The person pasting page 79 of an insurance contract into a chatbot is not defying anybody's policy. The roundup republished by Fast Company names that exact moment as the use case: you want an assistant to comb complex fine print because you are not a lawyer [6]. What teams tell themselves is that the memo covers this. What people do is paste the whole document and hope.
Weeve takes out names, addresses, company details and other identifying elements [1]. What it leaves is the substance, which for contract review is the point, because the question is usually about the clause and not the counterparty. For an HR complaint or a support transcript, identifiers are the smaller half of the problem, and the scrubber hands back a file that still says what happened to whom. There is also a cost on the return leg: once the model answers about Party A, somebody has to map placeholders back to real names, by hand, in a browser tab.
Being free and requiring no log-in [2] is what makes the step adoptable and also what makes it unprovable. No console exists, and no record that the scrub ran. So the honest home for it is the runbook next to the tab where the paste happens, not a policy PDF. Set that against the other control the same roundup mentions, which is switching off training in the chatbot's own settings [4]. That one you can check by looking at someone's screen, and it holds on the day they forget. Scrubbing earns its keep where the vendor's terms cannot be read at all: by the Cybernews analysis the roundup cites, 315 of the 500 top AI companies assessed, or 63 percent, do not disclose whether they train on user data [15][16].
For material that should not leave the machine, the local transcription options run offline [10][11], with the caveat that local models eat disk space and want at least 8GB of RAM, so a 2018 laptop may struggle [12]. Price that before standardising. Weeve's unlimited note taker is 120 euros a year, about $140 [9], while MacWhisper's premium lifetime license is 64 euros, about $75 [10], which puts the subscription at roughly 1.9 times the lifetime license in year one [17]. Basalt, which actually destroys redacted PDF text rather than drawing a box over surviving words, is $29 for a lifetime license covering up to three Macs [14], under $10 a machine [18].
Sort each recurring paste on two axes: whether the sensitive part is the identity or the substance, and whether the answer comes back into a system of record or gets read once and closed. Identity plus read-once is the Weeve case, and it is a real one. Identity plus system of record means somebody owns the re-identification step, and that is where the errors will land. Substance, either axis, means keep the file on the laptop or use an account with a contract behind it. Proton's AI Paper Trail sits outside that grid [13]; it reports what already went out, which is the question asked after an incident rather than before one.
