The hidden total cost of operating an AI agent

The visible bill for an AI agent is often model usage. The consequential bill is the service wrapped around it: preparing data, integrating tools, evaluating behaviour, reviewing uncertain work, responding to failures, managing suppliers and changing the workflow over time.

These costs are not all “hidden” because somebody concealed them. They are hidden because they sit in different budgets, emerge at different stages and are measured in incompatible units. A total-cost model gives them one decision structure without pretending every estimate is precise.

Fix the unit and time horizon

Name the service being costed. “The agent platform” is too broad; “triage and draft responses for UK enterprise support cases, with human approval before send” is costable. Define volume, case mix, environments, authority and service level.

Use two views:

• lifecycle view: discovery through exit over the expected commitment period;
• operating view: cost per month and per accepted workflow outcome.

Separate sunk, committed, fixed, step-fixed and variable costs. A specialist integration already paid for should not distort a forward decision, but it belongs in the historical programme total.

The GAO Cost Guide recommends establishing purpose, scope and a technical baseline before estimating, then documenting assumptions and updating with actuals. That baseline prevents a low-authority pilot estimate from being reused for a production service with very different obligations.

Build the lifecycle stack

Use a work breakdown that follows the service:

Cost layer	Examples
Decide	discovery, workflow mapping, alternatives, data readiness
Create	product, engineering, integration, policy and user experience
Prove	test cases, evaluation, security and privacy review, acceptance
Launch	migration, training, change support, production readiness
Run	model, platform, tools, storage, network, licences and support
Control	human review, monitoring, governance, access administration
Recover	incident response, reconciliation, remediation and communication
Change	model migration, retesting, policy and integration maintenance
Exit	export, transition, deletion, parallel run and contract close

Assign an owner and evidence source to every line. If the answer is “another team absorbs it,” that is still organisational cost. It may not be incremental cash, but it consumes capacity with an alternative use.

Model usage as a workflow, not a token

Inference may be billed by token, request, image, second or reserved capacity. The workflow adds retrieval, embeddings, reranking, tool APIs, databases, sandbox execution, logging and egress. Retries and long context can multiply the initial estimate.

Build from case classes:

Monthly usage cost = the sum, across case classes, of volume × attempts × component cost.

Add concurrency and peak requirements where they change capacity. Model low, expected and high cases for input length, output length, tool loops, failure rate and adoption. Include unused contractual minimums and tier transitions.

Do not optimise token cost in isolation. A smaller context may reduce a bill while increasing errors and reviewer time. Compare cost per accepted outcome.

Price human work honestly

Agents redistribute labour. Account for configuration, review, escalation, correction, supervision, support and governance. Measure minutes on representative work instead of assuming that a required approval is instantaneous.

Include the operating design behind “human in the loop”:

• who receives the work;
• what expertise is required;
• expected arrival and peak volume;
• review time by case class;
• availability outside ordinary hours;
• queues, handoffs and rework;
• training and calibration.

Human review can be a valuable control, not waste. The economic question is whether it is targeted and whether the combined system improves the outcome. A proposal that removes 1,000 minutes of drafting while adding 1,200 minutes of fragmented checking has not created labour capacity.

Include control and assurance as recurring work

Evaluation is not a single pre-launch event. Changes to models, tools, retrieval sources, policies and users can alter behaviour. Budget regression testing, periodic deeper evaluation, grader calibration, risk review and decision records.

The NIST AI RMF calls for measurement throughout the lifecycle and production monitoring. Translate that into named activities and frequencies. The level should be proportionate: a recommendation-only assistant and an agent that executes financial actions should not carry identical assurance cost.

Include identity administration, secrets rotation, retention controls, access reviews and incident exercises. If those capabilities already exist, estimate the marginal load and any step change in tooling or staff.

Price supplier and architecture dependence

Managed platforms can reduce internal operations but introduce contract management, price exposure and switching cost. GAO’s cloud acquisition review identifies data portability, application compatibility and interoperability as considerations in mitigating lock-in.

Estimate the cost to export data and configuration, replace proprietary components, retrain teams, retest the workflow and run systems in parallel. Do this before selection, while architecture choices are still reversible.

The UK Technology Code of Practice asks technology buyers to consider purchasing strategy, interoperability, security, privacy and lifecycle sustainability. Even where it does not govern a private buyer, it provides a useful checklist for costs that a product estimate may omit.

Add incidents and service degradation

Not every failure becomes an incident, but the operating model needs detection, triage, containment, recovery and learning. Estimate routine incident readiness as fixed cost. Model plausible events separately with ranges for investigation, downtime, reconciliation, user support, remediation and external specialist needs.

Avoid inserting a speculative catastrophic number to create false sophistication. Connect scenarios to the workflow’s real actions and data. Keep financial, regulatory, contractual and reputational consequences within the remits of qualified owners.

Also model degraded service: more handoffs, slower latency, a disabled integration or a fallback model. These states may occur more often than a major incident and materially change unit economics.

Use ranges and challenge optimism

Produce a central estimate only alongside sensitivities. The 2026 Green Book discusses risk, uncertainty and optimism bias, warning that costs are often underestimated and benefits overestimated. Its public-sector appraisal method is not a substitute for a company’s finance policy, but the challenge is universal.

Test at least:

• slower adoption and lower volume;
• more review and correction;
• price or exchange-rate change;
• longer integration and parallel run;
• major dependency replacement;
• delayed benefits;
• shorter useful life.

Show which assumptions change the decision. Obtain independent challenge from finance, operations and security rather than leaving the estimate solely with the sponsoring team.

Turn the model into an operating instrument

After launch, replace assumptions with actuals. Reconcile invoices, sample human effort, track accepted outcomes and attribute exceptional remediation. Review variance by case class and release.

Set action thresholds: a rise in retries may trigger engineering analysis; reviewer load may trigger scope change; a price increase may trigger routing or contract review; declining accepted outcomes may suspend the autonomy level.

Reconcile ownership as well as money. The product ledger may show model and platform spend, finance may hold licences, operations may absorb reviewers, and security may fund shared monitoring. Create one service view while preserving each cost centre’s authoritative records. This avoids double counting and makes the trade visible when a saving in one function creates work in another.

Retain the estimate that supported the decision, then compare forecast with actual. Explain variance through volume, price, case mix, schedule, quality and scope rather than silently rebasing the original business case. Failed assumptions are useful portfolio evidence when they remain visible.

Total cost is not an argument against agents. It is how leaders distinguish a system that is cheap to demonstrate from a service that is affordable to operate. The honest business case includes the work required to make the agent useful, governable and changeable—not just the intelligence it consumes.