What an AI Agent Really Costs: A Six-Line TCO Model.

IN ONE MINUTE

The licence or token bill is only one part of what an AI agent costs. A realistic total cost of ownership has six lines: model and inference, integration and build, data preparation, verification and human oversight, governance and security, and ongoing monitoring and change. McKinsey found that about 60% of an agentic task's cost goes into refining answers, and 93% of organisations it surveyed had exceeded their AI budgets. The fairest way to compare an agent with the current way of working is cost per outcome, against the fully loaded cost of the work it actually replaces.

THE OBJECTIVE

To give leaders a simple, complete template for costing an AI agent before approving it, and for comparing it fairly with today's process.

WHAT THE RESEARCH SHOWS

McKinsey's recent work on agentic economics (Jul 2026) is one of the most useful pieces on this topic:

• About 60% of an agentic task's cost is tied to refining answers: checking, repairing and re-running.
• The same task can vary in cost by a factor of 30 between runs, depending on the path the agent takes.
• Around 10% of users account for about 65% of token consumption.
• 93% of respondents reported exceeding their AI budgets, and a fifth had limited AI use because of operating costs.

Gartner (2025) predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.

None of this argues against agents. It argues for costing them properly.

THE FRAMEWORK: THE SIX-LINE AGENT TCO MODEL

Below is an illustrative monthly model for an agent handling 20,000 customer queries a month. The figures are examples to show the structure, not benchmarks.

Line	What it covers	Illustrative monthly cost
1. Model and inference	Tokens, API calls, hosting	£4,000
2. Integration and build	Engineering to connect systems, amortised over 24 months	£6,250
3. Data preparation and upkeep	Cleaning, knowledge-base updates, access controls	£2,000
4. Verification and oversight	Compute for refinement plus human review of exceptions	£9,000
5. Governance and security	Policy, audit, testing, compliance, risk controls	£2,500
6. Monitoring and change	Performance tracking, retraining, vendor management, user support	£2,250
Total		£26,000 a month (about £1.30 per query)

In many real cases, lines 4 to 6 end up larger than line 1. That is often where budgets are exceeded.

HOW AGENT COSTS BEHAVE

Agent costs don't behave like traditional software costs, and that changes how they should be managed:

• They are mostly variable. A licence is a fixed cost. An agent's inference and refinement costs rise with every query, every retry and every longer answer. Growth in usage is good news for value and a risk for the budget at the same time.
• They are uneven. McKinsey's finding that about 10% of users drive about 65% of token consumption means averages hide the real exposure. Budget for the heavy users, not the typical one.
• They can fall with scale, but only if managed. Build and integration costs are spread over more volume as usage grows. Inference costs only fall per unit if the design is optimised.

This is the same lesson as Issue 004: a budget without a ceiling and a named owner is a blank cheque.

SIX LEVERS THAT REDUCE AGENT TCO

Lever	What it means	Which TCO line it moves
Right-size the model	Use smaller, cheaper models for simple steps; reserve the most capable model for hard ones	Line 1
Reuse and cache	Store and reuse answers to common questions instead of generating them each time	Line 1
Set usage limits	Per-user and per-task caps, with alerts before thresholds are reached	Lines 1 and 4
Narrow the scope	Start with well-defined, high-volume tasks where the agent succeeds first time more often	Lines 1 and 4
Set review thresholds	Send only low-confidence or high-risk outputs to human review, not everything	Line 4
Build once, reuse widely	Shared integrations, data pipelines and governance controls across several agents	Lines 2, 3 and 5

COMPARING FAIRLY WITH TODAY'S PROCESS

Three principles help keep the comparison honest:

• Compare cost per outcome, such as cost per resolved query, not cost per token or per licence.
• Compare against the fully loaded cost of the work actually removed, including on-costs, tools and management time, rather than a headline salary.
• Include what people still do. Most agents change a role rather than remove it, so the human share of the work stays in the model.

BUILD, BUY OR PARTNER

The TCO model also helps with the sourcing decision:

Option	Where it tends to fit	What to watch in the TCO
Buy a packaged agent	Common, standardised tasks where speed matters more than differentiation	Usage-based pricing as volume grows; limits on customisation
Build on a platform	Core processes where your data and workflow are a source of advantage	Integration, data upkeep and the in-house skills to run it
Partner	Complex programmes where you need capability quickly and plan to build your own over time	Knowledge transfer, exit terms and who owns the intellectual property

WHAT THIS MEANS FOR CFOS AND PROCUREMENT LEADERS

Ask for all six lines in every agent business case. A proposal that only shows model or licence costs is incomplete. Require the full TCO before approval.

Contract for predictability. Negotiate usage caps, volume tiers and clear pricing for overages. Understand how costs change if usage doubles.

Track cost per outcome monthly. Put cost per resolved query, case or transaction into the management accounts, alongside the fully loaded cost of the previous process.

WHAT THIS MEANS FOR TECHNOLOGY LEADERS

Design for cost from day one. Model routing, caching and usage controls are far cheaper to build in at the start than to retrofit.

Make costs visible by agent and by task. Tag usage so every agent's cost can be traced to a use case and an owner.

Plan for change. Models, prices and vendors change quickly. Avoid designs that lock the organisation into one provider without an exit route.

WHAT THIS MEANS FOR BUSINESS OWNERS

Own the outcome and its cost. If you own the process, you own its cost per outcome, whether the work is done by people, agents or both.

Keep the human share in the plan. Most agents change roles rather than remove them. Budget for the review, exception handling and customer judgement that remain.

Scale only what is proven. Expand an agent's scope once cost per outcome is stable and better than today's process, not before.

THE LAST WORD

The token bill is the price tag. The six lines are the cost.

When your organisation last approved an AI agent, how many of the six lines were in the business case?

—

Until next week,

Artificial Ian

Tell your network: if someone you know is building an agent business case, the six-line template might save them a surprise later.

→ Found this useful? Forward it to a finance, procurement or technology leader.

→ Subscribe to The AI Executive Briefing, weekly in your LinkedIn inbox.

Sources: McKinsey, "Is that AI agent worth it? Agentic economics and the modern operating model" (Jul 2026); Gartner, agentic AI project prediction (Jun 2025). Cost table is illustrative.