Skip to content
SPEIA

Artificial intelligence

Artificial intelligence in your business tools, under control.

We build AI in where it saves time — reading an invoice, sorting a document, summarising a file — and nowhere else. It proposes; the server checks; a person approves.

In short

  • In service: invoice reading by AI
  • Claude or ChatGPT, the client's choice
  • Monthly cap: warning at 80%, stop at 100%
  • Every call logged, without the document's content

Facts checked on 11 November 2026.

In service

What AI does in our tools today.

One function, delivered and configurable, rather than a list of intentions. The same foundation runs in our own management tool and in the ERP delivered to a client.

  • Reading a supplier invoice

    A scan, a photo or a PDF: AI reads the invoice and prepares the entry. Nothing is recorded until a person has checked and approved it.

    Our management tool
  • An optional function

    Without a key set by the administrator, the function is off: the tool works entirely without AI. You switch it on when you have a use for it.

    The ERP delivered to a client
  • Claude or ChatGPT: you choose

    The provider — Anthropic or OpenAI — and the key are set in the settings. The key is yours, encrypted in the database; the right to use AI is granted by role.

  • A cap, not a surprise

    The administrator sets a monthly cap. At 80% the tool warns; at 100% it stops. A file already read is never paid for twice.

  • A log of every call

    Which task, which model, how long, what cost, by whom: everything is recorded — except the document's content, of which only a fingerprint is kept.

  • The same foundation for other tasks

    Sorting an incoming document, extracting fields, summarising a file: each new task goes through the same gate, with the same safeguards. We already do this sorting on our own mail.

    Our AI agents

Safeguards

Six rules that don't move.

They are written in the code, not in a brochure: an AI function that doesn't follow them cannot be plugged in.

  1. AI proposes, the server checks, a person approves

    No figure enters the database without a check written in code — an invoice's arithmetic is recomputed — nor without a person's click.

  2. A single gate

    No screen calls a model directly. Everything goes through a gateway that checks the user's right, the activation and the cap before the call, then validates the answer against a schema.

  3. The document is data, never an instruction

    A sentence slipped into an invoice or a PDF commands nothing: instructions are fixed and versioned, and the content being read is never part of them.

  4. The model has no tools

    It reads and it answers. It cannot write to the database, send a message or call another service.

  5. No silent fallback

    Credit exhausted or service down: a clear message, not another model swapped in without your knowledge.

  6. Every change is re-tested

    Changing an instruction or a model creates a new version, replayed on a set of documents whose expected result is known.

FAQ

Frequently asked questions

Does my tool need AI?

Not necessarily. Most of what makes a business tool useful — computing, chasing, alerting, producing a document — is programmed without a model. We start there; AI comes in when content must be read or understood that a rule cannot handle.

Where do documents read by AI go?

They are sent, for the reading, to the provider you chose — Anthropic or OpenAI — with your own key: its terms apply to what it does with them. On our side, the log only keeps a fingerprint, never the content.

What does AI cost in use?

The cost depends on the provider and the volume; it is paid on your key. The tool shows the month's consumption and stops at the cap you set.

Can AI get it wrong?

Yes. That is why it only proposes: a check in code verifies what can be verified, and a person approves before anything is recorded.

What if we started from the way you work?

One first conversation is enough to know whether SPEIA can help.