Cost per token tells you what a model charges. It does not tell you whether the work got done, how well, or at what real cost to the business. Cost per outcome, the total cost of a resolved ticket, a processed document, or a completed workflow, is the metric that actually shows whether an AI system is wasteful or valuable.
Walk into almost any company running generative AI in production today and ask how they measure success. Most will point to a token count. Tokens in, tokens out, dollars per million tokens, a dashboard tracking model spend by team. It feels rigorous. It is the number the AI providers hand you by default, and it is the line item finance sees on the invoice.
It is also, on its own, almost meaningless.
A token is a unit of text, not a unit of business value. Measuring cost per token tells you how efficiently a model chews through language. It says nothing about whether the output was correct, whether a human had to redo the work anyway, or whether the task the AI performed was worth doing in the first place. Worse, optimizing for cost per token quietly punishes the use cases that matter most. A complex, high-value workflow, one that requires an agent to reason across multiple steps, call tools, and check its own work, will always consume far more tokens than a simple one-shot question. If a team is judged on cost per token, the rational move is to avoid the ambitious, high-ROI work and stick to cheap, shallow prompts. That pushes teams toward exactly the wrong work.
This is the trap Sharma Vedula, Head of Solutions at Data Society, calls out directly:
“What companies need to measure instead is cost per outcome, not cost per token. Cost per the thing the business actually cares about. Cost per resolved ticket, cost per processed document, or even cost per completed workflow. Once you have that number, you can have a real conversation about where the system is wasteful and where it’s delivering value.”
That single reframe, from a technical unit to a business unit, changes almost everything about how a company should evaluate its AI investment. It is also, according to the data on enterprise AI adoption, the reframe most organizations still have not made.
The scale of the problem is bigger than most leadership teams realize. Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027, and the research points to the same handful of root causes every time: runaway costs from agents that run continuously and rack up compute bills no one budgeted for, governance gaps that let autonomous systems drift from their intended purpose, and, above all, unclear ROI. In NVIDIA’s own enterprise survey, “lack of clarity on AI’s ROI” ranks as the third most common barrier organizations report, cited by roughly 30% of respondents.
The real problem is how success gets measured. Companies are pouring money into pilots and production systems without a shared definition of what success costs and what success is worth. When the only number on the dashboard is token spend, a finance leader sees the bill going up and assumes the project is out of control, even when the system is quietly saving the business far more than it costs. Conversely, a cheap, low-token deployment can look efficient on paper while delivering almost nothing of value, because nobody is checking whether the resolved tickets were actually resolved correctly.
Cost per outcome fixes this by forcing the conversation back to the thing that was supposed to happen in the first place.
The idea is simple, but it only works if you get specific about the unit of value. “The thing the business actually cares about,” as Vedula puts it, is different for every workflow. A few concrete examples make the framework easier to apply.
Cost per resolved ticket. In customer support, industry benchmarks vary by source, but human-handled phone and chat contacts commonly run somewhere in the $13 to $25 range, while self-service or AI-assisted resolution can bring that down closer to $2 to $5 per ticket. But the comparison only holds if “resolved” means the same thing on both sides. If an AI system closes a ticket that then reopens because the answer was wrong, the true cost per outcome has to include that second contact. First-contact resolution rate, not deflection rate, is the number that keeps this metric honest.
Cost per processed document. For document-heavy workflows, contracts, claims, compliance filings, the outcome is a document processed correctly enough that no human had to redo the work, not just a document that got summarized. That means the denominator in your cost-per-outcome calculation has to account for the review, correction, and escalation steps that still happen downstream, not just the AI’s first pass.
Cost per completed workflow. This is the broadest and often the most revealing unit. A multi-step agentic process, say, an agent that reads an inbound request, checks it against a policy, drafts a response, and routes it for approval, might use ten or a hundred times more tokens than a single chatbot reply. Judged by cost per token, it looks wildly expensive. Judged by cost per completed workflow, compared against what it used to cost to run that process manually, it may be dramatically cheaper. Anthropic’s own engineering research backs this pattern up: a single agent typically uses roughly 4 times more tokens than a simple chat interaction, and a multi-agent workflow can use roughly 15 times more, which is precisely why token-based scorekeeping misleads leadership about what’s actually happening.
