AI costs: calculate the total cost of a process
Compare AI-supported work using accepted results, service fees, human review, maintenance and exit costs.
Article language: English

In this article
A subscription can look inexpensive until the bill includes time spent correcting answers, handling exceptions and maintaining connections. Calculate AI cost per useful, accepted result. Include data preparation, checking and subsequent handling over the same comparison period. Otherwise a saving can simply move from an invoice to another department's calendar, hoping nobody looks there.
You do not need an elaborate financial model at the outset. Start with one process, a common period and a simple cost breakdown. Separate actual data from assumptions and expenditure from time used. The method below compares ways of performing a task. It does not attribute savings to a company before they are measured and placed in a real work plan.
Choose a unit that can be accepted
Define what you count: a closed request, a checked description, an accepted summary or another recognisable unit of work. Do not count a rejected draft as an accepted unit. Record the quality standard that applies to both ways of doing the task.
The process boundaries must match. If the old measurement ended when a completed reply was sent and the new one ends with a generated draft, the results differ. Include the later review and corrections. For a first trial, an AI pilot plan helps set acceptance criteria before seeing the result; see this article on planning an AI pilot (in Polish).
Build a baseline from observation
Describe current stages and measure representative executions. Include easy, typical and correction-requiring tasks. For each, record the material type, active human work, waiting and acceptance outcome. Do not infer the whole process from the best day or a single demonstration. If the sample is small, say so.
Examine existing tool and maintenance costs too. Some will remain after AI is introduced. Others disappear only after a contract ends or a system is actually switched off. Show when each cost may change. A new feature resembling an old one does not by itself prove the existing service can be dropped.
Separate setup from recurring work
One-time work may include organising documents, configuration, preparing instructions, integration and training. Record the effort and the source of the rate used to value it. Specify the period over which you want to assess the investment. Showing setup separately permits a fair comparison of a short test with a solution intended for longer use.
Recurring work includes preparing each input, operating the tool, checking, correcting and passing on the result. Maintenance includes updates, access assessment, refreshing knowledge and responding to problems. Do not assume instructions written once will remain sufficient. A change in documentation or model may require more work. Name the person who will do it and include it in the budget.
Check how fees are charged
Collect current plan terms, number of accesses, usage-based fees, add-ons and minimum commitments. For an API, a connection used by an application, cost may depend on processed units and functions used. Use the current offer and your own account's actual data for costing, with the date checked. If the process reuses previous answers, include a check of AI cache freshness in the comparison; this article on stale AI cache responses (in Polish) discusses the issue. Fewer new calls must be accompanied by a check that the earlier result still fits the current case.
OpenAI's production guidance (in English) describes monitoring usage and forecasting cost from workload. It concerns the cost of a particular platform. In the company's balance add human work and other services. For user-assigned licences, a review of Copilot before renewal is a separate route (see this article on reviewing Copilot before renewal (in Polish)); for a new product, assess the AI provider and consult this article on evaluating AI providers (in Polish).
Make the cost of checking answers visible
Measure the time needed to compare a result with its source, assess completeness and correct errors. Separate checking simple formatting from review requiring specialist knowledge. If corrections are performed by a different role than before, record that change. Shorter work for one role can increase another's workload, especially with difficult exceptions.
Include cases where a result must be rejected and the task repeated. Omitting them understates the cost of an accepted result. Record the reason for rejection to distinguish unsuitable input, tool error and unclear acceptance criteria. Team training in AI verification helps develop consistent checking rules. See this article on training teams to verify AI answers (in Polish).
Show waiting and human time separately
Elapsed process time matters to the recipient, but it is not the same as active work. If a tool runs in the background, an employee may do another task. If they must monitor successive stages, waiting still limits availability. Record both so that every minute is not automatically converted into an employment cost.
Likewise, a shorter activity does not always reduce expenditure. Freed time may increase the team's capacity, improve timeliness or enable other work. These effects can be valuable but need their own description. Show a financial cost reduction only when an actual expenditure change is confirmed, not merely because a trial took less time.

