AI-native businesses price outcomes, not seats

Seat pricing has one assumption baked in: value scales with the number of humans logged in. AI products exist to break that assumption. When the product does the work, the customer's rational move is fewer seats producing more output — which means a seat-priced AI product gets punished for succeeding. Pricing has to follow the work, and the work is no longer measured in humans.

You can watch the migration happening in the pricing pages. Token-based pricing at the model layer, with its own asymmetries — output tokens priced several multiples above input. Per-action pricing for agents. Per-agent-per-month pricing, which is seat pricing where the seat is synthetic. Subscription-plus-overage hybrids. Autonomy-tiered pricing, where an agent that acts unsupervised costs more than one that drafts for review. Each of these is an attempt to re-anchor price to something that still correlates with value once headcount doesn't.

Metering becomes load-bearing infrastructure

Here's what the pricing-page view hides: every one of those models demands infrastructure that seat-priced SaaS never needed. If you bill on usage, you need event-based metering with idempotency keys so a retried request isn't billed twice. You need a gateway that attributes every LLM call across levels — API key, user, team, end customer — the way LiteLLM's spend logs do. You need an aggregation backbone (the reference stacks — Lago, OpenMeter — sit on ClickHouse for a reason) turning raw events into billable metrics. And you inherit a new class of failure mode that is simultaneously a revenue bug and a trust incident: duplicate events, stale price tables silently applied to a repriced model, tokenizer discrepancies where your count and the provider's count disagree. In a seat-priced company, billing is a finance concern. In a usage-priced company, billing is a distributed data pipeline, and it had better be more reliable than the product it meters.

Margin discipline is now an engineering requirement

The second thing seat pricing never taught SaaS companies: cost of goods sold. Classic software had near-zero marginal cost, so gross margin was a spreadsheet formality. Every agent run has real COGS — inference, tools, retries — and it varies per customer, per task, per day. That makes per-tenant gross margin a metric engineering has to instrument, not one finance can derive. Cost-aware pricing stops being a strategy-deck phrase and becomes an architectural requirement: the system must know what each outcome cost before it can price the outcome. Model-price deflation — roughly an order of magnitude per year on a cost-per-million-tokens basis — helps and hurts at once: your costs fall on a schedule, but so does every competitor's floor, so margin captured through laziness evaporates.

The concession is real, though: buyers hate unpredictable bills. Pure usage pricing shifts volatility onto the customer, and procurement departments push back hard — it's exactly why subscription-plus-overage and committed-volume discounts exist, and why they'll persist. And seats remain the honest model where the human genuinely is the unit of value: collaboration tools, communication tools, anything where the product amplifies a person rather than replacing a task. Outcome pricing is for products that do work, not products that host work.

But if you're building the former, the sequencing matters. Pricing model, metering pipeline, and cost attribution are one decision, and it's an architecture decision. Retrofitting metering onto a shipped product means reconstructing usage you never recorded. The companies that price outcomes credibly will be the ones that built the meter before they needed the money.