AIReturn

The CFO & the AI budget

AI Unit Economics: Cost per Outcome, Not Cost per Token

AI unit economics measures cost per outcome — per resolved ticket, merged PR, or advanced deal — net of rework, not cost per token. Here's the CFO math.

Rodrigo Paredes BassiPublished 11 min read

TL;DR

  • AI unit economics is the cost to produce one real outcome — a resolved ticket, a merged PR, an advanced deal, a completed process — net of rework, not the cost of a token. Tokens are an input price; outcomes are what the budget is buying.
  • Cost per token is falling and nearly irrelevant to return. The unit that matters is cost per good outcome — cost over accepted, quality-gated results, not tokens burned or raw outputs counted. Raw counts inflate under AI, so cost per resolved task can climb even while cost per call falls.
  • Computing it requires attributing AI cost by skill, by model, and by product — not one aggregate token bill — then dividing by outcomes that shipped, minus the ones that had to be redone.
  • Worldwide AI spending is forecast at $2.59T in 2026 (Gartner), yet 78% of finance executives still can't fully tie that spend to outcomes (CloudZero). Cost per outcome is the unit that closes the gap.
  • The decision tool is the Cost × Rework matrix: plot each team on cost and rework, and the quadrant tells you whether cheap tokens are actually buying cheap outcomes — or hiding a redo bill.

What are AI unit economics?

AI unit economics is the fully-loaded cost to produce one unit of real business output with AI — one resolved support case, one merged pull request, one advanced deal, one completed process — measured net of rework. It reframes AI cost from an input price (cost per token) to an outcome price (cost per good outcome — accepted, quality-gated), so a CFO can compare spend against value, not just against last month's invoice. This is the BOFU, dollars-and-cents view of the CFO's guide to the AI budget: once you've decided to fund AI, this is the unit you underwrite it on.

Definitions: the terms this post uses

Output-per-dollar-net-of-rework is AIReturn's core ROI unit: the business output a team ships per dollar of AI spend, after subtracting the cost of redoing flawed AI work. Cost is the denominator; rework-adjusted output is the numerator. Its inverse — dollars per unit of that output — is cost per outcome. AI-spend-under-management (AISUM) is the total AI spend AIReturn measures and holds accountable for a customer — token and usage cost in v1, extensible to AI-tool subscriptions and agent compute. It is the full denominator of the unit-economics equation: every dollar in, not just the ones on one vendor's bill. Cost × Rework matrix is a 2×2 that plots each team on AI cost (high/low) against AI rework (high/low), producing a per-team verdict — scale, keep, fix, or cut. For unit economics, it's the tool that separates "low cost per outcome" from "low token price with a hidden redo tax." Workslop (BetterUp/Stanford's term, borrowed) is AI-generated output that looks polished but lacks the substance to advance the task, forcing the recipient to redo it. It is the industry's name for the symptom; AI rework is AIReturn's name for the measurable cost that inflates cost per outcome.

Why cost per token misleads

Cost per token is an input price, and it's the wrong unit for three reasons. First, it's falling and abundant. Token prices have dropped steeply, so optimizing them is optimizing the cheapest part of the stack. A CFO who drives token cost down 30% and cost per outcome up has lost money while the dashboard looked greener. Second, cheap tokens can produce expensive rework. A model that drafts a support reply for a fraction of a cent, but produces a reply that gets reopened, escalated, and rewritten by a human, did not save money — it moved the cost from the token line to the payroll line. This is why FinOps for AI answers cost, not return: allocating 100% of token spend tells you the bill, never whether the work shipped. The redo cost is the symptom the industry now calls workslop — self-reported at roughly $186 per employee per month for those who receive it (BetterUp Labs + Stanford, via HBR, 2025). Third, tokens don't map to outcomes. Two teams can burn identical tokens and ship wildly different amounts of accepted work. Cost per token can't see that difference; cost per outcome is defined by it. The through-line: a low cost per token with a high rework rate is a high cost per outcome wearing a disguise.

