DEV Community

Micky Irons
Micky Irons

Posted on Edited on

NDA and AI Tools: Does AI Use Breach Confidentiality?

It can. Most NDAs limit disclosure to named recipients and a permitted purpose, so sending a counterparty's documents to an outside AI service is a disclosure to a third party unless the agreement allows it. A tool running on hardware you control, with nothing leaving, usually stays inside the terms. Read the permitted recipients clause first.

Which NDA clauses decide the answer?

Three clauses decide it: the definition of confidential information, the permitted purpose, and the permitted recipients list. Read them in that order and you usually have your answer in ten minutes.

The definition is normally wide. It catches the material in any form or medium, and it catches notes, analyses and anything derived from the material. So a summary your AI tool produces is itself confidential information, and so is a search index built from the documents.

The permitted purpose is rarely the problem. If you are allowed to evaluate a transaction, using software to help you evaluate it fits the purpose.

The recipients clause is where this comes unstuck. The standard form permits disclosure to your directors, officers, employees and professional advisers who need to know for the permitted purpose, on condition they are bound by obligations no less onerous than your own. A software company you send the documents to is not a director, an employee or a professional adviser. Unless someone has widened that clause, it is not a permitted recipient.

Two further clauses matter more than people expect. The return and destruction obligation asks you to hand back or destroy every copy on request. If a provider retains inputs for a period, or your prompts sit in a vendor's logs, you cannot honestly certify destruction. And the security obligation, usually a duty to protect the material with at least the care you apply to your own, is the clause a counterparty reaches for once something has already gone wrong.

Is an AI service a third party under the agreement?

Yes, where the material leaves your control and arrives in a system another company operates. The transmission is the fact that matters. Whether a human at the provider ever reads it does not change what you did.

Enterprise terms often promise that customer content is not used to train models. Take that at face value and it still does not answer the question in front of you. The documents have been transferred to another legal entity, processed on infrastructure that entity controls, possibly retained for a window for abuse monitoring, and possibly handled by its own subcontractors. Under English law, confidentiality rests on the agreement you signed and on the equitable duty that sits behind it. Both turn on whether you put the material into someone else's hands.

There is a second trap here. A provider's terms bind you and the provider. They do not bind you and your counterparty. Only the counterparty can permit a disclosure their NDA prohibits, and a warranty from a vendor is not that permission.

Where the documents contain personal data, a parallel set of obligations runs alongside the contract. The ICO's guidance on AI and data protection is the starting point. Its response on allocating controllership across the generative AI supply chain sets out the ICO's position that controller and processor roles fall according to what each party actually determines about the processing, rather than according to the label a contract gives them. That is a consultation output rather than statutory guidance, so treat it as the regulator's thinking and not as a rule. The practical point stands: you cannot contract your way out of a role you are in fact performing.

Do AI-specific clauses now appear in UK NDAs?

They do, and in three recognisable shapes. The first is a flat prohibition: no confidential information may be entered into any artificial intelligence, machine learning or large language model system. The second is a permission with a boundary, allowing processing on systems under the receiving party's sole control provided nothing is transmitted outside that control and nothing is used to train a model. The third is notice and consent, where you name the tools in advance and have them approved.

Older agreements say nothing about AI at all. Silence is not permission. The recipients clause was already doing this work before anyone drafted an AI paragraph, and it carries on doing it.

Watch for wording drafted for other purposes that catches AI by accident. Bans on automated processing, on data mining, or on moving the material to any system outside a named jurisdiction were mostly written with other risks in mind. They still bite.

Does an on-premise deployment stay inside the terms?

Usually, because the disclosure never happens. If the model runs on hardware your organisation owns or controls, the documents stay inside your network, and no copy reaches a vendor, then there is no third party to point at. What you have is your own staff reading your own files with better tools, which is what the permitted recipients clause already allows.

That is the reason the Mickai Sovereign Intelligence Operating System is built the way it is. SIOS runs on hardware the customer owns, it is capable of running offline, and there is no data egress. It ships with 63 studios, 14 of them production-ready at launch and 49 in development, and 50 specialised models we call brains. A closed beta is open, with one regulated company onboarding as a design partner.

One honest limit, because it bears directly on deal documents. A local OCR runtime has read scanned PDFs in our controlled tests, and the extraction and ingestion integration inside SIOS is still being completed. If your data room is largely scans, ask any supplier including us to show ingestion on your own files rather than accepting a general assurance.

