The Real Cost of Running an Enterprise AI Agent: Inference Is 5%

One production agent handling 20,000 tasks a month costs about $395,000 in its first year. That figure comes from a model built on published list prices and official wage statistics, shown line by line below. Inference is $20,300 of it, or 5%. Enterprise AI cost is dominated by review labour, evaluation and integration maintenance, and most budgets price none of the three.

What enterprise AI cost actually covers

Ask a finance team what an agent costs and you will usually get the API invoice. That invoice is real, checkable, and almost never the largest number.

An agent is not a model. It is a model plus a retrieval layer, a set of tool integrations, a policy for when a human sees the output, a test suite that proves it still works, and a person who gets paged when it does not. Every one of those has a price.

Gartner attributes its cancellation forecast to three causes, and the first is escalating costs. Anushree Verma, the Senior Director Analyst behind that prediction, put the mechanism plainly: early-stage experiments "blind organizations to the real cost and complexity of deploying AI agents at scale". The pilot is cheap because the pilot skips the expensive parts.

The rest of this post prices those parts. Every input is either a published list price, an official wage statistic, or an assumption I have labelled as one. You should replace my assumptions with yours. The structure is the useful part, not my volumes.

The inference line, priced from the published rate card

Start with the number everyone already counts, so the comparison is fair.

The published rate card

Anthropic publishes per-token prices for every model. Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output tokens, with cache hits billed at 10% of the input rate. Those are the numbers in this model.

Assume a completed task consumes 60,000 input tokens and 4,000 output tokens, with 70% of the input served from cache. That is a realistic shape for a triage agent that carries a long system prompt, a policy document and four tool results.

Uncached input costs $0.036. Cache reads cost $0.008. Output costs $0.040. Total inference per completed task is $0.084. At 240,000 tasks a year, that is $20,300.

For scale, the vendor's own worked example prices 10,000 simple support conversations at about $37 using its cheapest model. The agentic version costs 23 times more per task than the chat version. Agent architectures buy capability with tokens, which is why the falling price per token has not lowered anyone's bill.

Retries bill at the same rate

The $0.084 figure covers a task that succeeds. Failed runs, timeouts and abandoned tool loops bill identically, and they do not appear in any success metric.

Anthropic's own engineering team measured this effect on production traffic. Agents use about 4 times more tokens than chat interactions, and multi-agent systems use about 15 times more. Token usage alone explained 80% of the performance variance on their internal browsing benchmark.

Read that finding honestly. It means the token line is a lever you can pull, and pulling it down costs accuracy. The relationship between token spend and gross margin on AI features is not a rounding error for a vendor. For a buyer running one internal agent, it is.

Evaluation is a standing cost, not a project

An eval suite is the only thing standing between a model update and a silent regression in production. It is also the line most often funded once and never again.

Build it once, then maintain it forever

Price engineering time from official data rather than a guess. The US Bureau of Labor Statistics reported a median annual wage of $140,300 for computer and information research scientists in May 2025. Load that by 1.4 for employer costs and you get $196,420, or about $893 a working day across 220 days.

A 500-case suite covering the agent's real task distribution takes roughly 30 engineer-days to build, at $26,800. Keeping it current costs 2 days a month, at $21,400 a year. Running it twice a week and on every model change costs about $6,300 a year in tokens, including a judge model on each case.

First-year evaluation therefore lands at $54,500. That is 2.7 times the inference line. If I had to cut one item from this model, the eval suite is the last thing I would touch, because it is the only line that tells you whether the other lines bought anything. The mechanics of building one are covered separately in the guide to standing up an LLM eval suite.

Monitoring is cheap to store and expensive to read

Trace volume sounds frightening and prices out small. A year of agent traces at full prompt and response fidelity is measured in tens of gigabytes, and cloud log ingestion is billed in cents per gigabyte.

The expensive part is the person reading them. Drift, tool breakage and quiet accuracy decay do not fire alerts on their own. Somebody has to define what a bad week looks like, then look every week. That work is the subject of monitoring an agent for drift in production.

This model books a quarter of one engineer against the agent for on-call, dashboard maintenance and incident response. At the loaded rate above, that is $49,100 a year. Whether you hire for it or absorb it is a real decision, and the case for a dedicated role is set out in the piece on hiring an agent ops engineer.

One honest gap: several observability vendors do not publish a list price for their LLM or agent monitoring products. Third-party figures circulate widely and none of them traces to a vendor rate card. I have therefore priced monitoring as labour plus a general platform subscription, and left the tooling rate to be filled in from your own quote.

