← All perspectives

Perspective / AI solutions

AI unit economics: modelling the cost of a successfully completed business task

Token prices reveal only part of an AI workflow’s cost. Model retries, tools, infrastructure, human review and failure rates to understand the economics of a successfully completed business task at production volume.

10 min readReviewed

Silicon processor on a precision balance beside an hourglass.
Conceptual editorial image.

The useful cost of an AI workflow is its fully accounted operating cost divided by the number of business tasks completed to the required standard. Cost per token or per model call cannot answer that question alone. A cheaper model may generate more review, retries or failed tasks; a more expensive one may improve the overall economics. This article builds an illustrative cost model for a document-processing workflow, separates operating assumptions from observed evidence and shows which sensitivities a commercial sponsor should examine before scaling.

The executive view

  • Use the accepted business task as the denominator rather than the model call.
  • Include review, failures, retries and operating commitments in comparable cost boundaries.
  • Stress-test workload mix and service performance before approving scale.
01

Choose the unit of commercial output

Start with a task the business recognises. For a hypothetical supplier-onboarding service, one completed unit might be a validated supplier record with the required evidence, correctly linked documents and an approved exception decision where necessary. A generated summary is an intermediate output, not necessarily a completed onboarding.

Define the quality threshold and the point at which the unit is accepted. If a human must correct missing fields, that review effort belongs in the workflow. If an output is rejected and returned to the start, its consumed resources still belong in the cost base. Counting only successful inference calls removes the very failures that can dominate production economics.

Segment the workload. A straightforward domestic supplier, a complex corporate group and a record with missing documents can have different processing paths. A blended average is useful for planning, but the underlying mix must remain visible so that a change in customers does not appear to be an unexplained model-cost problem.

A usable denominator also needs a time boundary. A task started on Friday and completed on Monday should have one stable identity across the reporting periods. Report period costs alongside a matured cohort of initiated tasks, or explain the lag when dividing current spend by current completions. Otherwise a backlog clearance can make one month look exceptionally economical while the work was paid for earlier.

Separate accepted output from business impact. An onboarding record may pass its quality checks without the supplier ever trading. Procurement needs the processing unit to compare technical options; finance may also need cost per activated supplier or first completed purchase. Keep those measures connected without substituting one for another. Define exclusions, cancellations and duplicates before the trial. A system that rejects difficult records early could improve its apparent cost per accepted task while shifting more work back to the team. The rejected population therefore belongs beside the unit-cost measure, with its next operating destination visible.

02

Build a cost ledger beyond model inference

Separate variable consumption from fixed operating commitments. Variable costs can include model input and output, retrieval, search, code execution, storage operations, integration calls and review time. Fixed commitments can include hosting, monitoring, support coverage and the maintenance effort required to keep connectors and evaluation suites working.

OpenAI’s pricing documentation distinguishes model pricing from separately charged tools and infrastructure categories. Treat that as a reminder to inspect the actual billable components of the chosen architecture rather than copying one headline token rate into a business case. Prices, processing tiers and cache treatment should be captured with a date and currency.

Also record what the model excludes. Implementation cost may be evaluated through a separate investment appraisal or amortised over an agreed horizon. Security review, data preparation and business-process redesign do not disappear because they sit in another budget. Make the boundary explicit so that two candidate solutions are compared on the same basis.

Allocate shared infrastructure with a rule the team can reproduce. A retrieval service used by several workflows might be allocated by requests, storage occupied or attributable processing time. The best driver depends on what causes the cost. Record unallocated overhead separately rather than forcing it into a conveniently small category. For small pilots, fixed minimum charges can dominate; at volume, review capacity or an integration tier may become the constraint.

A reconciliation should connect task-level consumption to provider invoices and internal time records. Task telemetry is often estimated, while the bill is actual. Explain differences such as cached input, batch discounts, reserved capacity and currency conversion. Do not silently replace actual expenditure with a theoretical list-price calculation. The FinOps Foundation’s unit-economics capability provides a useful reference for relating expenditure to business units; the allocation model here is an analytical proposal. Its purpose is to expose which decisions can change spend and which costs will remain after a local optimisation.

