What does an accepted AI result cost?

A cheap model call is not necessarily a cheap result. A team may pay for the first response, retries, retrieval and a reviewer’s corrections—and still hold the work because it does not meet the required standard.

For one repeatable workflow, ask: What did we spend attempting a defined group of tasks, for each output that passed our agreed review?

This guide provides a worksheet for that calculation. It measures workflow-level cost per accepted output, not full ownership cost or organisation-wide return on investment.

Start here

• Define success: agree what one accepted task output means before testing.
• Count the work: include every attempt, relevant service charge and review minute.
• Check the conclusion: reconcile billing and test whether uncertain outcomes could change the decision.

Define acceptance before counting

Choose one task type—for example, drafting an internal status summary from approved inputs. Before comparing routes, record:

• Submitted unit: one eligible task, counted once regardless of calls.
• Required output: format, permitted sources and intended use.
• Acceptance rubric: factual support, completeness and required review checks.
• Decision authority: who may accept the output, and for what use.
• Cohort and cut-off: submission window, eligibility rules and deadline for resolving outcomes.
• Dispositions: accepted, held, rejected, withdrawn or pending.

Acceptance for internal use is not permission to publish, send or act on an output.

Count accepted task outputs, not successful calls. A task requiring three attempts before acceptance contributes one to the denominator. Held and rejected tasks contribute none. Report pending and withdrawn tasks separately; do not remove their incurred costs from the cohort.

Decide beforehand whether corrected outputs qualify. If a reviewer substantially rewrites the work, count that effort. If the result no longer meets the agreed definition of AI-assisted acceptance, classify it accordingly rather than relaxing the threshold after seeing the bill.

Draw the cost boundary

For the cohort, include three components:

1. Model and service charges: every billable attempt, including retries, fallbacks and failures where charged.
2. Other workflow charges: retrieval, search, tools and other paid services used in those attempts.
3. Human effort: actual review, correction, rework and final-disposition time, including work ultimately held or rejected. Apply a stated organisational costing rate; label hypothetical rates in examples.

Included cohort cost = model/service charges + other workflow charges + costed human effort

Cost per accepted output = included cohort cost ÷ accepted task outputs

Record every attempt, even if it has no charge. An unbilled failure can still consume reviewer time or trigger paid follow-up work.

State exclusions beside the result. These may include initial design, integration, training, unallocated subscriptions, maintenance, governance and overhead. Excluded costs have not disappeared: the calculation simply does not cover them. Costed staff time is also not necessarily an additional cash payment or a cash saving.

Keep usage and billing separate

Provider usage describes activity, not business acceptance.

The OpenAI organisation usage and costs reference separates usage measures from monetary cost objects. Its input-token measure includes cached input, so adding cached input again would double-count usage.

Anthropic’s Usage and Cost API documentation likewise separates usage and cost reports. Coverage depends on product and account: the documented Admin API is not for individual accounts, Enterprise uses a separate Analytics API, and Priority Tier costs are excluded from that cost endpoint. Neither provider report supplies your acceptance decisions or reviewer hours.

Charge categories also differ. Anthropic’s caching documentation distinguishes uncached input, cache creation, cache reads and output; its input_tokens field alone is not total prompt usage when caching applies. Preserve provider-specific categories rather than applying one token formula everywhere.

Keep measured consumption and charged money as separate records. Reconcile attributed service charges against account billing or invoices where possible. Credits, subscriptions, tiers, delayed charges and currency conversion can produce differences. Avoid counting both an estimate and its later invoice entry.

For currency conversion, retain the original amount and currency, conversion rate, rate date and method. Label unreconciled amounts as estimates. Do not distribute an entire account bill across a small test without a defensible allocation rule.

Keep a privacy-minimised trace

The register should connect tasks, attempts, costs and decisions without becoming another store of client content.

Record	Minimum useful fields
Cohort	Workflow/rubric version, eligibility, submission window, disposition cut-off
Task	Pseudonymous task reference, submission time, final disposition, decision time
Attempt	Task and attempt references, actual route/model/provider identifiers, timestamps, retry or fallback reason
Charges	Attempt or shared-service reference, charge class, amount, currency, billing period, estimated or invoiced status
Human work	Task reference, review/correction minutes, role or costing category, rate basis
Review	Rubric version, check outcomes, decision, concise reason code, controlled evidence reference if needed

Use actual recorded configurations, not marketing names, to identify routes. Link every attempt to its task and charge evidence where available.

