Suppose your client calls in six months with one question: can you demonstrate that our data is gone?
Not: did you delete it. Not: is it in the contract. But: can you demonstrate it.
Most professionals have no answer to that question. Not because they are careless, but because the tools they use do not provide the answer. And that is exactly the kind of problem that only becomes visible at the moment it is too late to do anything about it.
What the market promises today
Take a look at the frequently asked questions of any large AI platform. The answer to the deletion question reads roughly like this: you can request deletion at any time, permanent deletion can take up to ninety days, and you will receive a confirmation afterwards.
That is neatly worded, and it is probably true. It is also the most a cloud vendor can offer. But look closely at the closing piece: you will receive a confirmation.
A confirmation is a statement. Someone tells you that something has happened. Whether that is true, you cannot check — you have to believe it.
For a pizza delivery, that is fine. For a file covered by confidentiality, it is something else.
Why confirming is not the same as proving
There are three reasons why those ninety days and that confirmation are worth less in practice than they sound.
The vendor is judge in its own case. The party that performs the deletion is the same party that declares it succeeded. There is no independent check in between. That is not an accusation — it is simply how the arrangement is built.
The copies nobody mentions. Your document did not stay one file when you uploaded it. A text version was made, that version was cut into pieces, those pieces were turned into strings of numbers for the search function, a reference sits in a search index, a chat history hangs around it, and back-ups exist of all of the above. If the deletion promise is about “your file”, it may not cover its six derivatives.
Nothing is left to show later. Even if everything was neatly erased, two years from now you hold nothing. No document, no verifiable trail. Only an email from a vendor that may no longer exist.
What the law actually asks
The GDPR includes the right to erasure, and that right is well known. Less well known is the duty that surrounds it: you must be able to demonstrate compliance with the principles. Demonstrating is a different verb than doing.
For you, that means something uncomfortable. You are responsible for data you entrusted to an AI vendor, even when that vendor performs the deletion. Your client holds you to account — not them.
This article is general information, not legal advice. For your own situation, put it to a lawyer or your data protection officer.
The two ways it can be done
There are only two routes by which “gone” is also demonstrably gone.
Route one: make sure it was never there.
What was never sent to an external party does not need to be deleted there either. If your documents stay on your own machine, the deletion question suddenly becomes simple: you delete them, and you were the only one who ever had access. There is no vendor that has to confirm anything, because there is no vendor involved.
That sounds almost too simple, and yet it is the only route that solves the question structurally rather than contractually.
Route two: make erasure mathematically demonstrable.
Sometimes something does need to be kept — a log, a case-file structure, an accountability trail. For that there is a technique that has long existed in the security world but rarely shows up in AI tooling: cryptographic erasure.
The idea is surprisingly simple. You store the data encrypted, and every person or every case file gets its own key. To erase, you destroy the key. What remains is unreadable — not hard to read, but mathematically unreadable, including for yourself.
The benefit lies in what does remain: the trail that something existed, and that at a certain moment it was made unreachable. So you can deliver two things at once that normally clash — the data is gone, and you can demonstrate when that happened and who did it.
Seven questions for your vendor
Ask these questions before client data goes into an AI tool. They are short, and the way they are answered often says more than the answer itself.
- Where is my data physically stored, and which other parties get to see it?
- What exactly happens on a deletion request — which derivatives are included?
- Are search indexes, strings of numbers (embeddings) and chat history erased as well?
- How long do back-ups exist, and are they covered by the deletion?
- Do I get something I can still use to demonstrate it two years from now?
- Can someone other than you verify that it happened?
- And the most important one: why should I believe you?
Question seven is not a provocation. It is the question your client will soon ask you.
Curious where you stand today? Take the AI scan — five minutes, in your browser, nothing is sent.
Where we stand
We are building Provenize because we believe question seven deserves an answer that does not consist of trust.
Provenize is an AI memory layer that runs on your own machine. Your documents do not come to us — there is nothing at our end to delete. What there is, is a log in which every action is recorded and every line is locked to the previous one, so a silent change stands out afterwards. Erasure works through key destruction: the content becomes unreadable, while the trail that something was there — and when it was erased — remains verifiable.
And an outsider can recheck that trail themselves. Not with a statement from us attached, but with a check they run themselves.
To be honest about it: Provenize Desktop has not been released yet — the release is planned for late 2026. The product is designed to support you in meeting your obligations. It is not a seal of approval, not a certificate and not a guarantee; we deliberately do not use those words.
Want to think along before it is finished? We are looking for a limited number of firms and teams that use and shape the product before release. It costs nothing. What we ask in return is honest feedback: info@provenize.ai.
Working in law or compliance? Also read how the file stays in-house, or compare cloud AI and local AI honestly side by side.
v0.1 · September 2026 · Back to the blog overview