AI economics · October 2026
The Unit Economics of AI in Payments and Finance Operations
What an AI model really costs per case once checking, mistakes and fixed costs are counted, with figures from ScalePoint Research's model tests.
- The model's price is usually the smallest part of what AI costs per case. Checking time and the mistakes that get through are bigger.
- How accurate a model is decides how it has to be run: as is, with spot checks or with a person checking every case. That moves the full cost far more than the model's price does.
- In Test 1, a model at 4.6 cents a statement came out at about 5 cents all in. A model at 1.2 cents came out at about $14, because a person would need to check every statement.
- Agents call the model once per step and re-read the case each time, so a weaker model that takes more steps or starts again costs more per case.
- A cheap model is only cheap per case when it's accurate enough on that task. Hence a firm should test models on its own cases and measure the result in money.
When a firm puts an AI model into a payments or finance process, the first number it sees is the model's price, quoted per million tokens. That number is easy to compare, and it often decides the choice. It's also usually the smallest part of what the process costs per case. This article sets out what goes into the full cost per case, using figures from the ScalePoint Research model tests, and what changes when the model works as an agent. All figures are in US dollars.
The four parts of cost per case
- Model fees. What the provider charges for the text the model reads and writes, across every call it makes on the case.
- Checking. The time a person spends reviewing the model's work. How much checking is needed depends on how often the model gets things wrong.
- Mistakes that get through. The money lost when a wrong answer isn't caught, such as a fee overcharge never reclaimed or a payment booked to the wrong company.
- Fixed costs. Building the process, keeping a test set, monitoring results and the time of whoever owns the model, spread across the cases it handles.
The full cost per case is the sum of the four. The model fees are the only part on the provider's invoice, which is why they get the most attention.
Model fees
Models are priced per token, roughly three quarters of a word, with one price for text sent in and a higher one, usually, for text written back. Some models also write working notes before they answer, called thinking, and those are billed as output. Hence two models with similar list prices can cost very different amounts per case. The Hosting and terms page lists the prices for every model tested.
For one-step work, the model is called once per case. In Test 2, reconciling a month of card and payout settlements in one call cost between 4 cents and 90 cents a run, depending on the model.
An agent calls the model once for each step, and every call reads the case so far, so each step reads more text than the one before. A case that takes ten steps can cost much more than ten single calls. A weaker model that takes extra steps, or has to start the case again when its answer doesn't add up, adds to that. The reconciliation break example shows where those calls come from.
Checking sets the cost
How often a model gets things wrong decides how a firm can run it. In Test 1, reading UK card acquiring statements, ScalePoint Research used three levels:
- Use as is. 0.1% of fees wrong or less. No checks.
- Spot checks. Up to 0.5% wrong. A person checks one statement in ten.
- Full review. Up to 2% wrong. A person checks every statement. Above 2% the model isn't usable for the task.
The full cost per statement then adds the model's price, the checking time at an assumed $55 an hour, and the fees still wrong after checking, of which 10% are assumed to become a real loss. On those assumptions, at 10,000 statements a month:
| Model | Model price per statement | How it has to be run | Full cost per statement | Full cost a year |
|---|---|---|---|---|
| Grok 4.7 | 4.6 cents | Use as is | $0.05 | $5,500 |
| Claude Opus 5.5 | 15.5 cents | Use as is | $0.16 | $18,600 |
| Gemini 3.8 Flash | 2.6 cents | Spot checks | $5.00 | $600,000 |
| GLM 5.3 | 1.4 cents | Spot checks | $6.10 | $731,000 |
| Mistral Large 3 | 1.2 cents | Full review | $14.36 | $1.72 million |
| GPT-5.5 | 13.9 cents | Full review | $14.87 | $1.78 million |
| A person, no software | 45 minutes a statement | $41.25 | $4.95 million |
Mistral Large 3 is the cheapest model in the table to call, and on these assumptions one of the dearest to run, because every statement would need a person to check it. Grok 4.7 costs nearly four times as much per call and comes out at about 5 cents a statement all in. The assumptions are set out in the Test 1 paper, and its full cost calculator lets a reader put in their own volumes, staff cost and loss rate.
Mistakes that get through
Mistakes can cost far more than the model fees saved. In Test 2, nine models reconciled a month of files for an invented merchant group, with 17 planted errors worth $4,832. On the larger version of the test, DeepSeek V4.1 Flash saved 86 cents a run against Claude Opus 5.5 and missed about $1,400 more. Opus 5.5 missed about $13 a run on average.
Averages also hide the worst runs. Gemini 3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash and Gemini 3.8 Flash each had a run that missed about $2,600 or more. A finance team needs to know what a bad month looks like, which is why the Test 2 leaderboard shows each model's worst run next to its average.
The answer depends on the task
In Test 1, cheap models came close to the best: Gemini 3.8 Flash and GLM 5.3 got under 0.5% of fees wrong at 1 to 3 cents a statement. In Test 2, most cheaper models fell well behind. Reading a statement is mostly extraction and classification, while reconciliation asks a model to reason across several files at once. Hence a price gap measured on one task shouldn't be assumed to hold on another. Nine skills for testing AI models in payments sets out which skills each task needs.
Fixed costs and volume
Some costs don't change with volume: building the tools and the instructions, keeping a set of past cases with known answers, running the checks each month, and the time of whoever owns the model in the team. At low volumes these fixed costs dominate, and a cheap, simple set-up can be the right one. At high volumes the per-case costs dominate, and a few cents of model fees or a few tenths of a percent of accuracy add up to real money, as the table shows.
Price, hosting and switching
Model prices fall fast. Epoch AI found that the price of reaching a given level of performance fell by between 9 and 900 times a year, depending on the task. A model that's too expensive per case today may not be in a year, but that only helps a firm that can switch. A firm that keeps its own test set and instructions can rerun the test on a new model and compare the result in money before it switches. What changes when an AI agent works the queue covers what to keep in-house.
Where the model runs matters too. Some firms can only send customer or transaction data to a host with zero data retention, or to servers in the EU, and some prefer models with open weights that they can run on their own servers. Those rules narrow the list of models a firm can use before price comes into it. The Hosting and terms page shows where each tested model ran and on what terms.
What to ask before choosing a model
- What does it cost per case at our volumes, including checking and the mistakes that get through?
- How often does it get things wrong on our own cases, measured in money?
- Does that accuracy let us run it as is, or does it need spot checks or a full review?
- For an agent, how many model calls does a typical case take, and what happens to a case it can't finish?
- Can we switch to another model later, and do we keep the test set and the instructions?
- Where does our data go, and on what terms?
ScalePoint's AI Economics engagement works through these numbers for a firm's own processes, and ScalePoint Research runs private evaluations on a firm's own cases. The terms used here are explained in the glossary.