10 / AI at Work · Field briefing
Cheap tokens. Expensive work?
Count everything a useful result needs.
Per-token pricing is one component of operating cost. The work also involves preparation, tools, infrastructure, retries, checking and correction. For a workflow decision, measure cost per accepted task alongside quality and latency.
What does an accepted AI-assisted result really cost?
An explanatory schematic—not a live system or measured performance.
KAF World Digital · Prepared 4 October 2026 · First editorial edition. Worked examples are invented for teaching.
Try the example02 / Make the denominator honest
Make the denominator honest
Begin with a numerator you can explain and a denominator that matches the outcome. If a trial attempts 100 drafts and 80 meet the acceptance rules, the denominator is 80 accepted drafts. Counting all 100 attempts as successes would hide the cost of failure.
Consider an illustrative calculation, not a provider quote. Assume the trial incurs £12 of model and tool charges, £8 of allocated infrastructure and 120 review minutes at an assumed £30 per hour. Total measured cost is £80; divided by 80 accepted drafts, that is £1 per accepted draft. The example excludes setup and rejected-task recovery unless those costs are added, and does not assert a financial return.
Preparation and checking may dominate a small trial. Track what a person actually does rather than assigning all elapsed time to AI processing. Avoid double-counting work already included in another cost category. Report fixed setup costs separately, then show the volume over which any allocation is spread.
Compare changes while holding acceptance rules stable. A smaller model, shorter prompt or cached context may reduce charges, but it can also affect quality or tool usage. A longer prompt may cost more per call while reducing retries. A cheaper request is not automatically a cheaper completed task.
Caching and batching have conditions. Cache benefit depends on reusable context, supported pricing and actual hit rates. Batching suits work that can wait; it does not replace interactive response requirements. Measure the actual configured service rather than carrying one vendor's current rates into a universal claim.
The practical comparison is with the existing workflow and its required quality. Include review effort, missed work and failure consequences where they can be estimated. Keep uncertainty visible when a short trial cannot establish typical demand or rare-event costs.
Evidence for the technical mechanisms: Anthropic · Cost Optimization, Anthropic · Building Evals. The worked examples and practical framework are KAF editorial recommendations.
03 / Worked example
Calculate accepted-task cost
Measure what a useful result costs, including the checking. This interactive lesson uses prepared synthetic examples and local calculations.
Use the companion worksheet and the worked example. All assumptions are illustrative.
No account, model API or external service is connected. Inputs are not submitted or stored by this page.
04 / Go deeper
Useful words. Important details.
- Accepted-task cost
- Defined total cost divided by the number of outputs meeting the acceptance rules.
- Fixed cost
- An expense that does not vary directly with each attempt over the stated range.
- Marginal cost
- The additional cost of one more unit of work.
- Cache hit
- Reuse of supported cached context; savings depend on the service and actual reuse.
Use a defined accounting boundary
Total cost can include model/tools, infrastructure, labour and allocated fixed costs. State the period and avoid double-counting categories. A rejected attempt belongs in the numerator’s cost even though it is not an accepted result.
The zero case matters
With zero accepted tasks, cost per accepted task is undefined, not zero. The lesson deliberately displays that condition. It is useful evidence of a workflow that has not yet produced a qualifying output.
Optimisation has a quality constraint
Compare caching, batching, model choice and context changes against the same acceptance rules. Measure actual usage; provider prices and cache conditions change. Cheaper calls can become more expensive work if they increase retries or review.
The technical notes qualify the explanation; they are not product-specific assurance or a certification.
05 / Evidence & scope
Inspect the source.
Keep the context.
Sources were accessed on 4 October 2026. These technical and security references support the mechanisms described; they do not validate our fictional examples or establish comparative product performance.
- Anthropic · Cost Optimization
Baseline measurement, caching, context and quality/cost trade-offs. No notebook price is presented as a current quote.
- Anthropic · Building Evals
Task-specific evaluations and code, human and model-based grading. Provider examples are not independent comparative evidence.
What this briefing does—and does not—establish
The worked examples, suggested questions and worksheets are original KAF teaching material. Numbers are explicitly illustrative. No customer data, observed adoption results, market-size forecast or live provider pricing is presented.
Check the exact product, release, configuration and policy before using a workflow. Source documents can change; this is a dated educational briefing.
Keep a useful copy.
Read the full briefing offline, or use the companion worksheet to apply it.