Ask your system in plain language, and when AI proposes an action you are the one who confirms. AI does not sign. You do
For SMEs and software houses that want artificial intelligence inside the system they already use, on real data, without hallucinations presented as facts and without data leaving the company.
Do you recognise at least two of these?
- To know "how many of customer X's orders are late" someone has to open three screens and a spreadsheet.
- Replies to customers about delivery times are written by hand, looking up the same data every time.
- You tried a generic chatbot: it answered well on everything except your data, and once it invented a delivery date.
- Someone proposed a "copilot" that acts on its own, and the question left with you was: if it gets it wrong, who is responsible?
- The system holds ten years of data and nobody can ask it a simple question.
If you said yes at least twice, the problem is not AI: it is where it sits and who decides.
What does an answer that takes three screens really cost?
Do the sums on your company. Demo estimates, to be verified on your data.
- An internal question that takes three screens and a spreadsheet costs 10 minutes: ten a day is more than 400 hours a year of people paid to decide, not to search.
- A delivery date invented by a chatbot is worth a customer who does not call back.
- A wrong automatic action (a discount, an order, a cancellation) with no trace of who decided it is worth an endless argument and, in an audit, a finding.
What is an AI assistant on company data with human confirmation?
A language model connected to your system through a standard protocol (MCP, Model Context Protocol) that gives it two things: read tools on real data (orders, customers, batches, jobs) and action tools that do not execute: they propose. The answer cites the data it comes from; the action appears as a proposal, and a person confirms or discards it. Every confirmation stays on record with name, date and what changed. It is the same principle this website is written with: AI proposes, a person confirms, and it stays on record who did.
How does it work, in four steps?
- Ask in plain language, from the chat inside the system, from WhatsApp or with a voice message: "which orders this week have no delivery note?".
- The assistant queries real data with the tools you gave it, not with what it "remembers": the answer comes with references (order, customer, date).
- If an action is needed ("tell the customer we ship on Thursday"), it prepares it and queues it as a proposal.
- A person confirms with one click, corrects or discards. From that moment the action is theirs, and it stays on record who.
Why inside your system, and not a separate chatbot?
- No invented answers presented as facts: the assistant answers only with the data it read, and says so when it cannot find them. That is how hallucinations are fought, not with a disclaimer in small print.
- Human confirmation on 100% of actions, by construction: there is no "automatic" switch to leave on by mistake.
- Data stays in the company: the model can run on your server, or use an external provider only for the text, never for raw data. You decide, in writing.
- Born from modules in production: assistant with confirmation, automatic document reading, voice-dictated work reports are among the 62 ready modules, adaptable in about two weeks.
What do you need in the company?
A system with readable data (database, APIs or exports), a server or hosting for the module, and the five questions you ask by hand most often today: they are the starting point of the trial. No dedicated hardware. For a software house: I join your repository, the module is integrated into your Laravel code with tests and documentation, and it stays yours.
How much does it cost and how long does it take?
First we measure, then we build, and the price comes after the value.
- 30 free minutes on your documents or data: you get a document with today's numbers, yours even if you stop there.
- Clickable mockup in 15-20 days, from 1,500 to 3,000 €, fully deducted from the total if you go ahead.
- Written quote after the mockup, paid in stages with a working demo every two weeks. Your code, on your server, with tests and documentation.
By the day, for teams that want the work inside their own repository: from 850 to 1,100 € a day for applied AI and industrial systems, from 600 to 800 € for a senior Laravel developer, from 700 to 900 € for application security. No per-document fee, no per-user fee. The full price list.
The next step is small: 30 free minutes with ten documents or one month of real data. If it makes sense, a clickable mockup you see before investing. Starts are scheduled one at a time.