Back to all posts

Cost Per Outcome: The FinOps Metric Agents Need

2026-10-01
luodaInt
Cost Per Outcome: The FinOps Metric Agents Need

Cost Per Outcome: The FinOps Metric Agents Need

Token dashboards are easy. Showing the CFO what one successful business result actually cost — retries, tool calls, and human review included — is the hard part.

Pilot purgatory is often a process problem: agents rehearsing workflows that never hit a cost or revenue line. Once something does run in production, a second failure shows up. Teams can see total LLM spend and still cannot answer a simple question: what did one accepted outcome cost? That gap is where FinOps for agentic systems starts.

The useful unit

Cost per query prices a model call. Cost per task prices a run whether it worked or not. Cost per outcome puts every attempt in the numerator and only verified business results in the denominator. Finance lives on the third rung.

Tokens measure activity; outcomes measure return

BCG (Joppe Bijlsma, Djon Kleine, and Filippo Scognamiglio, July 1, 2026) argues that the first wave of AI governance was about access. The next wave is economics. Traditional FinOps was built for infrastructure and SaaS seats. Agentic workflows break that meter: prompt length, retrieved context, output size, model choice, reasoning effort, tool use, cache hits, and how many loops an agent runs before it stops all move the bill.

Their recommended ratio is return on AI (RoAI):

RoAI = Economic return / (Cost of human intelligence + Cost of tokens)

The optimization target is not cheapest tokens. It is cost per successful outcome at the quality, latency, risk, and human-review burden you actually require. BCG names four forces that make unmanaged agentic spend hard to forecast: breadth of adoption, task intensity, context-and-loops (agents carry instructions, history, and tool outputs forward and often re-send them each turn), and model mix (defaulting to the newest frontier model or largest context window is a design choice with a billing impact — and a weaker model that retries into success can cost more per outcome than a stronger one that finishes cleanly).

TechTarget (Kashyap Kompella, May 29, 2026) makes the same unit shift for ops: a chatbot is mostly prompt-response tokens; an agent plans, calls tools, retrieves data, coordinates sub-agents, and may retry before a case closes. The question changes from “How much does the model cost?” to “What does it cost to complete this business outcome reliably, securely, and at acceptable quality?”

On Hacker News, Deborah Jacob (botanu) put the instrumentation pattern in engineering language: an event_id for business intent (resolve a ticket, generate a report), a run_id per attempt, all runs sharing the event, a final success/failure/partial outcome, and cost per outcome = sum of all runs for that event, including failed attempts, with context propagated via OpenTelemetry baggage. The idea is not a product endorsement. It is the join finance needs: spend and results on the same business event.

The unit-cost ladder (and why vendors already climbed it)

CloudZero (Keith MacKenzie, September 28, 2026) lays out a three-rung ladder that matches how engineering and finance talk past each other:

| Rung | What it prices | How failures count | | --- | --- | --- | | Cost per query | One model call | N/A — a query is a single call | | Cost per task | One complete agent work unit | Failed tasks still sit in the denominator | | Cost per AI outcome | One business result that met a written definition | Failures live in the numerator, never the denominator |

An agent that attempts everything and resolves little can look fine on cost per task and terrible on cost per outcome. Only the second number survives a budget review.

Vendors already price the outcome. CloudZero’s September 2026 check of first-party rate cards (enterprise and regional pricing differ) lists examples such as Fin at $0.99 per outcome (50-outcome monthly minimum; a configured human handoff can count as billable), Salesforce Agentforce at $2.00 per conversation (bills whether or not it resolves), and Zendesk AI agents at $1.50 committed / $2.00 pay-as-you-go per automated resolution (resolution verified by a separate LLM check). Those are seller prices, not your fully loaded cost. Outcome-based pricing only tilts for the buyer once you know your own number.

CloudZero’s worked illustration (model spend only, before platform and engineering): about 12 model calls and 36,000 input / 6,000 output tokens per attempt on Claude Sonnet 5 at the prices they cite ($2 / $10 per million tokens) → roughly $0.132 per attempt; at a 70% full-resolution rate → about $0.19 per resolved ticket. They contrast that with $6–$8 of human handling time as the value side of the ratio — treat both as their worked example, not a universal support-ops table. Fill your own worksheet with your tokens, resolution rate, and loaded handle time.

Architecture multiplies the meter. Anthropic’s engineering write-up on their multi-agent research system (June 13, 2025) reports that agents typically use about 4× more tokens than chat, and multi-agent systems about 15× more than chats — worthwhile when the task value can pay for the lift, not as a default for every ops queue. CloudZero’s point follows: a 15× cost architecture needs outcomes worth roughly 15× more, and cost per outcome is the check.

