A seat was a sensible billing unit for software that waited to be used. You bought a license, you paid per head, and the meter followed the org chart. If nobody opened the tool that month, the vendor still got paid. The cost of the product sat with the customer, and the vendor's gross margin barely moved whether the tool ran twice a day or two hundred times.
Agents broke the arrangement before anyone repriced it.
A seat bills a person. An agent bills a habit. When the work runs at three in the morning, retries four times after a bad tool call, and burns reasoning tokens the buyer never sees, the meter stops tracking headcount. It tracks activity, and the activity does not clock out. Vendors who kept selling seats are now paying for that inference out of their own margin.
The curve that fell
Inference did not stay expensive. GPT-4 shipped in March 2023 at $30 per million input tokens and $60 per million output. GPT-4o, fourteen months later, listed at $5 and $15. GPT-4o mini came in at 15 cents and 60 cents. Stanford's AI Index put GPT-3.5-class inference down about 280 times between November 2022 and October 2024, and Andreessen Horowitz described a cost curve falling roughly 10x a year at a fixed capability level.
Cheap tokens are good news for the buyer and a slow squeeze on the vendor. For a vendor billing by the seat, the top line does not move while the bottom line does. The cheaper the token, the more work the buyer points at it, and the more inference the vendor absorbs. Price per unit fell. Volume per seat rose faster.
A seat bills a person. An agent bills a habit.
The bill arrives either way
Employers refund a failed human hour, absorb it, or write it off. The vendor bills a failed agent run in full. The model does not know the task failed. It charges for every token it produced, including the ones that led nowhere, then charges again on the retry.
Set those two costs side by side and the disparity lands on the vendor. A loaded junior hour runs $60 to $120. A single failed agent run costs cents. The exposure is the fleet: thousands of cheap failures looping on a broken tool, at a volume no per-seat invoice was built to describe. The buyer sees a small number per run. The vendor sees the aggregate, and the aggregate is the invoice the vendor pays.
Three ways to meter it
Software has settled on three meters for agent work, and the margin sits in a different place in each.
- Seats are the incumbent. Anthropic sells Claude Team at about $30 per user per month, and OpenAI sells ChatGPT Business on a similar per-seat line. This meter is the one the buyer already understands, and it is the one that hides the inference risk from the customer and loads it onto the vendor.
- Outcomes are the newest. Intercom charges Fin at $0.99 per resolution. Salesforce prices Agentforce conversations in Flex Credits, around $2 for a conversation. Zendesk has moved support pricing toward the same resolution unit. The vendor now shares the risk: no resolution, no fee. The buyer likes the alignment. The vendor has to price the failure rate into the unit, because a resolution that takes eleven tool calls costs less margin than one that takes two.
- Usage is the oldest and the most honest. Anthropic and OpenAI have billed API tokens since the beginning. Cursor had to explain to its users in mid-2025 that a flat seat could not cover variable compute, and rebuilt its plans around usage. Cognition moved Devin from a flat $500 a month to a $20 entry plus metered units in April 2025, for the same reason. The meter that matches the cost is the meter the seller resists, because it hands the buyer the lever.
Pricing the loop
The three-year trade turns on who carries the inference when the agent works harder than the seat assumed. A vendor that prices the loop, and puts a floor under the retry, keeps its margin. A vendor that keeps selling seats is buying the customer's volume with its own compute, one quiet run at a time.
The failure mode is quiet. Gross margin on a seat-based agent product does not break in one quarter. It drifts, month over month, as usage climbs and the listed price holds. The curve was there the whole time, visible in every invoice the vendor paid and in none the vendor sent.
Whoever prices the loop owns the category. Whoever keeps selling seats gets to keep the inference, and the inference is the part that grows.