03

Model attempts, review and accepted outcomes separately

Use a simple accounting identity: cost per accepted task equals total relevant workflow cost divided by accepted tasks. Behind that identity, retain separate counts for tasks initiated, model attempts, tool calls, review episodes, retries and final dispositions.

A task can involve multiple attempts without becoming multiple business outputs. Conversely, one batch call can process several records, each needing its own acceptance decision. The telemetry model should reflect those relationships rather than forcing a one-call-to-one-task assumption.

Anthropic’s evaluation guidance identifies task cost, latency and errors as measurable alongside quality. Connect that evaluation evidence to the cost ledger. A candidate model should be assessed on the same representative tasks and review standard as the baseline. Otherwise an apparent saving might simply reflect a more permissive definition of success or an easier test sample.

At instrumentation level, attach the business task ID to attempt IDs, tool executions and review decisions. Keep the acceptance event separate from a model's own completion message. A worker can report success while a downstream write fails; the accepted-task count should advance only after the required record and evidence are present. If final quality is sampled rather than checked on every output, show that distinction and the uncertainty it introduces.

Evaluation should include the actual routing policy. A small model that escalates uncertain cases to a larger one has different economics from either model assessed in isolation. Simulate the escalation thresholds, human queue and retry limits using the same task mix. Review failures individually before changing thresholds: a missing document, an incorrect extracted value and a tool outage need different remedies. Increasing model capacity may improve one problem while leaving the other two unchanged. This is why operational traces are decision evidence, rather than optional engineering detail beside the financial spreadsheet.

04

Work through a fully illustrative monthly comparison

Suppose two configurations each receive 10,000 onboarding tasks. Configuration A has higher direct processing cost but creates fewer review cases. Every figure below is an invented planning assumption, not a vendor price or a claimed client result.

Monthly measureConfiguration AConfiguration B
Direct processing£2,000£1,500
Retry and rework processing£300£300
Human review£6,000£9,000
Fixed operating allocation£2,700£2,700
Total relevant cost£11,000£13,500
Accepted tasks9,0008,800
Cost per accepted task£1.22£1.53

Configuration B saves £500 in direct processing but costs about 25% more per accepted task. The reason is visible: review effort increased and fewer tasks reached acceptance. This does not establish that a larger model is always better. It demonstrates why direct model spend is an inadequate optimisation target. Validate every assumption through representative trials before treating either configuration as investable.

The sensitivity can be expressed directly. For configuration B to match A's approximately £1.222 cost per accepted task while accepting 8,800 tasks, its total cost would need to fall to about £10,756. With its other costs unchanged at £4,500, review expenditure would need to fall from £9,000 to about £6,256. That is approximately a 30.5% review-cost reduction. The gap is large enough to define a concrete trial question rather than a general instruction to optimise prompts.

Do not infer a payback period from this operating comparison alone. Suppose, as a further hypothetical assumption, A requires £40,000 more implementation expenditure than B and saves £2,500 per month in total operating spend at equivalent value. Simple undiscounted payback would be sixteen months. If acceptance volume or quality differs, that comparison needs adjustment before use. Finance should also test discounting, transition duration and the possibility that the benefit appears as released capacity instead of reduced expenditure. The additional calculation makes the investment assumption explicit; it does not validate it.

05

Price review according to real work, not theoretical seconds

Human review is often modelled as a uniform checkbox. In practice, reviewers may need to locate evidence, understand an exception, correct a record and document the decision. A six-minute review at a fully loaded £30 per hour costs £3 before any further rework. A workflow that requires 2,000 such reviews therefore carries £6,000 of review effort.

Measure the distribution by case type. An ambiguous supplier record may interrupt a specialist, while a missing formatting field can be handled by an operations team. The appropriate labour rate and availability differ. Include queueing delay where the business has a time-sensitive service commitment.

Do not assume all saved minutes become cash savings. They may create capacity, reduce backlog, improve service or release time for higher-value work. Translate those benefits into the decision honestly. A staffing reduction, avoided recruitment and increased throughput are distinct commercial cases with different conditions for realisation.