The unit that matters: cost per outcome, net of rework

Cost per outcome flips the denominator from tokens to accepted results — outputs that shipped and stayed shipped, not outputs that were produced. The precise unit is cost per good outcome: raw output counts inflate under AI — a model will happily produce more drafts, replies, and PRs — so a count-based denominator flatters the number. Gate it to accepted, quality-passing work. Cost per outcome = total AI cost ÷ (outcomes produced − outcomes reworked) One measurement trap sinks this if you let it. The naive estimate — average tokens per task × price per token — runs off by roughly 3–8×, because a real resolved unit drags retries, tool-calls, and multi-turn escalation behind it. So cost per resolved task can rise even as cost per call falls. The denominator has to be built from the actual traced tokens per resolved unit, retries included — not a per-call average. Two moves make it real. Name the outcome per team — a deal and a PR don't share units, so AI unit economics is computed per function, never as one org-wide number. And net out rework — subtract the outcomes that had to be redone before the ratio, because an outcome that gets reopened isn't a delivered unit; it's a delivered-then-undelivered unit that still cost money to produce.

FunctionThe "outcome" unitCost-per-outcome question
EngineeringMerged PR / closed ticketAI cost per PR that merged and didn't get reverted
SupportResolved caseAI cost per case closed that stayed closed
SalesAdvanced / closed-won dealAI cost per deal that moved forward and didn't regress
OperationsCompleted process runAI cost per process that finished without a manual redo
The unit is the outcome each team is paid to ship — not tokens, not activity, not "hours saved."

You can't compute it from an aggregate token bill

Cost per outcome has a denominator (outcomes) and a numerator (cost). Most stacks can't produce the numerator at the right grain. An invoice total tells you the bill, not which team, workflow, or model generated it. To divide cost by a team's outcomes, you need AI cost attributed by skill, by model, and by product — including multiple products at once — not one blended line. AIReturn does this via OpenTelemetry (OTEL) instrumentation: it captures usage cost at the level of the work, so the spend that produced a support team's replies is separable from the spend that produced engineering's code. In AIReturn's model this is v1 scope — AI-usage (token) cost across Copilot, ChatGPT, Claude, Cursor, and in-house agents, summed as AISUM. Fully-loaded human/salary cost is v2 and deliberately out of scope today, which keeps HR-surveillance sensitivity out while still netting rework out as the human-correction signal. Granular attribution alone, though, is table stakes — FinOps-for-AI tools already do it. What makes the denominator CFO-grade is that AIReturn's cost reconciles to the provider invoice and every figure is confidence-labeled (exact or estimated); because providers expose no per-user dollars and floor at daily granularity, per-team cost is honestly modeled from traced usage, then reconciled. The differentiation is not the split — it is joining that reconciled cost to the good outcome it produced. Without that granularity, "cost per outcome" is an average of averages — the exact blur that leaves 78% of finance executives unable to tie AI spend to outcomes (CloudZero, 2026).

A worked example (illustrative, not a benchmark)

The numbers below are illustrative — chosen to show the arithmetic, not to report a real result or a target. They demonstrate one point: the cheaper token can lose on cost per outcome. For scale, two figures a CFO will recognize from the market: Intercom's Fin agent is priced at about $0.99 per AI resolution, against a fully-loaded human cost of roughly $5.60 per ticket (public vendor/industry figures, illustrative anchors — not AIReturn measurements). A real per-team number must still be computed from actual traced tokens per resolved unit, retries included, never a per-call average. A support team runs the same volume of AI-drafted replies through two models. Model A is premium and pricier per token; Model B is cheap per token.