Review labour is the largest line in this model

Every agent that touches a customer, a payment or a record has a review policy. Either a human checks a defined class of outputs, or a human samples them, or nobody checks and you have accepted the error rate. Most enterprises pick the first.

The arithmetic

Assume 30% of tasks route to a human, at 4 minutes each. That is 72,000 reviews and 4,800 hours a year.

BLS reported a median hourly wage of $21.53 for customer service representatives in May 2025. Loaded by 1.4, that is $30.14 an hour. The review line comes to $144,700 a year.

That is 7 times the inference line, from one assumption about how often a person looks. I think the single biggest budgeting error in agent programmes is treating review as a transitional cost that disappears after go-live. It does not disappear. It changes shape, and the shape it takes is set by your architecture, which is why the design of the human-in-the-loop step is a cost decision before it is a safety one.

Who reviews changes the answer

Swap the reviewer and the line moves violently. A compliance officer at the May 2025 median of $80,730 a year costs about $54 an hour once loaded. The same 4,800 hours then cost $260,800, which is 13 times the inference line and two thirds of the entire first-year budget on its own.

My view is that you should price the reviewer before you price the model. The model choice moves the total by single-digit percentages. The reviewer choice moves it by half.

Integration maintenance is the line nobody renews

An agent earns its keep by writing to systems, not by reading from them. Six connected systems is a modest count for a customer operations agent: CRM, ticketing, identity, warehouse, knowledge base and a payments or billing surface.

What 40 years of maintenance data says

Software engineering settled this question long before agents existed. Robert Glass, summarising the research in Facts and Fallacies of Software Engineering, put maintenance at 40% to 80% of total software cost, averaging around 60%. Nothing about connecting a model to six APIs exempts you from that finding.

This model books 60 engineer-days to build the integrations, at $53,600, and 3 days a month to keep them working, at $32,100 a year. Combined, integration is $85,700 in year one and the second largest line.

Agents make the maintenance worse in one specific way. A deterministic integration fails loudly when a schema changes. An agent often routes around the failure, produces a plausible answer from stale context, and reports success. The write path is where this gets serious, and the controls are covered in the piece on giving an agent write access to legacy systems.

The full cost model, and what moves it

Here is every line in one place, with the basis for each. Replace the assumption column with your own figures and the structure still holds.

Token prices are Anthropic list prices accessed 29 August 2026. Wages are BLS Occupational Outlook Handbook medians for May 2025, loaded 1.4 times for employer costs. The 1.4 multiplier, the task volume, the review rate, the review duration and the platform subscription are assumptions, not measurements.

What moves the inference share

Four changes, each applied to the year-one model on its own.

Nothing you can plausibly do to the model makes tokens the main event. Cutting review by two thirds saves $96,000 and raises the inference share to 7%, which tells you where the savings actually live. That is also the calculation behind which business functions pay back on agents first.

Where this cost model is weakest

This is a model, not a measurement. Here is what would make it wrong.

The assumptions doing the heavy lifting

Three inputs decide almost everything: the review rate, the review duration and the 1.4 employer loading. None is measured. All three came from me, and a reader with real operational data should overwrite them immediately.

The model also assumes a stable task shape. Real agents face a long tail of hard cases that consume far more tokens and far more review time than the median. A model built on medians understates the tail by construction.

And the design is deliberately conservative on architecture. If you run a multi-agent system on a frontier model, the picture inverts. At 4 times the tokens and 2.5 times the price, inference reaches roughly 35% of the total and becomes a line worth optimising. My argument is about the common case, not every case.

The counter-argument deserves its due. Review rates do fall as agents mature, integration work does amortise, and a team running 20 agents shares the monitoring and governance lines across all of them. Per-agent cost at that scale is materially lower than $395,200. What does not amortise is review, because review scales with output volume rather than with agent count.

The wider evidence points the same way. The FinOps Foundation's 2026 survey of 1,192 practitioners managing more than $83 billion of cloud spend found 98% now managing AI spend, up from 31% two years earlier, with AI value management the top skills gap. Organisations are not short of cost data. They are short of a model that connects the cost to the value.

Frequently asked questions

Where to start this quarter

Two things, both doable before the next budget cycle.

First, take one agent already in production and fill in the seven-line table above with your own figures. You will find the review rate is not written down anywhere, which is the finding. Measure it for two weeks before you argue about it.

Second, add a fully loaded cost-per-task line to whatever dashboard already shows your token spend. I would not approve an agent renewal without one. A per-task number that includes the human is the only version of enterprise AI cost that a CFO can compare against the thing it replaced.

Related reading