Cost per completed task: the AI unit economics that actually matter

TL;DR

Token price answers "what does a thousand tokens cost." It doesn't answer the question a CFO has: what did it cost to finish the job. McKinsey's own analysis of agentic workflows found that in a customer service example, token costs were only 20 to 25% of the variable cost of running an agent. Human oversight was 70 to 75%. In a bank account onboarding workflow that needed five to seven agents and several review teams, the fully loaded cost per completed account fell from roughly 50 to 150 dollars down to 10 to 30, a number no per-token price tag would have shown on its own. Agentic workflows also routinely use more tokens than a single chat reply, and that's fine: the token line is rarely the majority of the bill once oversight and infrastructure are counted, and it's the total divided by tasks finished that tells you whether the workflow paid off.

A small unit labeled as a single token next to a larger unit representing one completed task, made up of several smaller parts

What's wrong with measuring AI cost by the token?

Nothing, as far as it goes. A per-token price tells you what a thousand tokens cost this month, and that number is real and worth tracking. The problem is that it isn't the number a finance team needs, because a token is an input to a task, not the task itself. An agent that reads a record, calls three tools, checks its own work, and finishes correctly, and an agent that retries twice before getting there, can produce wildly different token counts for the exact same finished result. Comparing them on tokens alone rewards whichever one happened to use fewer, even if it's the one that failed and had to be redone by a person.

What did McKinsey find when it added up the full cost of a completed task?

McKinsey's QuantumBlack group published an August 2026 analysis of agentic workflow economics built from real deployments. The finding: completed-work cost, not token cost, is the metric that matters. In one customer service example they studied, token costs represented only 20 to 25% of the variable cost of running an agent. Human oversight, the review step where a person checks or approves the agent's work, accounted for 70 to 75%. In a more complex case, opening a bank account, the completed workflow required five to seven agents, multiple back-end systems, and two to four teams of people providing oversight along the way. Priced end to end, the fully loaded cost per completed onboarding fell from an estimated 50 to 150 dollars down to 10 to 30. Neither number shows up if the only thing being measured is the price per token.

Why can a model with a lower per-token price cost more per completed task?

Because the price of one attempt and the price of one finished result aren't the same calculation once failures are in the picture. DoiT's Cost Per Task framework makes this explicit with a simple relationship: cost per solved task equals the cost of one attempt divided by the rate at which attempts succeed. A model that's 20% cheaper per attempt but fails more often can end up more expensive per completed task than a pricier model that gets it right the first time, because the cheaper model's failures still have to be caught, and someone still has to redo the work. Reliability is a cost line, even when nothing on the invoice calls it that.

Doesn't an agent that makes twenty tool calls cost more than a single chat reply?

Often, yes. An agentic loop that reads a record, checks a policy, writes to two systems, and drafts a confirmation makes far more model calls than one chat completion, and the token bill for that loop can be several times larger than a single answer in a chat window. The reason this is still the better economic choice is what McKinsey's own breakdown shows: the token line is frequently the smaller part of the total cost, not the larger one. A workflow that finishes end to end, without a person manually carrying the output between systems, removes an oversight cost that a chat reply never had to pay in the first place, because a chat reply never finished anything a person didn't still have to act on.

Where does the human-oversight cost go, and why is it usually the bigger number?

It goes to whoever has to check the agent's work before it's trusted to stand: a functional or risk expert confirming a customer service resolution, a reviewer confirming a bank onboarding decision, an ops lead checking a workflow ran correctly. McKinsey's own figures put this at 70 to 75% of variable cost in their customer service example, and it scales with how deterministic the workflow is and how much compliance exposure it carries. This is also the cost line most within a company's control: a well-scoped workflow, with clear boundaries on what the agent is allowed to touch and a defined path for the exceptions that need a person, needs less oversight per run than a broad, loosely scoped one. Reducing exception rates through better workflow design, McKinsey notes, does more for total cost than optimizing which model is cheapest per token.

What does this mean for choosing which workflows to automate first?

Pick for volume and reuse before picking for token efficiency. McKinsey's analysis found that agent economics favor workflows with high run volume, since fixed costs like infrastructure and orchestration get amortized across more completed tasks the more often a workflow runs, and workflows that get built once and reused across several use cases compound that advantage further. A workflow that runs twice a year is a poor candidate no matter how cheap the model is per token. A workflow that runs a thousand times a month is a good candidate even on a pricier model, because the fixed cost per completed task keeps falling as volume grows.

How should this change what a chief financial officer asks about an AI project?

Not "what model are we using and what does it cost per token," which answers a question about a line item rather than a business outcome. Ask what the fully loaded cost per completed task is today, tokens, tool calls, and human review time combined, divided by the number of tasks that finished, and what that number would need to be to justify the workflow at the volume it runs. This number can be tracked quarter over quarter the way any other unit cost gets tracked, and it's the number that tells you whether spending more on AI is working, not just whether more people are using it.

Where does scoping which systems an agent can reach fit into this?

Directly, through the same oversight-cost line above. An agent connected to exactly the systems and skills its workflow needs, with access provisioned by role rather than left broad by default, produces fewer wrong or out-of-scope actions to catch, which is the same exception rate McKinsey ties to lower oversight cost. Access control and skills scoped to a specific, repeatable workflow belong inside the cost conversation, not outside it. They keep the largest cost line in McKinsey's own breakdown, human review, from growing in proportion to how many workflows a company hands to agents.

What's the one number worth asking for before approving the next AI budget line?

The fully loaded cost of one completed task on the workflow in question, run at the volume it will run, compared honestly against what finishing that task cost before. Not the price per token, which measures an input. Not the number of people using the tool, which measures adoption. The cost of one finished result, because it's the only number that answers whether the spending paid off.

FAQ

Is "cost per completed task" a term McKinsey invented?

No. McKinsey's QuantumBlack group applies this framing in its own August 2026 analysis of agentic workflow economics, and a separate research framework from DoiT formalizes a similar measurement approach for comparing models. Both make the same underlying point independently: token price is not the unit that determines whether an AI workflow paid off.

Does this mean cheaper models are always the wrong choice?

No. It means per-token price alone doesn't answer the question. A cheaper model with a high success rate on a given workflow can still be the right choice. The point is that success rate and oversight cost have to be part of that comparison, not just the price per token.

Is it true that agentic workflows use more tokens than a single chat reply?

Often, yes. Token cost tends to be a smaller share of the total bill than human oversight, per McKinsey's own breakdown of real deployments, which is why this is still usually the better economic choice.

What's the biggest lever for lowering cost per completed task?

Per McKinsey's analysis, reducing exception rates through better workflow design and scoped access, which lowers the human-oversight cost, tends to matter more than switching to a cheaper model.

Does this apply to small-volume workflows too?

Less so. McKinsey's own findings tie the strongest economics to high-volume, reusable workflows, where fixed costs amortize across many completed tasks. A workflow that runs rarely is a weaker candidate regardless of per-task cost.