The model bill is one line in an AI feature’s budget. Discovery, data preparation, integration, evaluation and human review determine whether the feature does a useful job. Price the complete workflow and state its assumptions before turning saved minutes into a return claim.

Separate feasibility from production implementation

A feasibility scope establishes the task, approved sources, data access and evaluation criteria, then tests a narrow end-to-end path. Production work adds authentication, interface behaviour, integration reliability, monitoring, fallback and handover. Neither scope has a universal price; missing data, uncertain business rules and complex permissions are estimation drivers.

  • Data work: source cleanup, extraction, ownership and permission mapping.
  • Integration work: APIs, retries, duplicate prevention and review or approval paths.
  • Assurance work: evaluation fixtures, difficult cases, deterministic checks and release evidence.
  • Operational work: configuration, account ownership, runbook, support and maintenance.

Estimate recurring cost from measured usage

Collect input and output usage from representative prototype runs. Estimate monthly requests, retrieval and storage, tool execution, retries and scheduled evaluations. Record the selected model and enabled features. Apply current rates rather than assuming a provider’s cheapest headline rate describes your requests.

Provider references checked 11 October 2026: OpenAI: API pricing and Anthropic: API pricing. Rates, model availability and feature charges can change. The arithmetic below uses invented scenario costs, not these providers’ quoted rates.

Illustrative costing formula; use actual usage and current rates for your configuration.
Monthly model cost =
  (input tokens / 1,000,000 × input rate)
  + (output tokens / 1,000,000 × output rate)
  + applicable cache, tool and other usage charges

Monthly operating cost =
  model + retrieval/storage + hosting/execution
  + licences + maintenance + additional review/exception effort

Distinguish spend monitoring from enforcement. Set application request-size limits, retry ceilings and job budgets, and test what happens when a limit is reached. A cost alert delivered after a burst does not necessarily prevent the burst.

A hypothetical intake estimate, with visible assumptions

Suppose a process handles 1,000 tasks per month and 60% use the pilot. Baseline handling is 8 minutes. Assisted handling takes 3 minutes plus 2 minutes for review and correction, so the net saving is 3 minutes per adopted task. That is 600 × 3 / 60 = 30 hours of potential monthly capacity.

At an assumed loaded staff cost of €30 per hour, that capacity is valued at €900. Suppose incremental operating cost is €500 per month and implementation is €8,000. Potential net monthly value is €400; simple payback is 20 months. These are fictional inputs for explanation, not a quote, client result or guarantee of cash savings.

For the same fictional assumptions, first-year simple ROI is (12 × €400 − €8,000) / €8,000 = −40%. That calculation treats released capacity as value and ignores financing and tax effects. It is not a forecast of cash return. Use your chosen observation period and count benefit and cost consistently when comparing options.

The €500 assumption excludes the review minutes already included in assisted handling; counting them again would double-count effort. Include any additional supervision or exceptions that those two minutes do not cover. If review/correction rises to four minutes, only one minute is saved per adopted task, or ten hours. At the same assumed staff cost, €300 of capacity does not cover €500 of operation.

Illustrative capacity calculation; local fixtures verify the arithmetic, including a pilot that loses time.
function monthlyCapacityHours(tasks, adoption, baseline, assisted, review) {
  return tasks * adoption * (baseline - assisted - review) / 60;
}
// Minutes per task; adoption is a fraction between 0 and 1.
// 1000, 0.6, 8, 3, 2 gives 30 hours.
// Validate finite, non-negative inputs in a production calculator.

Capacity, quality and cash are different outcomes

Time released creates value only when the organisation can use it. It is not automatically a payroll reduction. A support assistant may help staff answer more difficult requests; a reviewer may find a defect earlier. Quantifying those effects requires a baseline, observation window and evidence about what happened next.

Track correction time, error rate and adoption during the pilot. For conversion or revenue claims, establish a measurement method that accounts for traffic quality and other changes. Do not assign every extra enquiry to the AI feature merely because it launched at the same time.

What to bring to an estimate

Provide the task, sample approved inputs, expected volume, integration access, data constraints and who will review outputs. AI integration scope separates prototype and launch deliverables; AI maintenance covers recurring ownership. Share the assumptions and we can scope the first evidence-producing step.