Your AI Budget Needs a Cost per Completed Job

A cheaper model can still make an expensive process. The price per token is not the number your business lives with. The cost of finishing the job is.

Gartner warned in August that the inference cost per agentic workflow could rise more than fivefold through 2028. In September, IBM made a related point: companies often see the AI bill but struggle to connect the full cost to a business result. Neither finding says AI is a bad investment. Both say the unit of measurement needs to change.

The invoice is not the whole job

Take an incoming invoice. An AI system may read it, extract fields, check the purchase order, flag a mismatch, ask for an approval, and update an accounting record. A dashboard might show the price of the model calls. It will not necessarily show the staff time spent resolving the mismatch, the retry after a failed integration, or the invoice that had to be corrected a week later.

A workflow that costs less per call can cost more per completed invoice if it creates rework. A more expensive model may be worth using on the difficult cases if it reduces that rework. The right answer depends on your process, not a public price list.

Define the unit before comparing tools

Pick one unit of work that matters: an invoice posted correctly, a request resolved, a document approved, or a qualified inquiry routed to its owner. Write down what counts as complete. Then count the costs attached to that unit:

• Model and software usage across every step, not just the first response.
• Integration, monitoring, and maintenance attributable to the workflow.
• Human review, corrections, and escalations.
• Retries, duplicate actions, and work reopened after an error.
• The cost of delays or mistakes when they materially affect the business.

You do not need a perfect accounting system on day one. Start with a small sample of completed and failed jobs. Mark which steps consumed money, which consumed staff time, and which produced a usable result. Compare that with the current process. The baseline matters because an AI workflow can look efficient in isolation and still be worse than the system it replaced.

Route the hard cases differently

Not every step needs the same model or any AI at all. A fixed rule can reject a missing field. A lighter model can classify a routine request. A stronger model or a person can handle conflicting evidence. Design the route around the cost of being wrong, not around using the most capable model everywhere.

Set a ceiling for retries and tool calls. If the system reaches it, stop and hand the job to a person with the context already collected. Unbounded attempts are not persistence. They are an expense with no clear owner.

The question for your next AI budget review: What does one correctly completed job cost today, and what does it cost after the system is running?

If the answer is unclear, do not add another subscription. Measure one workflow end to end first. A lower model bill is not a win if the work costs more to finish.