Cost per outcome: the metric that turns agent spend into a decision

Ask a finance leader what the company’s AI agents cost and you’ll usually get an answer: a model provider invoice, a platform subscription, a number of seats. Ask whether those agents were worth it and the room goes quiet. The cost side is visible. The output side isn’t, and without both there is no decision to make.

That gap is already shaping budgets. About one in five organizations limit AI use because of token and running costs, according to McKinsey. Capping spend without knowing what it buys is a blunt instrument: it cuts the agents that earn their keep along with the ones that don’t.

Why token bills and seat counts don’t answer the question

Token spend measures consumption. It rises when an agent handles more work, and it also rises when an agent gets stuck in a retry loop. On a bill, those look the same. Seat counts and platform fees say even less: they’re fixed, so they can’t tell you whether a workflow got cheaper or more expensive this month.

What a budget owner needs is a unit cost tied to a business result. What does it cost, all-in, to get one supplier invoice posted, one lead qualified, one new hire’s accounts provisioned — without a person stepping in to finish the job?

Defining cost per completed workflow

The definition matters more than the dashboard. Small choices in the numerator and denominator can move the result by a factor of two.

cost per completed workflow = total agent cost ÷ completed workflows

total agent cost    =   metered cost of every run,
                        including failed and escalated runs
                      + delegated sub-agent costs,
                        rolled up to the parent run
                      + platform fees allocated by run count,
                        labeled "allocated"

completed workflows =   runs where the business outcome happened
                        (posted, updated, provisioned)
                        without a human takeover

Three rules keep the numerator honest:

• Count every run, not just the successful ones. Failed and escalated runs consumed tokens and credits. Leaving them out makes a struggling agent look cheap.
• Roll up delegated work. When an orchestrator hands a step to a sub-agent — a vendor-master agent, a document extraction service — that cost belongs to the parent workflow.
• Label allocated costs as allocated. Platform fees billed monthly have to be spread across runs, by run count for example. That’s a reasonable modeling choice, but it is a choice. Show metered and allocated cost separately so nobody mistakes one for the other.

The denominator is stricter than “runs that ended without an error.” A workflow counts as completed when the business outcome happened — the invoice was posted, the record was updated — and no person took over to get it there. An approval that policy requires is part of the process. A clerk re-keying an invoice the agent abandoned is a takeover.

The companion metrics

Cost per outcome on its own can be gamed, most easily by pushing hard cases to people. Read it alongside four companions:

Segment by task type, not by vendor

A blended average across all work hides the decisions that matter. Consider an example from a demo environment with simulated data, in which two agents process supplier invoices.

On standard purchase-order invoices, a cheaper vendor agent costs €0.57 per completed invoice against €1.85 for the in-house agent, at a similar completion rate. On multi-line invoices, the same vendor agent completes only 64% without a takeover, against 83% for the in-house agent.

Compared per vendor, the picture is muddled. Compared per invoice type, it’s clear: route standard PO invoices to the cheaper agent and keep multi-line invoices with the agent that completes them. The routing decision is made per task type, not per vendor, and the projected saving should be labeled as modeled until it has been observed.

Splitting by agent version matters just as much. In the same demo data, a new version raised cost per completed invoice from €0.70 to €2.40 because of a retry loop on multi-line invoices. A monthly bill would have shown higher spend with no explanation. Unit cost, split by version, pointed straight at the release.

Four pitfalls

Treating correlation as causation

Cost per merged pull request is a popular metric for coding agents. It’s useful for spotting trends, but a merged change doesn’t prove the agent created the value: an engineer may have rewritten most of it, and changes vary enormously in size. Report it with revert rate and review effort, and don’t present it as return on investment.

Comparing units that aren’t comparable

A one-line invoice and a forty-line invoice are both “an invoice.” A ticket resolved and a ticket deflected to a help article are both “handled.” Define the unit of work for each task type before comparing agents, vendors or teams.

Ignoring human time

Every takeover costs someone’s time, and that time rarely shows up in the agent budget. If you can measure human handling time, add it and label it as an estimate. If you can’t, keep the takeover rate next to cost per outcome so the trade-off stays visible.

Hiding attribution uncertainty

Linking a completed outcome back to the run that produced it isn’t always exact. When an ERP posting is matched to a run by a record ID the agent returned, confidence is high. When it’s matched by timing alone, it isn’t. Show the link method and confidence, and report evidence coverage beside the metric. A precise number built on uncertain links is false precision — we cover why in two-sided evidence.

A checklist for the quarterly agent review

1. Cost per completed workflow for each agent, by task type and version, against last quarter.
2. Metered and allocated cost shown separately, with sub-agent costs rolled up to the parent workflow.
3. Straight-through and takeover rates for each task type.
4. Evidence coverage, with every unconfirmed or contradicted action listed.
5. Mandate compliance and open exceptions, each with a named owner.
6. Routing changes: which task types should move to a different agent, and the modeled saving.
7. Spend against caps, and whether any cap needs to change.
8. A decision for each agent — renew, re-scope or retire — and the date of the next review.

How ThetaLayer helps

ThetaLayer reconstructs every agent run end to end and reconciles it against your systems of record. Cost per completed workflow includes failed runs, sub-agent costs and labeled allocations, and every completion is backed by evidence with a stated confidence. Scorecards break the metric down by agent, version, team and task type, across vendors, so routing and renewal decisions rest on comparable numbers. Book a demo to see it on one of your workflows.

Sources: McKinsey, The state of AI in 2026 (Aug 2026). Example figures are from a ThetaLayer demo environment with simulated data and are illustrative only.