Ask a large company how much it spends on AI, and you’ll usually get an answer within minutes. Model contracts, cloud commitments, consultants and licences are all easy to find.

Ask what one AI agent costs to resolve a customer support case, approve an invoice or complete another business task, and the room goes quiet.

Most organisations know the total platform spend. What they don’t know is the cost of a successful business outcome. Without that number, there’s nothing meaningful to compare against the process the AI agent supports, augments or automates.

The reason is fairly simple: AI platforms are built to prove capability and win adoption. Measuring unit economics comes later, when the first invoice arrives, and finance starts asking questions.

And those questions matter to platform owners, engineering managers and architects because they are the people expected to explain why one use case costs more than another and whether the investment is worth scaling.

Measuring may feel optional today, but finance departments are patient, not blind.

Why AI agent cost behaves differently from cloud cost

Traditional cloud FinOps already does a good job of attributing infrastructure costs to workloads. Serverless functions, warehouse scans and managed APIs all bill by usage, so this isn’t another story about capacity versus consumption.

AI agents add another layer.

The compute resources consumed by the same workload can vary enormously depending on what a user asks and how the agent decides to solve it. One long conversation with deep retrieval and several tool calls can cost more than a thousand short questions.

Some of that variation comes from design choices. Developers set the ceiling through model choice, context management, iteration limits, retrieval depth and retry policy. Users decide where each request lands within that ceiling.

Model inference is only part of the overall cost. An agent also uses warehouse queries, vector searches, document storage, network traffic, API gateways, runtime infrastructure and telemetry. This creates two distinct cost figures, each answering a different question.

• Marginal cost covers what one more request adds, including model calls, tool calls and retrieval. It tells you whether to serve the next request
• Fully loaded cost adds the runtime, vector database, observability and a share of the shared platform. It tells you whether the use case should exist

How to measure AI agent cost per business outcome

The goal is not just to measure what an interaction costs; it’s to measure what a successful outcome costs.

An agent that resolves a case for 20 cents is not cheaper if many of those cases need to be reopened. And a 15-cent interaction is not really 15 cents if someone still needs to spend ten minutes reviewing it.

To measure the cost of a successful outcome, first define what success looks like. That might be a support case resolved without reopening, a report accepted without revision, a document that passes validation or a transaction completed without human intervention.

Once you’ve defined success, these five steps help get you there:

1. Route every model call through a gateway
   Require each request to carry team, application, use case, agent version and a correlation ID, and have the gateway reject requests without them.
   Credentials identify the caller. Metadata carries the cost dimensions. Do not create a new credential every time finance wants another dimension.
2. Emit a trace for every interaction
   Record the model, input, output and cached token counts, and that metadata.
   This does not mean storing prompts and responses. Content capture is a separate decision with privacy and regulatory consequences, and cost attribution does not need it.
3. Price requests at the gateway
   Use a pricing table that you control and write the calculated cost directly onto each trace.
   That way, when a provider changes its pricing, your historical reporting stays consistent.
4. Join traces to business outcomes
   Export traces to your data warehouse and join them, on the correlation ID, to the table that already holds the result.
   That join is the whole exercise. Once connected, you can answer questions like “How much does it cost to resolve a support case?” or “How much does it cost to approve an invoice?”
   Allocate shared platform costs on top if you want the fully loaded figure.
5. Set budgets and limits per use case
   Alerts should reach the team that owns the spending, not a shared mailbox. Attribution without enforcement produces accurate reports on money already spent.

Where cost estimates go wrong

Almost every agent cost estimate I have reviewed contains the same error: it assumes one model call per user request.

That’s rarely how agents work.

An agent request is a loop: plan, call a tool, read the result, plan again. Each turn resends everything that came before.

Gartner puts agentic workloads at five to thirty times the tokens of an equivalent chatbot exchange [1].

A spreadsheet based on a single model call isn’t slightly optimistic. It’s wrong by an order of magnitude on the largest line of the bill.

Uber makes the point publicly. Agentic coding adoption there rose from 32% of engineers in February 2026 to 84% in March. Average monthly spend ran from $150 to $250 per engineer, with heavy users between $500 and $2,000, and the annual AI budget was exhausted by April [2].

What happened next matters more: rather than restricting usage, Uber added prompt caching, better model defaults and usage dashboards, and per-token cost fell while usage grew [3].

Several smaller errors usually travel with the first:

• Reasoning tokens are billed at the output rate on every major provider, so the visible answer is a poor guide to what you pay [4-6]
• Failed calls, retries and rejected responses cost full price and rarely appear in any estimate
• Prompt caching is also left out, although some providers can serve a repeated system prompt with tool definitions at a tenth of the normal input price [7]
• Retrieval deserves its own line too, since every query adds roughly the chunk size multiplied by the number of chunks returned, before any additional framing

Common mistakes

A few mistakes come up repeatedly:

• Reading provider invoices at the end of the month and dividing them across applications gives you a total, and you already have that total
• Treating model choice as settled. Prices fall, and new models appear every few months, so a decision made during the build decays quietly. Send a share of traffic to a candidate model, score it on quality and cost, and let the evidence choose
• Ignoring scheduled runs. The moment users can schedule an agent, someone will set it to repeat every minute against a large knowledge base and then leave for the weekend. A minimum recurrence interval takes an afternoon to build
• Leaving it until later. Adding attribution to a platform that already carries a dozen live use cases is far more work than building it from the start, and by then the arguments about who pays for what are already underway

Summary

An AI platform you cannot measure by successful outcomes is a platform you cannot defend, expand, or shut down with confidence. Decide what one successful interaction can cost, instrument for that number before the first agent goes live, and keep the difficult conversations short.

Pavlos Polydoras is a Lead Data & AI Architect at Holisticon Insight, where he designs and builds enterprise AI platforms and agentic systems for large organisations.

References

[1] Gartner, “Gartner Predicts That by 2030, Performing Inference on an LLM with 1 Trillion Parameters Will Cost GenAI Providers Over 90% Less Than in 2025,” 25 March 2026. https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performing-inference-on-an-llm-with-1-trillion-parameters-will-cost-genai-providers-over-90-percent-less-than-in-2025

[2] Forbes, “Uber Burns Its 2026 AI Budget in Four Months on Claude Code,” 17 May 2026. https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/

[3] Fortune, “After blowing through its entire 2026 AI budget in months, Uber CTO says the tokenmaxxing era is over,” 7 August 2026. https://fortune.com/2026/08/07/uber-ai-spending-tokenmaxxing-is-over-cto/

[4] OpenAI, “Reasoning models: API documentation.” https://developers.openai.com/api/docs/guides/reasoning

[5] Anthropic, “Extended thinking: Claude Platform documentation.” https://platform.claude.com/docs/en/build-with-claude/extended-thinking

[6] Google, “Gemini API pricing.” https://ai.google.dev/gemini-api/docs/pricing

[7] Anthropic, “Claude pricing: prompt caching.” https://platform.claude.com/docs/en/about-claude/pricing