Getty Images

Tip

Cost per task: A new method for AI spend management

Rather than relying on traditional metrics such as per-seat or per-token pricing, some businesses are measuring AI cost per successful task to manage their AI spending.

Donald Farmer
By
Published: 16 Oct 2026

The price of AI computation has fallen faster than the price of almost any other earlier technology, according to 2025 data from Epoch AI. Nevertheless, AI spending in the enterprise is steadily increasing in what Gartner recently called the inference paradox: falling costs per token are completely offset by the dramatically increasing use of tokens for complex, multistep agentic workflows.

This might not be as paradoxical as Gartner claims, though, because AI spend is not a straightforward economic question. As AI technology improves, its practical use cases also expand dramatically. The problem for the C-suite is that budgeting methods built for traditional enterprise software, even for SaaS subscriptions, do not work for AI -- which is metered by the token. Thus, businesses need to find a new method for managing AI spend by measuring AI cost per task.

Why the old budget model fails

Enterprise software has conventionally been priced per seat, so costs are linear, enabling finance teams to forecast with confidence. Even when priced by usage, traditional applications such as ERP or CRM have steady consumption patterns that are useful for budgeting.

With AI, cost depends on how much work the model does. Because work can vary from request to request, seat-based AI budgeting isn't possible. The evolution from AI chatbots to agents has made this problem even trickier. An agent turns what would be one prompt for a chatbot into a sequence of calls: it plans, retrieves documents, calls tools, drafts, checks the result and revises. In its aforementioned August 2026 press release, Gartner estimated that an agentic workflow costs at least five times as much as a basic chatbot interaction.

What makes agentic AI cost difficult to manage is that token pricing is not straightforward. Rate cards give the price per unit but not the number of units a piece of work requires. But AI providers also price input and output tokens differently, and reasoning models generate internal tokens that users are unaware of. So, too often, the monthly invoice provides a total for tokens without business context.

FinOps dashboards designed to manage cloud spend are of limited help because they track provisioned infrastructure -- but an agent is not provisioned so much as invoked. As I saw recently in a medical device company, the dashboard might show that engineering spent $10,000 on an API, but it can't say that $8,000 of that was one agent reading the same employee quality management handbook 4,000 times.

What cost per task means

Tasks are a fundamental unit of work: well-defined, verifiable and with acceptance criteria. In contrast, tokens are billing units. Therefore, it makes sense to consider the cost of AI per task: the full cost of completing one such unit of work successfully, averaged across all attempts, including the failures.

Include human review as part of measuring cost per task -- rather than assuming it is a payroll cost -- because without it, you cannot see what role people play in the total. Similarly, the cost of retrying failed tasks is easy to miss, yet, if you include it, you will reverse several "cheapest model" decisions.

Suppose a small model costs $0.20 a run and produces an acceptable extraction 15% of the time, while a larger model costs $0.80 a run and succeeds 78% of the time. The small model costs $1.33 per usable result, and the large one about $1.03, before counting the rework the small model creates. The cheapest token might produce the most expensive task.

How to measure cost per task

AI providers bill based on raw token consumption, not on business outcomes. To measure actual cost per task and control spending, implement a four-step framework:

  1. Tag traffic at the source. Automatically tag every AI request with its team, product feature and task at the exact moment it is sent. Once an API call is completed, you are too late to capture that metadata, and the business context is lost.
  2. Standardize multivendor data. Convert different usage metrics from different providers and self-hosted models into a unified format. Do this in your central data lakehouse.
  3. Benchmark medians vs. outliers. Measure standard task costs -- median costs -- against worst-case costs, such as your most expensive 5% to 10% of tasks. A large gap can reveal hidden failures, such as agents getting trapped in expensive, repeated loops.
  4. Report by high-impact activity. Rank workflows by total spend and closely monitor cost per user.

Tailored reporting views are also essential for getting the most out of cost-per-task measurement. Different leadership roles need their own views of the same underlying data. For example, business owners can compare cost per task against real business value, such as human hours saved or support tickets resolved. Finance teams can track cost per task in the lens of budget variances and cost-center trends. Engineering teams can monitor worst-case cost spikes, cache efficiency and agent loops to optimize performance.

Lastly, always bear in mind that every cost-cutting measure involves a tradeoff. So, evaluate your optimizations using one core metric: cost per successful task. If a change reduces API costs by 30% but lowers task accuracy, it might force human employees to step in and fix errors. In that case, you haven't saved money; you've just shifted the expense onto payroll.

Who owns the bill?

Unlike conventional enterprise applications, AI spend does not fit neatly into departmental lines. Finance might own the budget, and engineering might own the prompts and model choices, but multiple lines of business own the outcomes and their use. If all stakeholders share responsibility, no one really owns it, and the usual result is often a surprise invoice followed by a heated interdepartmental wrangle.

No AI workload should run in production without a named owner with profit-and-loss responsibility for it. This is usually a product manager or a line-of-business lead, not the IT department, because the owner must understand the value the task creates, which is a business value, not an IT calculation.

Reducing cost per task without buying the cheapest model

Cutting AI costs isn't merely about switching to cheaper, lower-quality models. It's about eliminating structural waste. Businesses can implement cost per task as a method of tracking AI spend by using five core techniques:

  1. Smart routing or inference tiering. Route routine tasks, such as categorization, to fast and low-cost models. Automatically escalate to expensive frontier models when task complexity or low confidence scores require it.
  2. Context caching. Reuse common system instructions and reference documents across tasks. This enables providers to cache static data for reuse.
  3. Context trimming. Stop sending full conversation histories or extensive documentation for every interaction. Filtering out irrelevant text can slash token count and improve accuracy.
  4. Workflow and batch optimization. Streamline your agent setup -- for example, let agents generate code instead of loading tool definitions -- and run non-urgent background tasks through low-cost batch processing.
  5. Guardrails and circuit breakers. Set strict caps on task iterations and per-task spend. Circuit breakers stop failing agents from looping endlessly and burning through budget overnight.

Donald Farmer is a data strategist with 30+ years of experience, including as a product team leader at Microsoft and Qlik. He advises global clients on data, analytics, AI and innovation strategy, with expertise spanning from tech giants to startups. He lives in an experimental woodland home near Seattle.

Next Steps

Dig Deeper on Enterprise AI Strategy