AI
The AI pricing problem: why 'per seat' is dying and what replaces it
Software has been sold by the seat for thirty years because people were the unit of work. AI breaks that assumption — and the industry's scramble for a replacement will change how you buy everything.
There's a spreadsheet somewhere in every software company titled something like "pricing v14 FINAL final" and, as of early 2026, it has never been more contested. The reason is simple to state and awkward to solve: AI has broken the thirty-year-old logic of charging per seat, and nobody has fully agreed on what comes next.
Why the seat made sense — until now
Per-seat pricing was never really about seats. It was a proxy for value: more employees using the tool meant more work being done in it, so charging per human scaled neatly with the benefit. It was predictable for buyers, compounding for vendors, and easy for procurement to audit. A genuine win-win, which is why it survived three decades of otherwise ruthless change.
AI breaks the proxy. When software does the work rather than merely hosting it, value no longer tracks headcount — it can be inverse to it. A support platform whose AI resolves 60% of tickets makes each human agent more productive and, awkwardly for the vendor, means you need fewer of them. Charging per seat for a product that reduces seats is a business model eating itself.
The cost problem underneath
There's a second, grubbier issue: marginal cost. Traditional SaaS cost pennies to serve another user, so flat pricing was safe. Every AI request, by contrast, burns real compute — the electricity and silicon behind that are their own story, told in our inference-chips explainer and our look at data centres and the UK grid. A flat-fee customer who hammers the model can be genuinely unprofitable. Vendors learned this the expensive way in 2023–24, when several launched unlimited AI features and quietly capped them within a year.
Charging per seat for a product that reduces seats is a business model eating itself.
The three contenders
Usage-based. Pay per token, per document, per minute of transcription. This is how the model providers themselves charge, so it maps cleanly onto vendor costs. The buyer's problem is predictability: usage bills spike, and finance directors hate surprises more than they hate high numbers. The cloud-computing decade taught everyone what an unmanaged consumption bill looks like.
Outcome-based. Pay per result: per resolved support ticket, per qualified lead, per completed hire. Intellectually elegant — you pay for value, not activity — and increasingly common where outcomes are cleanly measurable. The catch is attribution. Vendors want credit for every resolution; buyers notice the AI cherry-picks easy tickets. Expect contract disputes about what "resolved" means to keep employment lawyers comfortable for years.
Hybrid. The pragmatic winner so far: a per-seat platform fee for access plus a bundle of usage credits, with overage rates on top. It preserves predictability for the base load and meters the AI on top. As of early 2026 this is the emerging default across mainstream business software, and most "AI agent" products — the category we audited in our agents explainer — launch straight onto credits or per-task pricing without ever offering a seat.
The wildcard: local models
One more force pushes the same direction. As capable small models move onto laptops and phones — the shift covered in Small models, big shift — the marginal cost of routine AI drops towards zero for tasks that never touch a vendor's servers. That undercuts usage pricing for commodity work and pushes vendors to charge for what genuinely requires their infrastructure: frontier reasoning, proprietary data, integrations and accountability. The long-run shape is visible: cheap-or-free local AI for the routine, metered cloud AI for the hard, and vendors earning their margin on trust and workflow rather than raw tokens.
What buyers should do about it
First, ask every AI vendor three questions: what is the unit of charging, what does a typical customer of our size consume, and what happens — contractually, not reassuringly — when we spike. Second, treat AI consumption like a utility: set budgets, alerts and an owner, exactly as sensible firms already do with cloud spend. Third, in multi-year deals, negotiate the credit price now; unit costs of inference have fallen steadily and your contract should share in that decline, not lock in 2025 economics.
The seat isn't vanishing overnight — inertia is the strongest force in enterprise software. But its monopoly is over. The next time a renewal lands on your desk, the interesting number won't be the headcount. It will be the meter.