The Real Cost of AI in an Exchange — Tokens, Receipts, and OCR Economics
Every receipt OCR costs money. What a token is, why each receipt runs about 1.5 to 3 cents, and how cheap pre-classification keeps the cost of AI in currency exchange accounting software under control.
Most articles about AI talk about capabilities and dodge the uncomfortable question: how much does all this cost? For a money changer handling thousands of receipts a month, this question is not a side note — it lands directly on the profit margin.
Let's be honest and open up the economics of AI. Unlike a calculator that you buy once and is free afterward, AI costs money every time it works. Understanding this difference is the key to the right decision.
What a token is and why it matters at all
AI models break text and images into small units called tokens. Roughly every few characters is a token, and an image translates into a large number of tokens depending on its detail. The cost of AI is calculated based on the number of input and output tokens.
What does that mean? It means every time the model reads a receipt:
- The receipt photo is turned into tokens (input)
- The model returns the amount, date, and bank (output)
- You pay for the sum of these tokens
Unlike traditional software whose processing is effectively free, here every operation has a real, small cost. This "small" adds up at high volume.
The real number: how much each receipt runs
Let's not talk in the abstract. A full read of one receipt with AI OCR, as measured, costs about 1.5 to 3 cents. Now multiply this by real volume:
| Receipt volume | Cost at 2 cents per receipt |
|---|---|
| 50 receipts a day | ~$1 a day, ~$30 a month |
| 200 receipts a day | ~$4 a day, ~$120 a month |
| 500 receipts a day | ~$10 a day, ~$300 a month |
The numbers aren't astronomical, but they aren't zero either. And the important point is: if you hand every photo to the model indiscriminately, these figures grow for no reason.
The hidden trap: photos that shouldn't cost AI
In an exchange's WhatsApp group, not every photo is a receipt. A colleague's personal photo, a sticker, an unrelated screenshot, a congratulatory message. If the system hands every incoming photo to the expensive model, a large share of your cost goes to reading things that aren't documents at all.
This is where a simple trick changes the whole economics.
Cheap pre-classification: spending only where it's worth it
The solution is a cheap pre-classification stage before full extraction. Its logic is:
- A light, local check: before anything, a cheap filter determines whether the photo is "a financial document at all."
- A free skip: non-financial photos are set aside locally (with a local text-reading tool like tesseract) and never reach the expensive model — zero cost.
- Full extraction only for receipts: the expensive AI model is spent only on photos that really are financial documents.
The golden rule of AI economics: the cheapest token is the token that is never consumed. Any photo set aside before reaching the model costs zero.
With this, if half the group's photos are non-financial, half the potential cost is effectively eliminated — without a single real receipt being dropped.
Usage-based pricing and the monthly gift credit
Because AI is usage-based, its cost model must be usage-based too — not a vague fixed fee. The fair logic is:
- Each plan has a monthly gift credit; an itemized quota (a number of OCRs and lookups) that resets every month.
- As long as you are within the quota, you effectively pay nothing for AI.
- The excess is deducted from the services wallet, transparently and item by item.
This means a small exchange with a few hundred receipts a month probably stays within the quota and AI is effectively free for it; and a large exchange pays only for its real excess usage. The fairness is precisely that the cost is tied to real usage, not to a fixed number.
Transparency: know where your money goes
An important principle in AI cost is its visibility. If you don't know how many receipts were read each month and how many tokens were consumed, you can't manage it.
That is why a token usage report must be available — which in Nexto is trackable in the panel. You see how many OCRs were done, how much of the quota is left, and how much the excess was. AI cost becomes controllable when it becomes visible.
How to keep your AI cost low
A practical summary for any exchange:
- Let pre-classification do its job: non-financial photos should not cost the model.
- Use caching: repeating the same lookup (for example the same IBAN) should not cost again.
- Monitor usage: look at the token report monthly to catch an unusual spike early.
- Know your quota: know how much your monthly gift credit is and when you reach the excess.
These four things are the difference between AI that is cost-effective and AI that quietly eats out of your pocket. If you want to get familiar with the applications first, WhatsApp Receipt OCR is a good starting point.
Conclusion
AI is not free — every receipt OCR runs about 1.5 to 3 cents, because it works on token consumption. But this cost is controllable with two levers: cheap pre-classification that sets non-financial photos aside for free, and a usage-based model with a monthly gift credit that ties cost to real usage. An honest currency exchange accounting software, instead of hiding these figures, shows them to you transparently.
Want to see how token usage and AI cost control work in practice? Build a dedicated demo with sample data and see the usage report up close.
See all of this inside Nexto
A complete, private instance with sample data — no install, no credit card.