	99th percentile seat	Variance (99/median)
AI writing assistant	80K	400K	2.5M	31x
AI coding copilot	250K	1.4M	8M	32x
AI agent platform	150K	1.5M	18M	120x
AI customer support	60K	350K	1.8M	30x
AI research tool	200K	1.2M	6M	30x

A 30x variance ratio is roughly 10x what traditional SaaS sees. A 120x variance ratio (agent platforms) is roughly 30x what traditional SaaS sees. Per-seat pricing simply cannot absorb these variance levels without breaking on one side or the other.

The problem is not solvable by tiered seats either. A “heavy seat” tier at 5x the price still loses money on the 99th-percentile user (variance 30x means a 5x price still produces a 6x cost gap on the heaviest cohort). A “team seat” plan with shared usage is a step toward consumption pricing; at which point the seat construct is just procurement window dressing.

The two failure modes of per-seat AI

Vendors that try to hold per-seat pricing on AI products fail in one of two ways:

Failure mode 1: Price for the heaviest user. Vendor sets the per-seat price at $80/month based on what the heavy users cost to serve. The light-user market; which is the majority of total addressable users; refuses to buy because the value-per-seat at light usage doesn’t justify $80/month. Vendor gets stuck in a small high-end niche, rarely expands, and is eventually beaten by a usage-priced competitor that can serve the long tail at a price the long tail will pay.

Failure mode 2: Price for the median. Vendor sets the per-seat price at $20/month based on median usage cost. Heavy users overwhelm the unit economics; vendor’s cost of inference on the top 10 percent of users exceeds the revenue from those users, and gross margin compresses below 20 percent. Vendor either has to raise prices (and lose price-sensitive customers), introduce heavy-handed rate limits (and frustrate the heaviest users into churn), or migrate to usage pricing (which is what they should have done at the start).

Both failure modes are visible in the market right now. GitHub Copilot’s $19/month price is failing on the second mode for heavy users. Microsoft’s M365 Copilot $30/seat is failing on the first mode for light users. ChatGPT Enterprise’s $60/seat is failing on both modes simultaneously depending on the customer cohort. The compression in seat-based AI pricing is structurally inevitable.

What’s already moving away from seats

Several major AI vendors have already migrated, partially or fully, away from per-seat pricing:

• Anthropic Claude API: Pure usage-based since launch, no per-seat construct.
• OpenAI API: Pure usage-based; per-seat ChatGPT Enterprise is a separate, increasingly secondary product line.
• Cursor: Per-seat with a hard cap that triggers usage-based overage; effectively a hybrid that admits the per-seat ceiling.
• Replit Agent: Pure usage-based for the agentic workload.
• Vercel v0: Credit-based, which is a thin wrapper over usage pricing.
• Glean, Harvey, Hebbia: Hybrid commit-plus-overage models that resemble enterprise consumption pricing more than per-seat.
• GitHub Copilot: $19/seat on the surface, but Microsoft has been quietly adding usage caps and tiered “agent mode” pricing that are migration steps.

The vendors holding pure per-seat AI pricing in 2026 are mostly the ones with the smallest variance; narrow-use-case productivity tools where the gap between median and heavy users is smaller. Even those are migrating as their feature sets expand into agentic workloads.

What replaces it: consumption pricing

The dominant pattern is converging on a consumption-pricing model with three properties:

1. A metered unit of value. Tokens, calls, evals, jobs, or “credits” that wrap multiple metering primitives. The unit must correspond to cost-to-serve so unit economics work, and to value-delivered so customers can budget against outcomes.

2. A commit-plus-overage structure for enterprise. Customers commit to N units of usage per year at a discounted rate; usage above commit is billed at overage rate. This gives customers budget predictability and gives the vendor revenue predictability while still scaling revenue with heavy users.

3. A fair-use limit for self-serve tiers. Self-serve plans have a flat price up to a usage cap, beyond which usage-based billing kicks in. This preserves the “predictable monthly bill” experience for the median user while protecting the vendor from heavy users on flat plans.

The full playbook for migrating to this model; metering choices, commit-and-overage design, fair-use limits, and procurement adaptation; is covered in the AI consumption pricing playbook.

The seat construct survives, but as a governance feature rather than a pricing primitive. Customers still want to know “who in my org is using this,” and seat assignments are how that is tracked. But the bill is computed against usage, not seat count.

The 2026 to 2027 transition timeline

The transition from per-seat to consumption pricing is happening at three layers, on different timelines:

• API-first AI products (already migrated, 2023-2024). Anthropic, OpenAI, Mistral, Together; many priced usage-based from the start.
• Agentic AI platforms (migrating now, 2025-2026). Cursor, Replit, Vercel, Glean; moving from seat to consumption or hybrid models. The migration is ~70 percent complete by mid-2026.
• Embedded AI features in traditional SaaS (migrating 2026-2028). Salesforce, Microsoft, Google Workspace, Slack; these are the slowest because they have decades of seat-based procurement infrastructure to navigate. The migration started in 2025 and will be largely complete by 2028.

The forcing function in many three layers is the same: seat-based pricing produces structurally bad unit economics on heavy users, and the vendor’s CFO eventually demands the migration. The vendor that holds out longest gets the worst margin compression and the strongest competitor responses. The pricing-models-by-alignment-with-outcomes piece shows where consumption pricing ranks against alternative AI engagement structures.

