By the end of 2027, run-based billing will be a standard SKU across the leading enterprise AI agent platforms

The prediction

By December 31, 2027, a majority of the leading enterprise AI agent platforms will offer run-based pricing as a standard, publicly listed option — not a custom enterprise-only arrangement. By "run-based" I mean a priced unit that sits above the raw token and maps to work rather than to computation: pricing per completed task, per-agent-hour or per-run allowances, capped run budgets, or outcome-based fees. The platform may still expose a token meter underneath and may still sell seats, but a work-mapped unit will appear on the standard price sheet or documented pricing tiers of most major agent vendors, marketed as the way to buy and forecast agent usage.

If, at the end of 2027, the leading agent platforms are still billing predominantly by per-seat subscription and raw per-token metering — with no widely available, publicly listed run-, task-, or outcome-based unit — this prediction is wrong.

Why I think this happens

The prediction follows from a structural mismatch I laid out in the run-cost era of AI agents: an agent does not answer a request, it runs a task, and a single task can span hours and billions of tokens as the agent loops and resends its growing context at every step. Gartner has put agentic task consumption at roughly 5 to 30 times a single chatbot request, and the autonomous horizon that METR tracks is doubling every four to seven months, so tasks are getting longer and more token-hungry even as price per token falls.

That mismatch breaks both incumbent pricing models in opposite directions, which is exactly the condition that forces a new unit to emerge. Seat pricing decouples cost from headcount — one user can dispatch a task that burns more compute than the seat covers in a year — so vendors either cap capability or bleed on power users. Raw token pricing fixes the vendor's exposure by handing unbounded variance to the buyer, who then cannot forecast a line item that swings two orders of magnitude per invocation. Enterprises will not sign unbudgetable contracts, and vendors will not eat unbounded variance. The stable resolution is a unit that gives the buyer a forecastable number while letting the vendor pass variance through in aggregate — which is precisely what a task, a run budget, or an agent-hour allowance does.

The market is already moving this way. The capital committed in mid-2026 — Sail Research raising 80 million dollars specifically to make long-horizon agent inference cheaper, voice-agent companies reaching unicorn valuations on run-intensive production deployments — is a bet that run cost is the binding constraint, and pricing follows the constraint. This is the general case of a shift I already flagged in the end of the seat: once the cost of a product scales with work dispatched rather than with people, the seat cannot hold and something metered to the work takes its place.

The reason this is a tier-2 call at 68 rather than a near-certainty is definitional drift and vendor inertia. "Standard SKU" is a judgment call — a vendor can bury run-based pricing in enterprise-only contracts, or dress up a token meter with a forecasting dashboard and call it run-based. Pricing pages also change slowly, and some incumbents will cling to seats for as long as customers tolerate it. The direction is high-confidence; the pace of formal, publicly listed adoption across a majority of leading platforms within eighteen months is where the real uncertainty sits.