AI automation costs: beyond the model bill.
Estimate cost per completed business task, including retries, human review, exception handling, operations, and setup. Token spending is only one input. Compare the total with the same workload handled manually, and separate released capacity from cash savings.
Inside this article 6 sections +
A low model bill can hide an expensive workflow. A draft might cost very little to generate yet require several minutes of review, repeated retries, or a specialist to resolve exceptions. A useful estimate starts with the unit the business actually cares about: a completed request that meets the agreed quality standard.
This worksheet uses hypothetical inputs so you can see how the calculation behaves. It is not a quote, a vendor price list, or a prediction of savings. Replace every assumption with your own observations. Use one currency consistently; the example labels it as euros solely to make the arithmetic concrete.
Try the workload, not just the token price
Suppose a team handles 2,000 requests per month, averaging eight minutes each. An automated path still requires review on some requests and extra handling on exceptions. Review and exception time are separate, additive activities in this model, so an exceptional request may incur both. If your measurements already combine them, enter the combined time in one field and set the other to zero.
Illustrative model, not realized savings. Assumes all requests eventually complete. Inputs stay in your browser. Without JavaScript, the default worked example remains visible.
Make every assumption inspectable
The manual baseline is volume multiplied by manual minutes, divided by 60, then multiplied by loaded hourly cost. The automated path adds machine cost, review labor, exception labor, monthly operations, and an allocation of setup cost. Machine cost should include the average expense of unsuccessful attempts and tool usage, not only the final successful model call.
Manual = N × manual_minutes ÷ 60 × hourly_cost
Review = N × review_fraction × review_minutes ÷ 60 × hourly_cost
Exceptions = N × exception_fraction × extra_minutes ÷ 60 × hourly_cost
Automated = N × machine_cost + Review + Exceptions
+ monthly_operations + setup_cost ÷ allocation_months
Unit cost = Automated ÷ N
For the default inputs, model and tool usage totals €160, review labor approximately €933, and exception labor €700. Operations add €900 and setup allocation adds €1,000. The resulting €3,693 monthly estimate is lower than the €9,333 manual labor baseline. That arithmetic demonstrates the model; it supplies no evidence that your business will achieve the same review rate or completion quality.
Separate capacity from money recovered
The example releases 220 modeled labor hours each month. That can create room for faster responses, more complex work, or growth without equivalent hiring. It does not mean the payroll immediately falls by 220 hours of wages. Salaries, staffing commitments, and the timing of workload all affect whether released capacity becomes a cash saving.
The setup allocation is also different from cash flow. Paying an implementation bill in the first month and spreading it over twelve months changes the management view, not the payment date. Build a separate cash schedule if timing matters. Avoid presenting the cost difference as a return on investment until implementation costs, operating duration, benefits, and failure scenarios have all been defined consistently.
Change the assumptions that can overturn the decision
Increase the review rate to 100 percent and lengthen the review time. Does the proposal still look useful? Reduce monthly volume while leaving platform and operational costs fixed. Does the cost per request become unattractive? These are more revealing exercises than debating a small difference in token pricing before you know how the workflow behaves.
Quality deserves a separate gate. A system can lower handling cost while making unacceptable errors. Include costly exceptions in the evaluation set, track complaints and rework, and do not treat an uncompleted request as a cheap success. If some requests fail permanently, calculate unit cost using the number that actually completed, with failure handling included in total cost.
Buy evidence for the investment decision
Before commissioning a larger AI build, agree on the decision a smaller paid pilot must support. For example: can draft-assisted handling improve a particular support queue once active review, exceptions, and maintenance are included? Specify the queue, representative request types, observation method, and person who accepts the result. A pilot that quietly changes the mix of cases cannot answer the original cost question.
The useful delivery package includes the baseline observations, instrumented workflow, editable assumptions, and a record of unresolved costs. Ask for the raw operation-level observations in a form your team can inspect, with sensitive content removed where appropriate. Keep implementation effort separate from recurring provider usage and human work. Agree who pays for test usage and who may approve an increase in the experiment's budget.
Continue when the evidence supports a useful outcome at an acceptable operating burden. Narrow the workflow when exceptions dominate. Stop when the result depends on access, quality, or staffing that the business cannot provide. A recommendation to stop can be a valuable consulting result because the purchased outcome is a better investment decision. Only observed results should replace the assumptions in this worksheet; a polished demonstration is insufficient.
Collect the evidence a pilot needs
Instrument each operation with a stable ID. Record attempt count, provider usage, review duration, exception reason, and final outcome. Sample review time carefully: leaving a browser tab open is not the same as actively reviewing a draft. Compare similar categories of work so a shift toward easier requests does not masquerade as an improvement.
Discuss the result with the people doing the work. They can explain whether a draft saves thinking time, creates hidden checking work, or shifts difficulty to another team. Combine that evidence with the worksheet before expanding the project. Our AI integration service can help define a bounded pilot; the agent-versus-workflow guide helps choose a design to test.
Review effort is one of the cost model’s most important assumptions; examine how to measure it in Measure AI review time before claiming savings.