Not “stop AI”
Banning the tools moves the work to personal phones and removes the only visibility you had. The goal is safe use, not no use.
The name
itzal
it‑SAHL /iˈtsal/ Basque
- Meaning
- Shadow; shade. Itzala — the shadow.
- The part people get wrong
- The tz is a ts sound, as in cats — not a z. It is “it-SAHL”, never “IT-zull”.
- Where it comes from
- Euskara, the Basque language — spoken across the western Pyrenees and unrelated to any other living language.
Basque stress varies by dialect and is not strongly phonemic; the second-syllable stress above follows standard euskara batua. If you are recording anything, have a native speaker say it.
Why that word
The record is the shadow of the message.
When one of your people pastes a customer record into an AI tool, what reaches your security console is not the message. It is the outline of it: what kind of data was in there, how much, which rule it broke, and what was done about it.
Enough to act on. Nothing to leak. A shadow is cast by the thing, follows it exactly, and carries none of it away — and we could not find a more accurate description of a redacted event record than that.
It is also, we will admit, a good name for a product whose entire claim is that it sees without taking.
The objective
One thing, stated so you can hold us to it.
Most security companies describe an ambition broad enough that no outcome could ever contradict it. Here is ours, narrow enough to fail:
Make it possible to use AI at work without anyone having to read what your people write.
Not “watch employees”
A control that requires surveillance to function fails the works council, the DPO, and eventually the people it monitors.
Not “trust us with it”
A vendor asking to hold your sensitive data so it can protect your sensitive data has moved the risk, not removed it.
Just this
Detect on the device, act before the send, and report a shape rather than a substance. Everything else is downstream of that.
The trade-off
We took the harder path, and it costs us.
A DLP tool for AI usage has an obvious shortcut: send the prompt somewhere with a big model and let it decide. It works, it is easier, and it means the content you were trying to protect now sits in a second vendor’s systems with its own retention window and its own breach surface.
We chose detection on the endpoint, redaction before the record exists, and no mechanism to retrieve the original text — because a product that protects data by collecting more of it is solving the wrong problem.
That constraint is not free. Tuning is slower, because we cannot look at what we got wrong. Some surfaces stay uncovered longer. No heavy model fits inside an endpoint budget. The full trade-off, stated plainly.
Commitments
Four things we have decided not to do.
Each one would make the product easier to build or easier to sell. They are written down here so that breaking one is visible rather than gradual.
Not now, not later
- No “send us the sample” button. The text that caused a false positive stays on the device, even though having it would make detection better.
- No invented social proof. No customer logos, testimonials or counts until they are real and permitted.
- No quiet coverage gaps. What Itzal cannot see is published, including the parts that lose deals.
- No third-party routing. Nothing on this site — fonts, analytics, forms — sends your visit to anyone else.