Cost Per Seat

Cost per seat is the cost of serving one licensed user of a product over a period, calculated by dividing the relevant delivery costs by the number of active seats. For conventional software it is a fairly stable number dominated by hosting, storage and support, and it is the denominator underneath gross margin on a seat-priced plan. AI features break the stability. When part of the cost of serving a user is inference that scales with what they do, cost per seat becomes a distribution rather than a constant. A heavy user of an assistant can consume many times what a light user consumes on the same subscription, so a plan priced per seat can be highly profitable for most customers and loss-making for a few, with the average concealing both facts. This creates a specific commercial risk. Seat pricing implicitly assumes bounded consumption per user, which was reasonable when the marginal cost of a user was near zero. If the product invites unbounded use — generation, agents, long documents, repeated retries — the customers who love it most are the ones who cost the most, and adoption improves retention while degrading margin. The metric that reveals this is not average cost per seat but its distribution across accounts. The usual responses are all reasonable and all have side effects. Fair-use limits keep the pricing model intact but must be discoverable before a customer hits them. Credits or usage allowances make consumption explicit at the cost of a more complicated purchase. Tiering by intensity of use aligns price with cost but requires the tiers to be legible to a buyer. Charging separately for AI features keeps the seat price clean but invites comparison with cheaper standalone tools. Improving the number is engineering rather than pricing. Prompt and retrieval efficiency, routing simple work to smaller models, caching, deduplication and sensible output limits all lower cost per seat without touching the price. So does removing the pathological cases: a small number of users triggering runaway loops or resending enormous contexts can move an account-level average on their own. Track it per account rather than in aggregate, alongside seat utilisation. An account paying for many seats that few people use has a low cost per seat and a renewal risk; an account with high utilisation and heavy AI use has good engagement and a margin problem. Those are opposite situations that an aggregate number reports as the same average.