AI providers keep releasing cheaper models, and a newer model can cost a fraction of an earlier one. Finance teams see the price drop and expect the AI line in the budget to shrink.
It rarely shrinks the way they expect. The price of a request is the easiest number to see, so it gets the attention. The cost that decides whether AI pays is somewhere else: in the people who check the work, the cases the agent can't handle, the errors found late and the upkeep that keeps it right.
This piece is for the chief financial officer (CFO) deciding how to judge what AI costs the business.
The wrong number
A price per request tells you what one answer costs. It tells you nothing about whether the answer was used, checked, corrected or thrown away.
Take an AI agent that prepares supplier invoice matches. This is an illustration, not a client. The model's fee for each match is small.
Then a person reviews the match. Some matches go back for correction. A few reach the payment run wrong and are found at month-end. Each of those steps costs people's time, and none of it shows on the provider's invoice.
In a job like this, halving the model's price changes little, because the fee was never the main cost. Raising the share of matches accepted first time changes a great deal.
Price the job, not the request
The number worth tracking is the cost of one job done correctly. For the invoice agent, that is one invoice matched, checked and paid right. It has five parts:
- The model's fees. What the provider charges for the requests the job needs.
- People's review time. The minutes a person spends checking the agent's work.
- Exceptions. The cases the agent hands to a person because they fall outside its rules.
- Rework. The errors corrected later, and what they cost by the time someone finds them.
- Upkeep. Keeping the agent's rules, knowledge and connections current as the business changes.
Count all five, in hours and in rand, for the month before the agent starts and for each month after. Divide by the jobs done. That gives you one figure you can compare with the way the work is done now.
Read that way, a cheaper model is good news only when it does the job as well as the dearer one. If it lowers the share of work accepted first time, the saving on fees goes straight into review and rework.
Cheap requests change behaviour
Falling prices bring a second risk. When each request feels free, it is tempting to put AI into everything: one more summary, one more draft, one more automatic reply. Each addition is cheap on its own. Together they fill someone's review queue, and that person's time is the dearest part of the system.
The discipline that keeps this in check is the one you already use for staff. Every AI agent has a defined job, an owner and a measure. Work without an owner and a measure does not get added just because it is cheap to run.
What this means for the budget
Three things follow for the finance function.
- Ask for the cost per job, not the price per request. A supplier or an internal team that shows only model fees has shown you one part of the cost.
- Set the baseline before anything starts. Without the cost of the job done the current way, no later figure means anything.
- Treat a model change as a change to test. A switch to a cheaper model is a decision to check against the record of accepted work, not a saving to book in advance.
Where this argument has limits
Fees matter at high volume. For work that runs constantly with little checking, the model's fee can be the largest line. There, the price per request is worth watching closely.
Some models do the job better, not just cheaper. A newer model may handle cases the old one couldn't, which cuts exceptions. That is a reason to test new models, judged by the cost per job.
Measuring takes effort. Counting review time and rework needs a record of what the agent did and what people changed. If that record doesn't exist, building it comes first.
One question to ask
For each AI agent in your business, ask: what does one job, done right, cost us now, and what did it cost before the agent? If nobody can answer, you are watching the price list, not the cost.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets. If you want to know what your AI work costs per job, start a conversation.
Start an Agentic conversation →
Sources:
- OpenAI, GPT-6.1 Terra model documentation (developer documentation)
- OpenAI, GPT-5.6 Terra model documentation (developer documentation)
Together these support the line "a newer model can cost a fraction of an earlier one". OpenAI's documentation lists GPT-6.1 Terra at half the per-token input price of GPT-5.6 Terra, and at less than half its per-token output price. Neither page says that one model replaces the other. Both pages were checked on 2026-10-27.