Review queues behave differently from a fractional labour estimate. If 20% of 10,000 tasks require three minutes of review, the direct workload is 100 hours. At a hypothetical loaded rate of £45 per hour, that is £4,500. It is a capacity estimate, not necessarily a removable payroll cost. Scheduling, training, absence cover and specialist escalation may require a larger staffed commitment. A queue concentrated into two mornings can need more concurrent people than the same work distributed evenly.

Measure the reviewer interface as part of the configuration. Showing the extracted value beside the source passage may reduce inspection time; requiring a person to search an entire document may erase the inference saving. Record correction time, evidence retrieval and the time spent resolving disagreements. Sample apparently accepted outputs to look for missed defects, with appropriate handling of sensitive records. Otherwise a fast reviewer can appear efficient because errors are left undetected. Changing review design can therefore improve both cost and quality, but those improvements need observed evidence.

06

Stress-test volume, quality and latency together

Run sensitivities around the assumptions most likely to move the decision. Typical candidates include task volume, exception rate, input length, cache effectiveness, review duration and accepted-output rate. Show both a central scenario and plausible adverse combinations rather than varying one convenient input at a time.

Volume can spread fixed costs while increasing operational pressure. A queue that is economical at average demand may fail its service target during a month-end spike. Repeated tool timeouts can raise cost and latency together. Model those coupled effects, including the consequences of rate limits, unavailable dependencies and manual fallback.

Also examine the cost of quality improvements. Better document extraction, narrower tool contracts or an additional deterministic check might reduce review more effectively than changing models. The architecture is the unit being optimised. Establish whether the proposed change improves total task economics while preserving permissions, accuracy and the required time to completion.

07

Set commercial gates for a controlled scale decision

A production business case should state the accepted-task definition, measured baseline, cost boundary, review model and downside scenarios. Name the assumptions still based on estimates and the evidence required to replace them. Avoid presenting a highly precise return figure when the success rate or review burden remains uncertain.

Approve a bounded operating envelope: expected workload mix, monthly exposure, maximum review backlog and required quality. Establish triggers for pausing, rerouting or revisiting the model when those conditions change. Unit economics should be monitored after release, not stored permanently in the original approval spreadsheet.

The executive question is whether the workflow produces a worthwhile commercial outcome at a sustainable total cost. That may justify a higher inference bill, a simpler deterministic process or a narrower AI scope. The useful technical capability is being able to explain that choice with traceable workload evidence rather than treating a lower token price as the business result.

Further questions

Questions this raises

How do you calculate the cost per successful AI task?

Divide the relevant total workflow cost by tasks accepted to the agreed quality standard. Include resources consumed by failed attempts, retries and required human review within the chosen accounting boundary. Keep the denominator tied to a business acceptance event and a defined reporting cohort. If a task remains pending, it should not disappear from the picture: report pending volume, age and expected resolution alongside completed outcomes. A low unit cost accompanied by a growing unresolved backlog may indicate that work is being deferred rather than handled more economically.

Is the cheapest AI model usually the most economical?

Not necessarily. Lower inference cost can be outweighed by additional review, failures or latency. Compare complete configurations on representative tasks using consistent acceptance criteria. The economical configuration depends on the workload and the intervention policy. Test model routing, retrieval, retry limits and the review interface together. An inference-price reduction can be attractive, but acceptance quality, review time and support effort determine whether the whole service has actually improved.

Should saved staff time be counted as a cash saving?

Only when the operating plan can realise that saving. Released capacity, reduced backlog, avoided hiring and lower expenditure are different benefits and should be modelled separately. For example, a team may use released capacity to reduce a backlog without changing headcount. That is a useful operational benefit, with a different financial treatment from an eliminated contract or avoided hire. Give each benefit an owner, a baseline and a plausible route to realisation.

References

Sources & further reading

  1. OpenAI API pricing Ongoing pricing; checked 17 November 2026
  2. Demystifying evals for AI agents 9 January 2026
  3. FinOps Framework: Unit Economics

Connected capabilities

AI solutionsStrategy & digital transformation

From thinking to doing

What could this mean
for your business?

Your next chapter

What could your
business do better?

Tell us where you want to go, what’s getting in the way, or what you need to deliver. We’ll start there.

Start a conversation

Start a conversation

Tell us what
you have in mind.

Interactive preview. Your information stays in this page and is not sent.