IN SHORT
- The AI that builds your system and the AI that runs inside it are separate things with separate costs.
- Inside a system it is good at reading: pulling fields out of an invoice, turning a photograph into a description, sorting incoming email, summarising long notes.
- The rule that makes it safe is the same everywhere: it proposes, a person confirms. Never let it be the last step before money moves.
- It is metered per use, on the same credit balance, and it stops answering when credits run out while the rest of the system carries on.
Two different AIs, and why the difference matters
One AI takes your description and builds software. A separate one lives inside the finished software and does a job every day, such as reading an invoice. They are billed differently and they fail differently.
Almost every confusion in this area comes from treating them as one thing. The building AI is used in bursts, while you are working, and then stops. The AI inside your system runs whenever the system runs, once per document or per message, for as long as the system is in use.
That means a system with AI inside it has a small ongoing cost that a system without one does not. It is usually tiny compared to the typing it replaces, and it is worth knowing about before you design around it.
What it is genuinely good at inside a business system
Reading unstructured things and turning them into structured ones. Fields out of a document, a category for an incoming message, a summary of a long note, a draft of a reply, a search that understands a question rather than matching words.
Notice what these have in common: a person still decides. The model does the reading, which is the boring part, and hands a human a filled-in form to approve.
| The job | What goes in | What comes out |
|---|---|---|
| Supplier invoices | A PDF or a photograph | Supplier, number, date, net, VAT, total, as fields |
| Site photographs | Pictures from a phone | A draft description attached to the job card |
| Incoming email | The message | A category, a priority and the record it belongs to |
| Long meeting notes | Free text | A summary and a list of what was agreed |
| Searching your own records | A question in Bulgarian | The records that answer it, not the ones containing the word |
The rule that keeps it safe
It proposes and a person confirms. Never put a model at the last step before money moves, a document is filed, or something is sent outside the company.
This is not caution for its own sake. Models are wrong occasionally and confidently, which is a worse combination than being wrong obviously. An invoice read with the decimal point in the wrong place looks exactly as convincing as one read correctly.
Structurally the fix is simple and costs nothing: the extracted fields appear on a screen with the original document beside them, and somebody presses confirm. That takes four seconds instead of four minutes of typing, keeps the benefit and removes the risk. It is also the shape the AI Act asks for in Article 14, so the safe design and the compliant design are the same design.
What it costs
Each call is metered in credits from the same balance the building uses. Reading a document is a small number of credits; the exact figure depends on the size of the document and the model.
Because it is metered per use rather than per seat, the cost follows volume honestly. A practice processing thirty invoices a month spends thirty documents worth. Nothing is charged for the months when it is quiet.
When credits run out, the AI features stop answering and everything else carries on: screens, logins, records and scheduled jobs. That is deliberate, so an empty balance never takes a business offline, and the warning arrives before it happens.
Where the data goes
The platform, the database and what you build sit in the European Union, in Frankfurt. The exception is the model processing itself, which may happen outside the EEA under the European Commission's standard contractual clauses.
Data the AI never sees never leaves. That sentence is the practical one: if a field is not part of what you send for processing, it is not part of the question. Designing around that is straightforward - send the invoice, not the whole customer history.
The platform keeps a per-project record of AI use: which provider, which model, how many requests, how many tokens, what it cost, day by day. If you need an extract for an internal audit or because a client is asking, ask and we will produce it.
What it is bad at, said plainly
Being reliably right about numbers with legal consequences, understanding handwriting nobody can read, and knowing your company's private conventions unless you have told it.
Do not put it in charge of deciding whether an invoice may be paid, whether a contract clause is acceptable, or whether someone qualifies for something. Those are decisions with consequences, and the correct design is that the model prepares the information and a person decides.
It is also worth being honest about the cases where the old way wins. If four fields need typing and they are always in the same place on the same form, typing them is faster than reviewing an extraction. Use it where the input is messy, not where it is already tidy.
QUESTIONS
Do I need AI inside my system at all?
Often not. Most of the value in a first system comes from records, roles, reminders and reports, none of which involve a model. Add it when there is a specific pile of documents or messages somebody is retyping.
Which model does it use?
The platform picks a suitable model for the job, and the usage record shows which provider and model answered each request. You are not tied to one vendor by the description you wrote.
Is my data used to train anything?
No. What you build and what you put in it are yours; it is not read, extracted or used for training. The security page states the full position, including the one exception about where model processing physically happens.
Can it read Bulgarian documents?
Yes, including scanned ones and photographs, though quality follows the picture. A flat, well-lit photograph of an invoice reads reliably; a crumpled one at an angle does not.
RELATED
Documents, photographs and AI inside - the tutorial, with a working example
Security - where the data sits and what leaves the EU
The AI Act - which duties are ours and which are yours
The automations that pay first - how to decide whether this one is worth doing