Search across your documents
Ask a question in plain language, get the passage and the file it came from. No more hunting through shared drives.
Artificial intelligence
AI solutions for business, on your own data
One use case at a time, built on your documents and measured against real work. Developed in Paris, connected to the tools your teams already use.
2008 — year we started
Paris 8e — team on site
Source code — delivered, it is yours
The problem
Start with a task, not with a technology
Most AI projects fail in the same way: a general assistant is bought for everyone, it knows nothing about the company, and three months later nobody opens it. The useful version is narrower. One task that people actually repeat, done on your own documents, with an answer you can check against a source rather than trust on faith.
We start by picking that task with you, and by writing down how we will know it works: time saved per file, error rate, share of answers with a citation. Then we build the smallest thing that does it, connected to your own content — procedures, contracts, tickets, invoices, product data — and we measure it on real cases before anyone deploys it to a whole team.
The technical choices follow the task. Sometimes a hosted model is right; sometimes an open-source model running on a server you control is better, because the documents cannot leave. Where the data is stored, who can query it and what is logged are decided during scoping, in writing, along with the cases the system must refuse to answer.
Use cases
What we build most often
Each one ships on its own. You judge the first before deciding on the second.
Document extraction
Invoices, delivery notes, contracts and forms turned into structured fields your software can read, with a human check on anything uncertain.
Assistants for a team
Support, sales or back office assistants that answer from your procedures and hand over to a person when they should.
Wired into your tools
Results land in your CRM, ERP or ticketing system through an API, instead of in a chat window someone has to copy from.
Where your data sits
Hosted or self-hosted, in Europe or on your own server. Retention, access and logging agreed before the build starts.
Measured quality
A test set from your real cases, a score before and after each change, and a log of every answer for review.
How we work
How an AI project runs
Four stages, in this order. You sign off each one before the next starts.
- 1
Pick the use case
We look for a repeated task with a measurable outcome, and we say plainly when a plain script would do the job better.
1 week - 2
Prototype on your data
A working version on a sample of your real documents, scored against a test set your team helps build.
2 to 4 weeks - 3
Build and integrate
The interface, the connections to your systems, the permissions, the logging and the fallback when the model is unsure.
4 to 8 weeks - 4
Deploy and monitor
Gradual rollout, training, then monthly review of the logs and the score. Models change; the tests catch it.
Ongoing
Examples
Things we have built or prototyped
Not a menu to choose from, but a sense of the size of a first project:
Internal assistant over your documentsInvoice and document extractionInbound email triageSemantic searchMeeting summariesTranslation and rewritingModel hosted in EuropeSelf-hosted open-source modelConnection to your CRM or ERPREST API and webhooksAnswer logging and reviewQuality dashboard
Example project
A concrete example
Example — to be replaced with a real project
An internal assistant over twelve years of procedures
Procedures lived in eight hundred documents across three shared drives, and new staff asked a colleague rather than search. We built an assistant that answers from those files and shows the source paragraph, with the drive as the single place anything is updated.
9 min → 40 s
to find a procedure
100 %
answers with a cited source
7 weeks
from prototype to production
Budget
What a first AI use case costs
We price one use case at a time. The figure depends on your data and on the integrations.
First use case — from
6 000 €
Indicative figure for a standard scope. Every quote is issued as a fixed price after the scoping phase, with no per-seat fee.
- Use case selection and test set
- Prototype scored on your data
- Integration with your tools
- Source code delivered, no lock-in
How it is priced
We start from the use, not the technology
An assistant, an automation or a document pipeline has no list price: it depends on your data, the tools to connect and how much control you need over the answers. We scope, then we price.
Custom quote
On quote
- Scoping of the real use case
- Assistants, automation, document processing
- Integration with your existing tools
- Firm fixed price after the scoping phase
- No quantified performance promise
We do not put a figure on accuracy, time saved or return on investment before we have seen your data. Whatever can be measured is measured during scoping.
Once it is running
An AI system needs watching
Integrations change, the way people use it shifts, answers drift. The follow-up is priced with the project.
AI Care
On quote
No commitment
- System monitoring
- Integration maintenance
- Workflow adjustments
- Technical support
- Priced together with your project
FAQ
What people ask us most
Will it invent answers?
A model can, which is why we build so that answers come from your documents and arrive with the source passage attached. The system is set to say it does not know rather than guess, and every answer is logged so you can review what it actually said and correct the content behind it.
Does our data leave the company?
That is your decision, taken at scoping and written into the specification. We can run an open-source model on a server you control, so nothing leaves, or use a hosted model in Europe when that is acceptable to you. Retention, access rights and logging are agreed before any build starts.
How do we know whether it works?
We build a test set from your real cases before writing anything, and we score against it after every change. You see the same numbers we do: accuracy, share of answers with a citation, and time on a task. If the score does not move, we say so instead of shipping.
Is our company too small for this?
Size matters less than repetition. If a task is done fifty times a week on documents you already own, a narrow use case pays for itself quickly. If it happens twice a month, it usually does not, and we would rather tell you that in the first call than sell you a pilot.
See also
Related services
Let's talk
Bring us one task that eats your week
Thirty minutes is enough to say whether it is a good first AI use case, and to give you an honest range.