Cost Per Workflow: Why Token Pricing Can't Prove Your AI Is Profitable

A simple AI task cost about four cents in 2023. Run the same category of task through an agentic system today — tools, reasoning steps, a few retries — and it costs around $1.20. Thirty times more, in three years, while every model provider’s price-per-token chart has been pointing the other way.

That’s not a typo, and it’s not mismanagement either. It’s arithmetic. Total spend is price per token times volume, and volume — the number of model calls a single task now quietly triggers — is growing far faster than price is falling. Seventy-three percent of enterprises say their AI costs blew past what they’d projected. Almost none of them mismanaged a token price. They just never measured the thing that actually moved.

AI unit economics requires pricing a complete workflow — inference, retrieval, orchestration, and every tool call in between — against the outcome it produced, not just the price of a token. Token pricing gives you a falling unit cost per million tokens; it doesn’t explain why a single agentic task can trigger ten to twenty model calls before producing one finished result. Without workflow-level attribution, you can watch the token price fall while the real cost per completed task rises thirty-fold, and never know why.

The metric that broke

Cost per token tells you what a model charges. It has never once told you what your business got back for that charge, and that gap has gotten expensive enough to show up on an earnings call.

MIT research, cited in CloudZero’s 2026 AI cost analysis, found that 95% of enterprise AI pilots show no measurable P&L impact — not because the models are bad, but because nobody can trace the spend to a result. Inference now eats 85% of the average enterprise AI budget, per the FinOps Foundation’s 2026 State of FinOps survey, and a single agentic task can trigger ten to twenty model calls before it produces one finished unit of work. A dashboard that reports blended price per million tokens is reporting on an ingredient. It was never built to report on a dish.

Even the FinOps Foundation has moved on from token-level reporting as the finish line — its 2026 survey found 98% of FinOps teams now manage AI spend, up from 31% two years ago, and the certification track has expanded specifically because “how many tokens did we use” stopped being a useful question the moment AI workloads stopped behaving like simple API calls.

What we mean by AI unit economics

It’s the discipline of pricing AI spend against a completed unit of work — a resolved ticket, a processed claim, a merged pull request — instead of against tokens, API calls, or a lump monthly total.

Where token tracking prices the model’s output, unit economics prices the whole chain that got you there: inference, retrieval, orchestration, every tool call in between, all divided by a denominator finance already reports on. That’s the difference between AI-cost monitoring and something a CFO can actually put in a board deck. If you can’t state a defensible cost and margin per unit of AI-delivered work, you can’t tell a profitable initiative from a liability you haven’t noticed yet.

An example, because the abstract version never convinces anyone

One of our customers runs an AI coding agent that opens and merges pull requests. The API bill for a month: $4,200. On its own, that number tells finance nothing — not whether it was worth it, not whether it’s improving, nothing.

Divide it by the pull requests the agent actually helped merge, and something more honest shows up:

What a Merged Pull Request Actually Costs
──────────────────────────────────────────────────────────────
Cost component                     Monthly value
────────────────────               ───────────────────────────
Model inference (input + output)    $2,480
Retrieval and context calls         $890
Failed / abandoned attempts         $560
Orchestration and tool calls        $270
──────────────────────────────────────────────────────────────
Total workflow spend                 $4,200
Pull requests merged                 80
Cost per merged PR                   $52.55
──────────────────────────────────────────────────────────────

$52.55 a PR is a number you can hold up against an engineer’s fully-loaded hourly cost and actually argue about. $4,200 is just a bill. Correlation One’s 2026 enterprise AI cost playbook makes the point sharper than we could: a workflow whose token bill doubled while it handled triple the volume actually got cheaper — a fact that’s invisible to anyone who isn’t dividing by something.

The part that doesn’t show up on the model’s invoice

The model bill is rarely the whole story. In retrieval-heavy and agentic setups, the surrounding harness — vector databases, embeddings, rerankers, the orchestration runtime, caching, egress — routinely makes up a real chunk of total feature spend, and none of it appears anywhere near the model provider’s invoice.

Google Cloud’s Pravir Gupta put it well at FinOps X 2026, as reported by SiliconANGLE: it’s the part of the iceberg under the water. An agent spins up a sandbox, writes a script, calls three other tools, and every one of those steps has a cost and no natural owner. Cockroach Labs documented one 2026 incident where a misconfigured agent approached half a billion dollars before anyone caught it. If your attribution stops at the token count, you’ve just built a more sophisticated version of the same blind spot.

This is the layer our Meter module was built to close — attribution at the token, agent-run, and workflow level, so harness spend lands on the same ledger as inference instead of hiding in a gap between two invoices. Because that ledger is FOCUS 1.4 native rather than bolted on afterward, it sits alongside your cloud and Kubernetes cost data instead of becoming one more spreadsheet someone reconciles by hand. For regulated customers, the same attribution runs through our BYOC Secure Relay Proxy, so none of that billing data leaves your own perimeter — which is exactly the deployment one of our banking customers needed to govern AI at institutional scale.