Two more cautions. An exhaustive recipients list naming only individuals can still be read narrowly, and a flat prohibition on automated processing bans the local system as surely as the remote one. Owning the hardware removes the disclosure. It does not rewrite the clause. There is more on the deployment side in private AI over internal knowledge and sovereign AI.

What should you ask before a deal starts?

Ask while the NDA is still a draft, because the recipients clause is cheap to change then and expensive to argue about later. Five questions cover most of it.

Does the recipients clause contemplate systems as well as people, or only people? Does any AI clause distinguish a system you control from a service you transmit material to, or does it treat both the same? Can you satisfy the destruction obligation for every copy, including indexes, embeddings and caches? What standard does the security clause set, and can you show you meet it? And who inside your organisation authorises the use, on the record?

If you want a single line to propose, ask for processing on systems under the receiving party's exclusive control, where no confidential information is transmitted outside that control. Counterparties tend to accept it, because it gives them a defined boundary instead of silence and guesswork.

How do you evidence compliance if you are challenged?

With records made at the time, not with an assurance written afterwards. The useful record is one the other side can check without having to trust you.

In SIOS that is the Open Audit Record. Every consequential action is sealed under ML-DSA-65, the post-quantum signature scheme NIST published as FIPS 204 in 2024. An auditor takes an exported record away and verifies it offline with a public key, using tools that are not ours. Consequential actions also wait for a named person to approve them, so the record shows who authorised the work and when.

Be precise about what that buys you. The record is tamper-evident, not tamper-proof. I do not claim a record cannot be altered. I claim that altering it makes verification fail, so the attempt surfaces instead of passing quietly. That is the property that matters when someone asks whether the log they have been handed is the log that was written. A tamper-proof claim would be unfalsifiable, and no auditor should accept one.

For the record on who is saying this. Mickai LTD is a UK company, Companies House 17166618, holding 104 filed UK patent applications carrying 2,340 claims. None of this is an argument against the companies building the compute and cloud layer, which remains the sensible choice for work that is not regulated. The argument is with the assumption that a regulated business must rent its intelligence, ship a counterparty's documents offsite, and take a vendor's word for what happened to them.

None of it settles the question for you either. The wording of your specific agreement governs, and a record proves what happened rather than whether you were permitted to do it. This article is a general explainer and not legal advice: take advice on the agreement in front of you.

Frequently asked questions

Does using AI on a counterparty's documents breach our NDA?

It depends on where the tool runs. Sending the documents to an outside AI service is a disclosure to a third party, and most NDAs permit disclosure only to named individuals who need to know. Running a model on hardware you control, with nothing transmitted out, is not a disclosure at all. Check the recipients clause first.

What NDA wording actually covers AI tools?

Usually the permitted recipients clause, even in agreements that never mention AI. It lists who may receive the material, normally directors, employees and professional advisers who need to know. Newer agreements add an explicit AI clause, either banning input into any AI system or allowing it only on systems under the receiving party's sole control.

Is a private, on-premise AI tool a disclosure to a third party?

Normally not, because nothing leaves your control. If the model runs on your own hardware, inside your own network, and no copy reaches a vendor, there is no third party receiving the material. Your staff are reading your own files with better tools. A clause banning automated processing outright would still catch it, so read the wording.

Should we ask permission before using AI on deal documents?

If the tool sends material outside your control, yes, and get it in writing. If the system runs entirely on your own infrastructure, permission is usually unnecessary, though naming the approach during negotiation costs nothing and removes an argument later. The cheapest fix is a line in the draft NDA permitting processing on systems you exclusively control.

What happens if the other side finds out we used AI?

That depends on what you did and what you can show. Unauthorised disclosure to a third party is a breach, and the usual remedies are injunctive relief, damages and a demand for return and destruction of every copy. If the processing stayed on your own systems, the argument turns on the wording rather than on any transfer.


Related briefings

Legal practice

Data protection and UK GDPR

Part of a series of 60 briefings on deploying and governing AI in UK regulated organisations, archived with a DOI at 10.5281/zenodo.22975756.

Evaluating AI for a regulated organisation? Mickai runs on hardware you own, offline. Consequential actions wait for a named person to approve them, and what the AI did is sealed into a signed record an auditor can check without us. Applications for the invitation-only closed beta are open. Apply for the closed beta.

Written by Micky Irons, founder and chief executive of Mickai LTD.

Top comments (0)