Model A (premium)Model B (cheap tokens)
AI token cost / month$8,000$3,000
Replies produced10,00010,000
Cost per token (relative)higher~60% lower
Rework rate (reopened / rewritten)10%35%
Accepted outcomes (net of rework)9,0006,500
Cost per accepted outcome$0.89$0.46
Human redo cost @ $25/redo$25,000$87,500
Fully-loaded cost per outcome$3.67$16.35
Illustrative figures. Cost per accepted outcome = AI cost ÷ accepted outcomes. Fully-loaded adds the human redo cost (reworked outcomes × redo cost) ÷ accepted outcomes; salary/labor cost is out of AIReturn's v1 scope and shown here only to make the rework tax visible.
On cost per token, Model B wins by a mile. On AI cost per accepted outcome, Model B still looks cheaper ($0.46 vs. $0.89). But once the rework it generates is priced in, Model B costs roughly 4× more per real outcome. The cheap tokens bought an expensive redo bill — invisible to any tool that stops at the token line.

The Cost × Rework matrix: reading cost per outcome as a decision

Cost per outcome is a number; the Cost × Rework matrix is the decision. Plot each team (or each model choice) on two axes — AI cost and AI rework — and the quadrant tells you what the ratio means.

Low reworkHigh rework
Low AI costKeep — cheap outcomes, low redo. Efficient.Fix — cheap tokens, expensive rework. The disguise: low cost per token, high cost per outcome.
High AI costScale — expensive but clean; return compounds. Fund it.Cut — high cost, high redo, little to show. Stop.
The bottom-left of a token-cost dashboard and the top-right of a cost-per-outcome view can be the same team. The matrix is what makes that visible.
The matrix is why unit economics can't be run on price alone. Model B from the example lands in Fix, not Keep: the low bill is real, but the rework makes the outcome expensive. A CFO who allocates on cost per token funds the disguise; one who allocates on cost per outcome, net of rework, funds the return. That is the step-by-step framework for measuring AI ROI applied to a single line item, and it's what turns a token invoice into a sample per-team AI ROI report.
Honest limit: rework is concrete in engineering — churn, reverts, reopens, and review rounds are already logged. For support cases, deals, and process runs, the proxies for rework are still maturing, and isolating AI-caused rework from a team's baseline redo rate is the hardest open question in this work. Name the boundary; don't over-claim a precise dollar of rework in every function yet.

FAQ

What are AI unit economics? AI unit economics is the cost to produce one unit of real business output with AI — a resolved ticket, a merged PR, an advanced deal, a completed process — net of rework. It reframes AI cost from an input price (cost per token) to an outcome price (cost per good outcome — accepted, quality-gated), computed per team, so spend can be compared against value. Worldwide AI spend is forecast at $2.59T in 2026 (Gartner), which makes the unit the budget rides on worth defining precisely. Why is cost per token a bad AI metric? Because it prices an input, not a result. Token cost is falling and abundant, so optimizing it optimizes the cheapest part of the stack. Worse, cheap tokens can produce output that gets reworked — moving cost from the token line to payroll. A low cost per token with a high rework rate is a high cost per outcome in disguise. Measure cost per accepted outcome instead. How do you calculate cost per outcome for AI? Cost per outcome = total AI cost ÷ (outcomes produced − outcomes reworked). Name the outcome each team ships (deals, PRs, resolved cases), attribute AI cost by skill, model, and product — not one aggregate bill — then subtract the outcomes that had to be redone before you divide. Run it per team, since a deal and a PR don't share units, and read it against the team's own baseline, not an industry average. Build the cost from actual traced tokens per resolved unit (retries included), not a per-call average, or the denominator can be off by multiples. Do cheaper tokens lower AI cost per outcome? Not necessarily. Cheaper tokens lower the input price, but if the cheaper model produces more rework, cost per accepted outcome can rise. In an illustrative comparison, a model that was ~60% cheaper per token cost roughly 4× more per real outcome once its higher rework rate was priced in. Judge models on cost per outcome net of rework, not on token price. How is AI unit economics different from FinOps? FinOps allocates and optimizes the cost — it answers "what did AI cost, and where." Unit economics divides that cost by outcomes net of rework — it answers "what did each result cost, and was it worth it." FinOps is the denominator done well; unit economics adds the numerator. That's why 78% of finance execs can still allocate every dollar and not tie it to outcomes (CloudZero, 2026).

Sources

See the return on your own numbers.

Get a clear, per-team return-on-AI picture in weeks — output net of rework, reconciled to your invoice.