Confidentiality

Give an AI model a sentence with every name redacted and the sense of the original has gone: it has a structure with holes in it, and it will fill them with guesses. Text that is placeholder after placeholder fares little better. So the names you protect are replaced by stand-ins that read like the real thing — a person of the same gender, an organisation of the same kind, in the right form for the target language — and figures by believable ones, scaled consistently. The AI model still reads a document that makes sense, and translates it as one.

Each name in your profile gets a stand-in, with its form in the target language too, so the translation reads correctly when the real name goes back in. By default, entries shorter than five letters or without a capital are left as they are, and you can change both for a project. The profile editor counts any name that has no stand-in yet, so you can see which ones. Email addresses, phone numbers and the formats you write rules for carry nothing the translation needs, so a rule finds them, with no model involved, and a numbered placeholder goes in their place. All of it goes back into the translation, and a segment where something did not come back cleanly is flagged for you to check, never quietly repaired.

Point AV at files you have already translated, or at the ones in front of you, and an AI model you choose finds the names in them and proposes a stand-in for each; you go through the list in the profile editor, and a name you mark as not personal data stays out when you run it again on new files. Or start an empty profile and type the names yourself. For another target language, copy the profile, and AV tells you which names need a form of their own.

A clock time, and a reference like IFRS 16 or a clause number, carry nothing a client needs hidden, so they go as they are; so do whole numbers below a limit you set in the figures profile — a count, a day of the month. Place names can stay in the clear down to the level you choose for the project, countries but not cities, say, so the text keeps its geography without giving away an address.

Copy a section, a table or the whole document exactly as an AI model would see it, with every name, figure and pattern replaced. Or export the source file itself, a Word or PowerPoint document, the same way, in Word with every replaced word highlighted, and show it to your client before anything goes anywhere. A copy is for reading: what you paste back keeps its stand-ins.

Translation, terminology, memory alignment with an AI model, code placement, match fixing, the agent console — each one, from rules you configure once. Protection that covers only the pathway somebody remembered is not protection.

Each of those pathways audits the request it actually built, against every entry it was supposed to remove, and finds every hit: which entry, and where in the request it sat.

A hit does not become a note you find later. The send is refused, and the message names the terms that were still in the request and where each one sat, and says that nothing was sent to the AI model. There is no setting that turns this into a warning, and no pathway it skips — the agent console and code placement included. Before a segment goes out for its codes to be placed, a formatting marker that falls inside a protected name is moved out of it, and the request is read a second time without the markers, so a name split by bold type cannot slip past.

And it refuses the quiet failures as firmly as the loud one. A profile you named that will not load, a profile that loads but protects nothing in this document, a profile that belongs to another project, a pattern or number profile the project names and cannot find: each one stops the send with a message saying which, and nothing leaves. Sending without stand-ins is something you decide, in the project's settings. It never happens because a profile went missing.

Records that would hold your client's text are off unless you turn them on. When AV refuses a send or finds a protected name in a reply, it tells you on screen; it keeps a record of what happened, in the project's own folder, only if you have chosen that in Settings, and first-time setup asks, with off as the safest answer. The capture tools we use to test table translation and code placement are not in the released application at all. And the support bundle you send us never carries any of it.

Features that need to remember their work keep working files: name-detection runs you can undo, the wizards' results, and a cache of the vision model's structure check. That cache sits outside the project folder.