Development

How to Estimate the Ongoing Cost of an AI Feature

Estimate the ongoing cost of an AI feature from the work it performs, not from a single example prompt. Count requests, model input and output, retrieval, retries, storage, infrastructure, and the human effort needed to review or correct results.

A useful estimate exposes assumptions and shows a range. It should become more accurate after a pilot, rather than become a number that everyone is afraid to revise.

Estimate ongoing AI feature costs by workload

Define one successful business outcome. For an internal knowledge assistant, that might be an answered question with permitted, relevant citations. For an automation, it might be a completed workflow requiring no corrective action.

Then list the operations behind that outcome:

Cost componentMeasurementCommon omission
Model callsCalls and tokens by modelPlanning calls and retries
RetrievalQueries and reranking workAccess filtering and index refresh
StorageSource, index, logs, backupsRetention growth
InfrastructureCompute and service usageBackground processing
Human workReview and correction minutesEscalations and support

Do not compare two designs using different definitions of success. A cheap response that requires manual repair can cost more per useful outcome.

Keep provider prices as replaceable inputs

Record the provider, model, billing unit, currency, and date for each price. Consult the provider's current official pricing before making a purchasing decision. This article intentionally supplies no supposedly permanent price table.

Some services bill input and output separately, and additional features may have their own units. Cached input, batch processing, image inputs, and tool use can change the calculation. Use the actual billing categories of the selected service.

For cost-management practices rather than a provider-specific quote, see the FinOps Foundation's AI cost and usage guidance. Verify the service invoice against your own usage records.

Build a transparent worksheet

Start with this model, replacing each term with your measured workload and current unit price:

model cost = sum(calls × mean billable units per call × price per unit)
service cost = retrieval + storage + compute + monitoring
human cost = reviewed outcomes × mean review minutes ÷ 60 × hourly cost
monthly operating estimate = model cost + service cost + human cost
cost per useful outcome = operating estimate ÷ accepted outcomes

Keep units consistent. If a rate is quoted per million tokens, divide the token count by one million before multiplying. Include retry calls either in call volume or as a separate term, but avoid counting them twice.

A hypothetical pilot might have 1,000 accepted outcomes and 1,300 total model calls. That ratio matters more than assuming every question produces exactly one call. These numbers illustrate a worksheet; they are not observed results from a customer deployment.

Model quiet, expected, and busy periods

Estimate ordinary demand, a plausible busy period, and a failure scenario. A retrieval outage can increase retries and human escalations while accepted outcomes fall. Average prompt length does not describe a single unexpectedly large document.

Set bounded context, tool-call, and retry budgets in the implementation. Define behavior when a limit is reached. A graceful escalation is often easier to operate than a workflow that keeps spending until it happens to succeed.

Separate implementation costs from recurring operations. A one-time document cleanup project is not the same as monthly ingestion, but both belong in the overall investment decision.

Reconcile estimates with evidence

Log request identifiers, model names, billable usage, outcome status, and relevant latency without copying sensitive content into unrestricted logs. Compare aggregate usage with invoices. Keep enough detail to explain variance by workload or model change.

The AI coding tool ROI article covers a different decision: whether coding assistance creates value. This worksheet estimates the operation of an AI product feature. Avoid using subscription adoption metrics as a substitute for its running costs.

Can I promise a fixed monthly AI bill?

Only when the commercial and technical limits support that promise. A budget alert alone does not prove that spending stops. Test the actual limits, fallback behavior, and responsibilities before offering a commitment.

For help bounding a pilot's economics and acceptance criteria, scope an AI feature review.