Shadow AI enterprise risk is not a line in next year's budget. It is happening on the laptops of your people, today, in the hours between meetings. An analyst pastes a client memo into a public chatbot to summarise it. A junior associate uploads a deal sheet to draft a first pass. A partner feeds a confidential file into a browser tab to save twenty minutes.
None of it appears in any ledger. None of it leaves a trace you can read. That is the true shape of the threat. Not that your staff use artificial intelligence, but that they use it without an audit trail. The data left the building. You cannot prove where it went, what it touched, or who can recall it.
The old discipline of shadow IT was a problem of unsanctioned tools. A team bought a subscription without telling procurement. It was a governance failure, but a legible one. You could find the invoice. You could revoke the licence.
Shadow AI is different in kind. There is no invoice. The tool is free, instant, and already open in the browser. The data does not sit in a rogue database you can audit later. It is absorbed into a model you do not control, on infrastructure you cannot inspect, under terms you did not negotiate. The perimeter you spent a decade hardening is bypassed by a copy and a paste.
For a family office, a regulated firm, or any organisation holding privileged material, this is not a productivity question. It is a liability question. When the regulator asks where the client data went, "we do not know" is not an answer. It is an admission.
The instinct is to issue a prohibition. Block the domains. Send the memo. Add a line to the handbook.
This fails, and it fails quietly. The work pressure that drove staff to public models does not disappear when you forbid the tool. The behaviour migrates to personal devices, personal accounts, the phone in the pocket. You have not closed the risk. You have blinded yourself to it. The audit trail goes from thin to none, and the leadership that issued the ban now believes the problem is solved.
A prohibition you cannot enforce is worse than no policy at all, because it manufactures the illusion of control.
The only durable answer is to give your people the thing they wanted, which is fast, capable reasoning over their own material, and to keep every byte of it on hardware you own.
That is what ARK O/S is for. The vault is a folder of plain files on a machine the firm owns. The model runs on the desk by default, and it answers with the network off. Every answer names the note it came from. Nothing is sent anywhere to be answered, and we hold no copy, because nothing of the firm's passes through us.
Capability alone is not governance. Governance is the trail. What your people make in the vault is sealed the moment it is made: a hash of the text and a signature from a key made on their own Mac. Change one character and the seal breaks. Anyone you show the record to can check it against a public register, without asking us and without trusting us. The seal proves the text has not changed and that one key signed it. It does not prove when, and it does not prove who a person is. It is a narrow claim, and a narrow claim is one counsel can rely on.
Shadow AI is dangerous because it is invisible. A vault on your own hardware is safe because it is legible. The same staff doing the same work, but now every answer has a source, every record has a seal, and all of it is yours to audit.
Say it straight, because the firms we want will ask. The vault for one person is free, forever, with no account and no telemetry, and the first public build lands in the week of 9 November. The vault shared by a whole firm, at £240 a seat a year, is not on sale yet. It goes on sale when two things ship: the import of PDF, Word, mail and meeting transcripts, and the vault opened by several people. We take no money for a thing that has not shipped.
The longer argument, for the CISO, the General Counsel and the CTO, is in the Shadow AI white paper at /the-shadow-ai-white-paper/. It is free to read and it is sealed.
Shadow AI enterprise risk does not close when you forbid the tool. It closes when the sanctioned tool is better than the shadow and bound to a record. You do not win this by saying no. You win it by offering something the shadow cannot: the firm's memory, on the firm's hardware, with a trail. The firms that survive the coming reckoning will not be the ones who banned the technology. They will be the ones who governed it.