Build a simple balance without double counting
A working period cost can be the sum of service fees, recurring work, maintenance and the share of setup cost assigned to the period assessed. Divide that value by the number of accepted work units. If no unit passed acceptance, the cost per accepted unit cannot be calculated. Show the total trial cost, the number and reasons for rejection, and the next step instead of entering zero. This is our management-calculation proposal; accounting and tax treatment should be determined separately with the responsible people.
Watch items shared by several processes. The same subscription or administrator should not be counted in full repeatedly. Record an allocation rule or show the shared cost separately. If the share is hard to establish, present a range of scenarios with explicit assumptions. Apparent precision to the last cent does not resolve uncertainty about actual usage.
Worked example: preparing a summary
Consider a process where an employee selects a document, prepares a note and sends it for approval. With AI, input must also be prepared and the answer assessed. Create separate spreadsheet rows for those activities. For every execution, record whether the result passed acceptance, needed correction or was redone manually.
After collecting data, compare the same document category and criteria. If the draft appears faster but the recipient spends longer checking sources, the difference becomes visible. If difficult documents need a separate route, count them separately. This example assumes no favourable result. Its purpose is a calculation that can support another trial, a narrower scope or stopping.
Test sensitivity to unknowns
Choose assumptions that could change the decision: volume, rejected-answer share, checking time, usage fee and maintenance effort. See how the balance changes with different values. Label each scenario as an assumption, not a certain forecast. Then identify the measurement that would reduce uncertainty most. If budget volume comes from a model, assess the forecast of service demand over time separately instead of treating its output as confirmed orders; this article on testing service demand forecasts over time (in Polish) discusses the topic.
If the result depends mainly on the number of exceptions, collect a representative sample. If it depends on removing an old service, check the contract date and technical dependencies. Outage cost and manual takeover can be described as an operational scenario without inventing a probability. An automation process audit can help identify exception sources and work queues; this article about auditing process automation (in Polish) discusses the approach.
Add the cost of changing or ending the solution
Check the work needed to export material, move instructions, disconnect integrations and return to another method. Include temporary parallel service operation if actually needed. Plan acceptance of data and documentation so ending a relationship does not leave the next person to rediscover the process.
A model change can also require another quality check. Assign an owner and time rather than treating every update as cost-free. A comparison procedure is described in this article on AI answer regression tests (in Polish). Rules for deleting copies and closing access should follow a data retention plan covering the whole task; see this article on AI data retention and deletion (in Polish).
How do shared costs differ from the cost of another task?
When considering expanded use, check what actually changes if another group of tasks is added. An existing subscription may already cover some team access, but higher volume may require another plan, extra review or support. Both observations belong in the calculation. Total solution maintenance cost and the cost of expanding its scope answer different decision questions.
In a worked example, one administrator supports several integrations. A share of their work can be allocated to maintenance using an explicit rule. For a new task, however, establish whether current availability is enough. If an extra on-call shift is needed, it is a new expansion cost. If not, do not add one merely to preserve a proportion in the spreadsheet. Recording assumptions helps avoid both errors.
What should the budget approver see?
Prepare a short decision record and a source spreadsheet. State the period, scope, number of accepted results, service cost, human effort and main limitations. Add a quality assessment and the date of the next measurement. The decision-maker should know whether you propose maintaining current work, expanding the trial or changing the method. A cost total without a recommended action leaves the analysis incomplete.
The spreadsheet should allow each material item to be traced to an invoice, usage report, measurement or clearly labelled assumption. Record whether rates include tax and whether different currencies were converted using a common rule. Do not mix costs from different periods without explanation. If the planned change depends on terminating a contract, include that date as a condition instead of showing a benefit from day one.
Base the time item on measuring AI work after including verification; this article on measuring AI work time (in Polish) discusses the approach. Record rejected results and exceptions too. A calculation containing only successful trials may look pleasant but be of little use when planning staffing.
When is the calculation ready for a decision?
The balance is ready when data sources, assumptions, the accepted-unit criterion and the effect of major unknowns can be reconstructed. It should identify the action supported by the numbers and when reassessment will occur. Sometimes the most sensible decision is to collect a missing measurement before expanding the scope.
A cost review can start with one task and its invoices, without declaring a predetermined saving. The Aeneia Audit offering presents that starting point.