By end of 2027, per-seat AI pricing as the primary pricing model will be a minority position in the market. Seats will persist as a governance overlay on consumption pricing for enterprise procurement reasons, but the pricing primitive will be units of usage.

Frequently asked questions

Why does per-seat pricing fail for AI products?

Two reasons. Cost-to-serve scales linearly with usage in AI products because most call has marginal inference cost; unlike traditional SaaS where cost-to-serve is roughly flat per seat. Usage variance across seats is 10 to 50x, often higher for agentic workloads, versus 2 to 4x in traditional SaaS. Per-seat pricing cannot absorb the variance: the vendor either prices for heavy users (and loses the light-user market) or prices for median users (and loses money on heavy users).

What is the typical usage variance across seats in AI products?

The 99th percentile seat consumes 30 to 120x what the median seat consumes, depending on product type. AI writing assistants and customer support tools sit at the lower end (around 30x). AI agent platforms sit at the higher end (around 120x). The variance ratio is roughly 10 to 30x what traditional SaaS sees, and per-seat pricing breaks at this variance level.

What’s wrong with tiered seat pricing as a fix?

Tiered seats compress the variance problem but do not solve it. A heavy-tier seat at 5x the price still loses money on the 99th-percentile user when variance is 30x. Tiering creates more pricing complexity for a smaller share of usage; pure consumption pricing solves the variance problem at the unit level instead of trying to bucket users into tiers.

What replaces per-seat pricing?

Consumption pricing with three properties: a metered unit of value (tokens, calls, evals, jobs), a commit-plus-overage structure for enterprise (predictable budget plus elastic upside), and a fair-use limit for self-serve tiers (flat price up to a cap, then usage-based above the cap). The seat construct survives as a governance feature for tracking who in an organization is using the product, but the bill is computed against usage.

Will per-seat AI pricing disappear entirely?

Not entirely. Per-seat survives in narrow-use-case productivity tools where usage variance is genuinely small (basic AI writing assistance, simple summarization). It also survives as a governance overlay on consumption pricing for enterprise procurement reasons. What disappears is per-seat as the primary pricing primitive for AI products. By end of 2027 it will be a minority position.

Which vendors have already migrated to consumption pricing?

Anthropic, OpenAI, Mistral, and Together are pure usage-based since launch. Cursor, Replit, Vercel v0, Glean, Harvey, and Hebbia have hybrid or pure consumption models. GitHub Copilot, Microsoft M365 Copilot, and ChatGPT Enterprise still hold per-seat as the headline price but have added usage caps and tiered overage models that are partial migrations.

What about enterprise procurement that’s built around seat counts?

Enterprise procurement adapts on a multi-year timeline. The transitional pattern is commit-plus-overage: customer commits to a unit-volume per year (which procurement treats much like a seat-count commit), with overage billed elastically. This gives procurement the predictable annual budget they need while giving the vendor consumption-aligned unit economics. By 2028, most enterprise procurement will have policies for unit-based commits.

How do I price a B2B AI product correctly today?

Use a hybrid model. Self-serve tier with a flat monthly price and a fair-use cap. Mid-market tier with a small commit and overage above commit. Enterprise tier with a substantial commit, overage rates, and seat-based governance overlay. This structure works for the full range of customer sizes and avoids the per-seat failure modes. The detailed playbook is in the consumption pricing playbook.

What’s the unit economics target on a consumption-priced AI product?

Gross margin of 60 to 75 percent on the metered unit, with the metered unit priced at roughly 4 to 7x cost-to-serve. The wider gap accommodates support, infrastructure, customer success, and the share of the price that funds non-inference value (eval discipline, observability, integration). The cost-per-action framework covers the unit economics in detail.

Will consumption pricing produce volatile revenue?

Less than vendors fear. Commit-plus-overage gives 60 to 80 percent of revenue predictability through commits, with 20 to 40 percent elastic upside through overage. Aggregate quarterly revenue is typically less volatile than seat-count revenue because overage growth on heavy customers smooths over seat-count adjustments on light customers. The volatility concern is mostly transitional anxiety; mature consumption-priced AI businesses have revenue forecasts with comparable accuracy to seat-priced ones.

Key takeaways

• Per-seat pricing fails for AI because cost-to-serve scales linearly with usage and usage variance across seats is 10 to 50x, versus 2 to 4x in traditional SaaS. The two underlying assumptions of per-seat pricing both break.
• The two failure modes of per-seat AI pricing are pricing-for-heavy-users (loses the light-user market) and pricing-for-median-users (loses money on heavy users). Both are visible across major AI vendors in 2026.
• The replacement is consumption pricing with three properties: a metered unit of value, a commit-plus-overage structure for enterprise, and a fair-use limit for self-serve tiers.
• The transition is happening at three layers on different timelines: API-first AI (already migrated), agentic platforms (migrating 2025-2026), embedded AI in traditional SaaS (migrating 2026-2028).
• Per-seat pricing survives as a governance overlay for enterprise procurement, but by end of 2027 it will be a minority position as a primary pricing primitive for AI products.

The death of per-seat AI pricing is not a near-future prediction; it is a present-tense observation that is roughly 60 to 70 percent complete in the API and platform layers and 20 to 30 percent complete in the embedded-AI layer. Vendors that lean into the migration capture the heavy-user revenue and serve the light-user market simultaneously. Vendors that hold out get squeezed on margin until the CFO forces the change.