Shifting from cost per token to cost per outcome takes more instrumentation than reading a usage dashboard, but the framework is straightforward.
First, define the outcome unit before you build anything. Decide, in writing, what counts as a resolved ticket, a correctly processed document, or a completed workflow, including the quality bar. An outcome that has to be redone by a human later doesn’t count as an outcome. This step is where most measurement efforts quietly fail, because it requires the AI team, the business owner, and often finance to agree on a definition none of them fully owns alone.
Second, total the full cost of delivering that outcome, not just the model bill. That includes token spend, yes, but also the infrastructure running the pipeline, the tooling and orchestration layer, any human review or escalation time, and the engineering hours spent maintaining the system. A workflow that looks cheap on the model invoice can be expensive once you add the analyst who has to check every third output.
Third, establish the baseline you’re replacing. Cost per outcome only means something in comparison. What did it cost, in dollars and time, to resolve that ticket, process that document, or complete that workflow before the AI system existed? Without that baseline, a number like “$2.40 per resolved ticket” floats in space. Against a $14 baseline, it’s a clear win. Against a $1.50 baseline, it’s a problem to fix.
Fourth, track the metric over time and by segment, not just in aggregate. The same workflow can have wildly different cost-per-outcome numbers depending on complexity, and averaging them together hides where the system is actually wasteful. A support system that costs $1 per outcome on simple password resets and $40 per outcome on complex billing disputes needs different fixes for each, and a blended average of $6 tells you neither.
Once that number exists, the conversation Vedula describes becomes possible: where is the system wasteful, and where is it delivering value. Teams can start asking sharper questions. Is a high cost-per-outcome workflow expensive because the task is genuinely hard, or because the prompt, the model choice, or the agent’s tool use is inefficient? Is a low cost-per-outcome number real, or is it hiding rework that hasn’t been counted yet?
Measuring cost per outcome is not primarily a finance exercise or a procurement checkbox, though that’s the piece most conversations about AI ROI skip. It requires people across the organization who can define outcome units clearly, instrument a pipeline to capture the full cost picture, read the resulting data critically, and translate what they find into decisions about model choice, workflow design, and where human review still belongs.
That is a data literacy and AI fluency problem as much as it is a metrics problem. Teams that only know how to prompt a model can’t tell you why a workflow’s cost per outcome tripled last quarter. Teams that only know finance can’t tell you whether a spike in token usage reflects a harder task or a broken agent loop. The organizations that get this right are the ones investing in both: people who understand how these systems actually work under the hood, and people who know how to connect that understanding to the metrics leadership is asking about.
The real conversation about AI spend starts with the right number, not a better dashboard for the wrong one. Data Society’s AI Advisory work helps organizations build a real cost-per-outcome model, and AI upskilling programs build the cross-functional fluency to act on what it shows.
Frequently Asked Questions
Cost per outcome is the total cost of delivering a result the business actually cares about, such as a resolved ticket, a processed document, or a completed workflow. Unlike cost per token, which only shows how efficiently a model processes text, it reveals whether the work got done correctly and at what real cost.
Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, inadequate risk controls, and unclear business value. In NVIDIA’s enterprise survey, about 30% of respondents named lack of clarity on AI’s ROI as a top barrier, making it the third most common challenge.
Anthropic’s engineering research found that a single agent typically uses about 4 times more tokens than a chat interaction, and a multi-agent system uses about 15 times more. Judged by token spend, those workflows look wasteful and push teams toward cheap, shallow prompts, even when the completed workflow may cost far less than the manual process it replaces.
Start by defining the outcome unit and its quality bar in writing, then total the full cost of delivering it, including tokens, infrastructure, tooling, human review, and engineering time. Compare that figure against what the work cost before AI, and track it by segment over time, because a blended average hides where the system is wasteful.
Teams need people who can define outcome units, instrument pipelines to capture the full cost picture, and read the resulting data critically. That makes it a data literacy and AI fluency problem as much as a metrics problem, and the organizations getting it right pair technical depth with business fluency.