Doing this yourself, without our product

You don’t need our platform to start. Four steps:

1. Count the whole chain. Tasks per day, times steps per task, times tokens per step — not just the final call that produced the answer.
2. Add the harness on top. Retrieval, orchestration, and tool calls usually add more than people expect. Treat it as a real cost line, not a rounding error.
3. Pick a denominator the business already uses. A merged PR, a resolved ticket, a processed claim. Never a token count — nobody on your board thinks in tokens.
4. Check it quarterly. Model routing and pricing tiers move faster than an annual budget cycle. A number you calculated once in January is a guess by June.

What this actually means for the board

Nobody on your board thinks in tokens, and handing them a blended per-million-token rate to approve is handing them a description of an ingredient and asking them to price the meal.

Cost per workflow is the unit of analysis they already use everywhere else — cost per unit, weighed against the value that unit produced. The more mature organisations we work with report this quarterly, next to cost per loan or cost per ticket, and reconcile AI spend against the same P&L lines finance already tracks.

The 73%-over-budget statistic, and Gartner’s forecast that 40% of agent projects get cancelled on cost grounds by 2027 — those aren’t stories about bad technology. They’re what happens when nobody can answer a simple question: what did it cost us to produce one unit of the thing we built this for?

Your token price is falling. Whether that matters depends entirely on a number most dashboards don’t show you.

See your own cost per workflow. Our Meter module attributes AI spend down to the workflow level — harness costs included — inside the same FOCUS-native ledger as your cloud and Kubernetes estate. Book a 30-minute walkthroughand bring one AI workflow you genuinely can’t price yet. We’d rather work through a real one than a hypothetical.

Frequently asked questions

What is cost per workflow? Cost per workflow is the practice of dividing the full cost of an AI task — inference, retrieval, orchestration, every tool call in between — by the completed unit of work it produced, instead of reporting a blended price per token. Where token pricing measures what a model charges, cost per workflow measures what a business actually got back for that charge. It’s the metric that lets a CFO tell a profitable AI initiative from one that’s quietly losing money, because it prices the whole chain rather than one ingredient in it.

Why is my AI bill rising even though token prices are falling? Because total spend is price per token multiplied by volume, and volume is growing faster than price is falling. A single agentic task can trigger ten to twenty model calls — reasoning steps, tool calls, retries — before producing one finished result, and each of those calls consumes tokens even though the price per token keeps dropping. The falling unit price and the rising total bill are both true at once; they’re just measuring different things.

How much of enterprise AI spend is actually the model bill? Inference is typically the visible line item, but in retrieval-heavy and agentic setups the surrounding harness — vector databases, embeddings, rerankers, orchestration, caching, egress — routinely adds a substantial share of total spend on top of it, and none of that appears on the model provider’s invoice. If your AI cost tracking stops at the token count, you’re likely missing a meaningful chunk of the real bill.

What should enterprises measure instead of tokens? A denominator the business already reports on: a merged pull request, a resolved support ticket, a processed insurance claim. Never a raw token count, since no board or CFO thinks in tokens. The specific unit doesn’t matter as much as picking one and dividing consistently — that’s what turns AI cost into a metric finance can actually use in a board deck.

Why do 73% of enterprises exceed their AI cost projections? Because most forecasting still assumes AI workloads behave like the chatbot-era request-response pattern — one prompt, one predictable cost. Agentic workloads don’t: they plan, call tools, spawn sub-agents, and iterate without a human in the loop, which is why token consumption on an agentic task can run five to thirty times an equivalent chatbot interaction. A budget built on the old pattern is structurally likely to miss.

How does DigiUsher attribute AI cost at the workflow level? DigiUsher’s Meter module attributes spend at the token, agent-run, and workflow level, joining inference cost to the harness spend around it — retrieval, orchestration, tool calls — inside a single FOCUS 1.4 native ledger shared with cloud and Kubernetes cost data. For regulated industries, that same attribution runs through a BYOC Secure Relay Proxy, so billing data never leaves the customer’s own perimeter.

Related reading

• $52.55 Per Merged Pull Request: How One 1,000-Engineer Firm Priced Its Own AI Spend — July 21, 2026
• AI Unit Economics: Why Your AI Invoice Can No Longer Answer the Board’s Question — July 16, 2026
• Token Economics Gave AI Spend a Vocabulary. It Still Has No Owner. — July 9, 2026
• AI Cloud Margins: The New Battlefield for Enterprise Profitability — March 12, 2026

© 2026 DigiUsher. All rights reserved. Privacy Policy · Terms of Service · Trust Center