Illustrative scenarios, details changed
Four companies. One method.
We never name clients or share their solutions. These scenarios show how the method plays out, step by step.
We never name our clients or share their solutions. These four scenarios are illustrative: industries are real, details are changed, the method is exactly how we work.
Case 01 · Manufacturing
Keeping the knowledge of senior service engineers
Equipment maker, several hundred employees, field service across Europe.
- 01 Audit
Scattered manuals, service reports and personal notes mapped. We marked what must never leave the company.
- 02 Pilot
Search across the documentation on one server with an open model. Every answer links to the page of the manual.
- 03 Model
A compact model trained on the history of service tickets suggests likely causes from a description of symptoms.
- 04 Machine
Ticket, diagnosis, draft spare parts order and customer report connected into one line. An engineer confirms every step.
What changed
Junior engineers work with the experience of senior ones. The know how became a company asset instead of data in someone else's cloud.
Case 02 · Financial services
Checking client documents without the cloud
Regulated financial company, mid sized, under DORA and GDPR.
- 01 Audit
Back office mapped, data classified, approach agreed with the data protection officer.
- 02 Pilot
A model recognises the document type and flags what is missing. Personal data is masked before any processing, and every decision is logged.
- 03 Model
A small classifier trained on the company's own forms runs without expensive GPUs.
- 04 Machine
Documents, check, risk signal, specialist queue. A person always makes the decision.
What changed
Specialists spend their time on difficult cases instead of sorting. Every AI action is visible in the log for the regulator.
Case 03 · Legal
Partners' experience available to the whole team
Law firm with several dozen lawyers, professional secrecy, strict walls between client matters.
- 01 Audit
Access rights per matter reviewed, risks of leakage between clients assessed.
- 02 Pilot
Search across the archive that respects permissions: a lawyer only sees the matters they are cleared for.
- 03 Model
A model trained on partners' markups flags risky clauses in the style of the firm.
- 04 Machine
Incoming contract, risk markup, precedents, draft changes, review by a lawyer.
What changed
Partners' experience is available to the whole team. Client data never leaves the firm and never mixes between matters.
Case 04 · Software company
Launching an AI assistant without making the news
SaaS company whose AI assistant is used by its customers and can read their data.
- 01 Audit
The agent's tools and permissions mapped against the OWASP Top 10 for LLM Applications.
- 02 Attack
A set of attacks: instructions hidden in documents, system prompt leakage, access to another customer's data. A report with priorities.
- 03 Protect
Input and output filters, rules for what the agent may do, and the same attacks run again.
- 04 Watch
Live monitoring, new attacks every month, reports the company can show its own customers.
What changed
Vulnerabilities were closed before launch, and the company can prove its assistant is protected to customers and auditors.
Private AI without leaks starts with an audit
In two weeks we show where data already leaves through external services and propose the architecture for your perimeter.
- First meeting and scenario review free of charge
- You keep the audit report whatever you decide
- NDA signed before any access to your data