Attribution before optimization

BCG’s operating model is blunt: see, shape, prove (or stop). Token costs do not belong in one opaque IT line. They distinguish three accounting homes: tokens that build reusable capability (investment-like), tokens that run internal work (opex, budgeted per function and workflow), and tokens inside a product or customer interaction (COGS, visible in gross margin). AI inference breaks the classic SaaS “next customer is nearly free” curve — each interaction carries model cost.

Their “see” dashboard questions are the ones ops and finance can share:

  • Which workflows consume the most tokens, and what outcomes do they produce?
  • What is the cost per successful outcome?
  • What share of spending has an assigned owner?
  • What is internal opex versus product COGS?
  • What percentage of input tokens is cached?
  • How much work defaults to frontier models?

Spending without an owner is ungoverned. Spending without a defined outcome is unproven. Spending without an assigned P&L line invites mismanagement.

TechTarget maps a fuller agentic cost stack most pilots undercount: model use (including reasoning tokens and long-context premiums), orchestration and retry loops, data and retrieval, tool and API calls (agents become transaction multipliers on CRM/ERP), infrastructure, governance and assurance (human review, audit, incident response), and the operating model itself (FinOps, evals, process redesign). If everything lands in “AI platform” or “innovation,” trade-offs disappear.

The FinOps Foundation’s FOCUS billing specification is the industry attempt to make that stack legible across vendors. FOCUS 1.4 is the published line as of mid-2026; the 1.5 release scope (status as of September 17, 2026; ratification targeted December 3, 2026) is where AI metering gets sharper: model-identity properties on SKU price details, principal/actor columns so a charge can name a person, service, or agent, and worked examples for token and generation billing. Practitioners do not need to wait for every column to ship — but if your AI bill still cannot say which workflow or agent burned the tokens, you are optimizing anecdotes.

Define the outcome before you measure it

CloudZero’s checklist: outcome name and owner; closing condition and time window; how human-assisted results count; explicit exclusions (abandoned, spam, duplicates, test); review date. Zendesk’s billable unit is a request “resolved by the AI agent, without any escalation to a human agent,” verified separately. Change the definition and the metric moves — version it like code, then optimize cost.

Shape the denominator: route, cache, stop, simplify

Once attribution exists, BCG and TechTarget converge on the same levers — none of which require a new model vendor:

  1. Eliminate unnecessary model use. Structured lookups, rules, arithmetic, and deterministic policy checks belong in software, APIs, or reusable skills — not in a planner that rediscovers a known path.
  2. Route by task complexity. Lighter or open-weight models for low-complexity steps; frontier models for hard reasoning. Include reasoning-effort and context-window defaults in the dashboard — overprovisioning is often a setting, not a requirement.
  3. Reuse context. Prompt caching, batch/async for non-latency-sensitive work, stable policy prefixes, and curated knowledge packs cut repeated-context waste. FOCUS 1.5’s cached-versus-fresh token work exists because cache hit rate is one of the largest optimization levers once it is visible in the bill.
  4. Cap loops. Step limits, tool budgets, and escalation gates are financial controls as much as safety controls. Runaway agents are a FinOps incident, not only an SRE one.
  5. Train for token discipline. Smallest useful answer, constrained output, retrieval that does not paste the wiki. Interfaces should default to concise outputs and clear stopping rules.

BCG’s governance punchline for standing reviews: assess the business numerator against the full denominator — tokens plus human time to initiate, review, correct, approve, and operate. An agentic coding workflow that dumps lines of code is activity; the numerator should count code that ships and survives review. A service agent that closes tickets but drives escalations is the same pattern. A marketing agent that generates assets nobody uses is activity, not return.

What to ship this week

You do not need a FinOps platform rewrite. Pick one production or near-production agentic workflow with a countable business result.

  1. Write the outcome definition. Closing event, human-assist rule, exclusions, owner. Freeze it for the measurement window.
  2. Join spend to that outcome. Tag or project-separate API usage by workflow; propagate a business event_id across retries and tool calls; include failed attempts in the numerator.
  3. Baseline cost per successful outcome. Model spend first if that is all you have; then add retrieval, tools, and review time so the number does not lie to finance.
  4. Apply one shape lever. Route routine steps off the frontier model, enable caching on stable prefixes, or add a hard step/retry budget — then re-measure the same outcome definition.
  5. Run a prove-or-stop review. Scale, optimize, rescope, or kill. “The agent is busy” is not a decision.

If you cannot say what one accepted result cost last month, you are not ready to scale the agent — you are ready to instrument it.

Want help turning one agentic workflow into attributed spend, a written outcome definition, and a cost-per-outcome baseline finance will stand behind? Talk to us.

Sources