AI in your business.Not in your secrets.
You want AI to read along in your email and files, but not in everything. We build a boundary that does not depend on a good instruction: confidential clients, files or folders are technically shielded, and every attempt is logged.
It ran at our place first.
At our place dozens of AI tasks read email, calendar and notes every day. When we took on clients who work with us under confidentiality, we could not rely on 'the AI knows that'. So those domains are on a list: the mail tools show no content from those senders, every attempt to write about them is blocked before it happens, and the work for those clients runs in a separate, secured AI environment where data security is contractually guaranteed.
Does this sound familiar?
- 01You want AI on your mailbox, but some clients or files must absolutely stay out of it.
- 02An instruction like 'do not use this' is a request, not a lock.
- 03After a data breach you need to be able to show what the system did and did not see.
- 04Staff paste client data into their own ChatGPT or Claude account, where that data is not protected by default.
What you get.
- A list of shielded clients, domains or folders that you manage yourself
- Email and documents from that source are unreadable to AI, not merely 'do not use'
- A block before writing, not a check afterwards
- A log of what was blocked
- For confidential work a separate, secured AI environment with contractual guarantees on data security, or a model on your own server
Operating it yourself says more than a description.
Let the assistant summarise an inbox in which one sender is shielded, then try to make a note about that client. You see what the AI never gets to see and what lands in the log. This is a demo on sample data, not client work.
Demo · prepared in advance · illustrative · no live model
Your inbox (fictional)
Delivery week 41 moved
planning@supplier-example.example
Draft VAT return for review
accounts@firm-example.example
Proposal and ratesshielded
info@klant-onder-nda.example
Question about the new quote
sanne@bakery-example.example
This week in the trade magazine
newsletter@trade-magazine.example
Shielded
Clients, domains and folders that you put on this list yourself. The AI technically cannot get into them.
- klant-onder-nda.exampledomain
- legal-example.exampledomain
- /Clients/Confidential/folder
Block log
No blocks yet in this session.
Request versus lock
An instruction is a request. A boundary in the technology is a lock.
A model that is asked not to use something can slip up. A mail that is technically never passed on cannot end up in there.
Confidential work runs through a separate, secured AI system where data safety is guaranteed by contract. That is different from a standard personal ChatGPT or Claude account.
Illustrative demonstration with examples prepared in advance. No live AI call and no API key needed; the answers are fixed and contain no invented figures or names.
Four steps, in this order.
We sit in
We map out which clients, files and folders are confidential, and what AI can already reach today.
We build alongside
The first version runs separately from your systems and writes nothing into them. If you stop, nothing has changed.
You keep the last step
Whatever the system produces is first ready as a draft. Only when you give the go-ahead does anything leave the building.
We connect only once it works
Once it proves right in practice, and only then, we connect it to your systems for good.
Lead time: one to three weeks for a working first version.
Shield list · Block before writing · Log · Business AI environment · Open models on your own server
What people ask us about this.
Q1Why is my own ChatGPT or Claude account not enough?
With a standard personal account it is not settled by default what happens to your data: conversations may be stored and, depending on the settings, used to improve models, and there is no data processing agreement. A business environment guarantees data security in a contract. For confidential work that is the minimum.
Q2Isn't an instruction to the AI enough?
An instruction is a request. The model tries to follow it, but guarantees nothing. We shield the source technically: the AI never sees the content, and an attempt to write about it is stopped before anything is stored.
Q3Can I decide myself what is on the list?
Yes. You manage the list of shielded clients, domains and folders yourself. Every block goes into a log, so afterwards you can show what the system did and did not see.