The metric
What is cost per outcome?
Cost per outcome is AI spend divided by a unit of work the business already counts, over the same period. Not per seat, which is flat whether a licence is used or not. Not per token, which rises when a model is inefficient. Per thing produced.
It is the only AI metric that settles a budget argument, because it is the only one where both sides of the table are looking at the same fraction. Finance has the numerator. Engineering has the denominator. Neither number means much alone, and together they are a decision.
Why the usual metrics fail
Per seat and per token both answer a different question
| Metric | What it tells you | Where it misleads |
|---|---|---|
| Cost per seat | What you committed to buy. | Identical for a seat used daily and a seat never opened. It cannot see the waste it is made of. |
| Cost per token | How efficiently you are buying inference. | Falls when a model becomes verbose and cheap. More tokens is not more delivered work. |
| Total AI spend | The size of the bill. | Grows with headcount and with success alike, so it cannot distinguish them. |
| Cost per outcome | What a dollar of AI actually bought. | Gameable if the denominator is optimized directly, which is why it never travels without its quality signals. |
How to compute it
Four decisions, made explicitly
- Pick a denominator the business already counts. Merged pull requests, resolved tickets, shipped releases. If you have to invent the unit to compute the metric, the metric will not survive its first review.
- Attribute the whole numerator, including metered spend. Seats attribute through the licence roster. Tokens need a tag or a key per team. Spend with no owner is where waste survives, because nobody has to defend it.
- Use the same period on both sides. A month of spend over a quarter of output is not a ratio, it is a press release.
- Publish the quality signals beside it. Rework, bugs per change, incidents, review depth. A cost-per-outcome figure that improved while all four got worse describes a trade, and reporting only the first half is the standard way this metric lies.
Each figure in that calculation carries a proof level: the spend is Counted, a between-team difference is Compared, and any dollar value placed on an outcome is Estimated and shown as a range.
Definitions
The terms, precisely
- Cost per outcome
- AI spend divided by a unit of delivered work over the same period. The denominator is a thing the business already counts, such as a merged pull request, a resolved ticket, or a shipped release, rather than a proxy such as seats or tokens.
- Cost per merged PR
- Total AI spend attributed to a team, divided by the pull requests that team merged in the same period. The most common concrete form of cost per outcome in an engineering organization.
- Cost per seat
- AI spend divided by licensed users. It measures what you bought, not what it produced, and it is flat whether a seat is used daily or never opened.
- Cost per token
- AI spend divided by tokens consumed. A useful efficiency measure for a routing decision and a poor measure of value, because more tokens is not more delivered work.
Common questions
Using cost per outcome
- What is cost per outcome?
- AI spend divided by a unit of work the business already counts, over the same period. In an engineering organization that is usually merged pull requests, resolved tickets or shipped releases. It answers what a dollar of AI bought, which is the question cost per seat and cost per token both dodge.
- Why not measure cost per seat?
- Because it is flat. A seat costs the same whether someone uses it every day or has never opened it, so the metric cannot distinguish a well-adopted tool from a wasted one. It is a budgeting number, not a value number, and treating it as the latter is how idle seats survive three renewals.
- Does this mean you should push engineers to merge more pull requests?
- No, and a team that optimizes for the denominator will produce a better number and worse software. Cost per outcome only stays honest paired with the quality signals that move in the other direction: rework, bugs per change, incidents, review depth. A throughput figure without those is a target waiting to be gamed.
- What is a good cost per outcome?
- There is no cross-company benchmark worth quoting, because the denominator is defined differently everywhere and pull requests are not a standard size. The number is useful as a TREND inside one company on a stable definition, and as a comparison between matched teams in the same company over the same period. Anyone selling you an industry average has picked a definition you did not agree to.
- How do you attribute shared AI spend to a team?
- Seats attribute directly through the licence roster. Metered spend needs a tag or a key per team, which is worth setting up because unattributed spend is where waste survives. Where attribution is genuinely impossible, allocate it explicitly and label the result as modeled rather than folding an assumption into a number that looks counted.
Get the numerator before you argue about the denominator
Most companies cannot compute this because the spend side is spread across vendors and unattributed. That part is countable, and 10 to 15% of seat and subscription spend usually turns out to be recoverable on the way.
Related: proof levels · the AI cost management guide · what the research says about AI and developer speed · what an AI spend audit is