Do not copy prompts, source documents, client names or generated client text into a general analytics log. Keep necessary review evidence in an appropriately controlled system. Restrict access and retention; pseudonymous references still need protection. A reason code such as “unsupported claim” is often sufficient for the cost register.

Flag missing review time, unmatched charges and unresolved outcomes. Gaps are findings, not invitations to invent allocations.

Fill in the worksheet

Fictional illustration only: all amounts below are assumed AUD. They are not provider prices, actual labour rates, exchange rates, measured business results or promises.

Worksheet line	Fictional example	Your cohort
Task type, rubric and dates	One fixed workflow/window	____
Eligible submitted tasks	100	____
Model calls, including repeat attempts	125	____
Accepted task outputs	80	____
Held or rejected tasks	20	____
Pending or withdrawn tasks	0	____
Model/service charges	A$36	____
Retrieval/tool charges	A$4	____
Review and correction time	3 hours	____
Hypothetical reviewer costing rate	A$60/hour	____
Costed human effort	A$180	____
Included cohort cost	A$220	____
Cost per accepted output	A$220 ÷ 80 = A$2.75	____
Exclusions and reconciliation gaps	State explicitly	____

The 125 calls include attempts beyond the first call for some tasks. Charges and review time associated with all 100 tasks stay in the numerator, including the 20 held or rejected tasks. Removing unsuccessful work would understate the cost of obtaining accepted results.

Now test two separate sensitivities:

• Lower acceptance: with the same A$220 cost but 64 accepted outputs, cost per accepted output becomes A$3.44, rounded from A$3.4375.
• More review: with 80 accepted outputs and one additional review hour at the assumed rate, cost becomes A$280, or A$3.50 per accepted output.

These are arithmetic scenarios, not forecasts. They show how acceptance and review effort affect the result without any change in model price.

If nothing is accepted, report the incurred cost and zero accepted outputs. Cost per accepted output is undefined; do not divide by zero or describe the workflow as cheap.

Compare routes on equivalent work

Use authorised routes only. A low observed cost cannot justify an unapproved model, provider or data-handling arrangement.

Keep eligibility, instructions, permitted sources, rubric, reviewer standard and cut-off consistent. Where appropriate, test both routes on the same protected task set. Otherwise, use comparable assigned cohorts and disclose material differences.

Each route may have its own evaluation result, but two tested outputs for one task are not two delivered business outcomes.

Report submitted tasks, dispositions, accepted outputs, calls, charge categories, reviewer minutes, included cost and cost per accepted output for each route. Show uncertainty from small samples, missing time or unsettled charges.

A lower figure is useful only if the routes met equivalent standards and enough outcomes are resolved to support comparison. Provider totals may include other workflows; explain unmatched amounts rather than calling them savings.

Decide—or hold for better evidence

A useful decision note says:

For [workflow and dates], under [rubric], the tested route produced [accepted] accepted outputs from [submitted] tasks. Included cost was [amount], or [amount] per accepted output, covering [charges and effort]. It excludes [items] and retains [evidence gaps]. We recommend [next step within existing authority], subject to [check].

Hold a route or spending decision if criteria changed mid-test, pending outcomes could reverse the ranking, material charges remain unattributed, review effort is missing or cohorts differ substantially. Record what evidence would clear the hold.

This worksheet informs a decision; it does not authorise purchasing, deployment or publication.

Keep the next question distinct

Return to the parent routing guide for candidate authorisation and routing policy. This worksheet measures the cost of accepted outputs within those boundaries; it does not set them.

For setup, infrastructure, maintenance, governance and overhead, use a total cost of ownership assessment. For organisation-wide benefits, opportunity costs and realised value, use an ROI assessment. Neither is interchangeable with cost per accepted output.

Start with one bounded workflow, preserve the quality threshold and make uncertainty visible. A defensible number is more useful than a tidy one.

Source and scope note

The linked provider documentation has a recorded check date of 9 October 2026 and may change. The figures are hypothetical. No account billing, live route performance, Digital Sanctum deployment, exchange rate or client result was verified. Recheck the applicable product terms, account records and current rates before using this worksheet for a real decision.