TL;DR — Key Takeaways
-
Your AI invoice tells you what you spent — not what you got. Flat-rate subscriptions can conceal underused capacity, making seat yield a potentially useful metric for understanding AI utilization.
-
One agency measured 16× seat yield from its AI subscriptions. RS Gonzales recorded $36,820 in AI usage valued at published API rates against a $2,300 monthly seat bill, although that figure does not represent actual savings.
-
Stop letting paid AI capacity sit idle. A proposed approach called seatsourcing routes work through opted-in colleagues’ existing AI seats, aiming to increase utilization without sharing credentials.
Most finance teams learn what their AI did last month from the bill. Deloitte’s third-quarter CFO Signals survey, published on 21 October, put the question to 200 North American CFOs. Of those, 53% said their organizations mainly review AI usage costs after the fact, through vendor bills, reports, dashboards or manual analysis. Only 20% use advanced controls, such as tracing spend to tokens or workloads.
For metered AI, that is workable, because the bill moves with use. For AI bought by the seat, the bill is almost the same every month. The invoice tells you that you paid for the same seats again. It cannot tell you whether they did a week of work or a year of it.
I am the chief operating officer of a U.S. digital marketing agency. Most of our team’s AI work runs on flat-rate Claude seats; a few automations still call a metered API, a small line on the bill. So I needed a number the seat invoice could not give me. I call it seat yield.
Two Ways to Pay for the Same Work
A business can buy AI in two ways. With the API, you pay for every request, and each new job adds to the bill. With a seat, you pay a flat monthly fee per person, and the vendor sets usage limits. The fee is the same whether the seat runs one job that month or fifty.
The first model is easy for finance teams to read; it behaves like a utility bill. The second is harder. Its cost is fixed, so the useful question is how much work the spending carries.
Here is the term as I coined it:
Seat Yield: the value of the AI work a seat delivers, priced at published API rates, divided by what the seat costs.
The comparison itself is not new; at least one FinOps glossary sets API list prices against what a firm pays, under other names. What was missing for us was one name and one ratio for a monthly report.
Our Baseline Month
Over the 30 days to 13 October 2026, the team’s Claude Code use was worth $36,820 at Anthropic’s published API prices; the agency paid $2,300 for its seats that month. That is a seat yield of about 16×.
The top half of the ratio came from the token counts in each person’s Claude Code logs for those 30 days. We priced them at Anthropic’s published per-token rates for each model. The bottom half is the seat bill.
Three warnings come with that figure. They are more important than the figure.
First, it is list-price value, not money kept. Had we paid per request, we would have used less. Seat yield is a measure of work carried by a fixed fee; it is not a bill we escaped.
Second, seats have caps. Every plan limits usage, so the fixed cost holds only inside those limits. A seat yield that keeps climbing will meet that ceiling sooner or later.
Third, it was measured before the change I describe below.
How to Compute It Monthly
A finance team can do this without new software:
1. Export the month’s usage in tokens, by model, from the AI tool’s admin console or its logs.
2. Price those tokens at the vendor’s published API rates for the same models, and note the date you read the price list.
3. Divide by the month’s seat invoice, including any extra-usage charges.
4. Report the ratio next to the bill every month, by team if you can.
A falling seat yield on a flat bill means paid capacity is sitting idle. A rising one means the same fee is carrying more work. Either is more useful than an invoice that says the same thing twelve times a year.
Why Seats Sit Idle, and What We Changed
A seat has one weakness: it is assigned to one person. When its owner is in meetings or on leave, the seat waits. Useful AI agents are also stuck on single laptops. On one day in October, the five most-used skills in our company were each held by a single person.
So I coined a second term: Seatsourcing.
Seatsourcing (n.): routing AI work to the Claude seats a team already pays for, colleague to colleague, so flat-rate seats work as one fixed-cost AI workforce. Every seat stays with its owner, and no login is ever shared.
In practice, a colleague picks an agent and a client in a simple form. The job waits in a queue. A teammate’s workspace that holds those agents, and has opted in, picks it up. Jobs run on each team member’s own Claude seat during the shared working day. Credentials are never shared; Anthropic’s terms say you may not share your login, and we don’t. The report goes back to the person who asked.
The queue went live on 31 October 2026. Its effect on our seat yield is not measured yet, and I won’t guess at it. The arithmetic is simple, though. The seat bill does not move when a colleague’s job runs on my seat, so each extra job lowers the average cost of a job. Seatsourcing raises a team’s seat yield: more of the work runs on seats already paid for.
What Finance Should Ask For
I am not neutral here. The agency I help run pays Anthropic for its seats. I also founded a company that provides marketing agencies with permanent team members trained in AI, including agentic AI specialists. But the metric is not tied to one vendor. You can price any flat-rate AI licence this way, as long as the vendor publishes per-use rates for the same models.
If your team reviews AI after the fact, as most of Deloitte’s CFOs do, add one line to that review. Put seat yield next to the invoice. The invoice tells you what you paid. Seat yield tells you what you got for it.
