Comparing AI options without hiding the operating cost

Research note. Prepared 8 November 2026. Original decision method informed by operational guidance; no supplier price survey or measured savings claim.

The price of a model call or accelerator is one line in the cost of delivering a task. The rest may include integration, infrastructure, support, monitoring and human checking.

A useful comparison holds the task and service expectation steady while making those responsibilities visible.

Define the unit of useful work

Start with an outcome the organisation can recognise: a document processed to the required standard, a support case resolved with appropriate review or an inspection completed within the required time.

Record the quality and operating conditions that qualify the result. Cost per attempted task can be misleading if one option produces many more failures or requires substantially more correction.

Compare operating arrangements

The table identifies cost questions. It does not imply that one arrangement is cheaper or more suitable in every case.

Use a transparent calculation

For a defined period, estimate:

Total operating cost = usage and infrastructure + allocated implementation cost + operations and support + review and rework + other directly attributable costs.

Then divide by the number of tasks completed to the agreed standard. State the period used to allocate implementation and hardware costs. Keep one-off and recurring items visible rather than hiding them in an unexplained unit price.

Avoid double-counting staff time. If a managed-service quote already includes support, identify what remains with the organisation before adding another support allowance.

Show which assumptions can change the result

Test low, expected and high workload cases using the same quality requirements. Include a case in which human review takes longer than expected and one in which a supplier or model change creates additional evaluation work.

For edge hardware, the workload must fit the intended configuration. For self-hosted models, idle capacity and the skill needed to operate the service can matter as much as peak throughput.

Do not assign false precision to unknowns. Show a range and identify the measurement that would narrow it. A short trial that measures correction effort may be more useful than a detailed spreadsheet built on an unsupported assumption.

Separate cost from suitability

An option can be inexpensive and still fail a necessary data, performance or operating requirement. Identify those conditions first. Compare cost among the options that plausibly meet them, and make conditional suitability visible.

Exit costs also belong in the decision: moving data, rebuilding integrations, replacing hardware or re-evaluating a new model. You do not need to predict every future change, but the major dependencies should be understood.

Research context

NCSC guidance places security responsibilities across development, deployment and ongoing operation. Those responsibilities create real work and should appear in the operating plan and cost assumptions. NCSC secure AI guidance.

A useful final comparison states the preferred option, the assumptions that support it and the findings that would change the choice. That gives the organisation a decision it can revisit as real usage becomes clearer.