DEV Community

Micky Irons
Micky Irons

Posted on Edited on

Telling Clients About AI Use: Do Solicitors Have To?

No single rule forces disclosure. The SRA expects it to be clear to clients where they are dealing with AI, which in practice means a general clause in the client care letter plus specific disclosure where AI does substantial work on a matter. Mickai seals every consequential AI action into a signed Open Audit Record.

What does the SRA expect on transparency?

The SRA has not written a rule that says you must announce AI use. It has said something narrower and more useful. In its compliance tips for solicitors, last updated on 9 February 2026, the regulator states: "It should always be made clear to clients where they are interfacing with AI."

Read that wording carefully. It turns on interfacing. If a client is typing into a chat window on your website, or reading a reply a model drafted, they should know that is what is happening. It does not say every internal use of a model needs its own notice.

The SRA's warning notice on the misuse of AI works the same way. It adds no disclosure duty. It restates that using these tools does not reduce or transfer the professional and regulatory responsibility you already carry for the quality and accuracy of your work, and it warns that feeding client material into public AI services puts that material somewhere you no longer control.

Underneath both sits the Code of Conduct. Paragraph 8.6 requires you to give clients information in a way they can understand, and to put them in a position to make informed decisions about how their matter will be handled. Whether AI involvement is information a client needs in order to decide is a judgement for the firm, matter by matter. In my experience the honest answer is usually yes, and the firms that say nothing are not making that judgement. They are avoiding it.

Should AI use go in the engagement letter?

Yes, and it should be short. A standing clause in the client care letter is the cheapest way to meet the transparency expectation, and it stops you having the same conversation from first principles on every new file.

The clause is not a disclaimer and should not read like one. It is a description of how the firm works. Clients accept that their conveyancing file passes through a case management system. They will accept that certain drafting and review tasks are assisted, provided you say which tasks, say that a named solicitor reviews the output and remains responsible for it, and say where the data goes. The firms that get pushback are usually the ones whose wording is vague enough to sound like an attempt to shift risk onto the client.

One warning. If your letter says the firm does not use AI, and a fee earner uses it anyway, your problem is no longer disclosure. It is a misleading statement to a client, sitting on a file. A standing clause that accurately describes assisted work is safer than a denial you cannot police.

When does a matter need specific disclosure?

When the model does work the client would reasonably expect a qualified person to have done, or when it touches material the client would not expect to leave your building. Those are the two tests I would apply, and they catch different things.

The first is a substitution test. A model summarising a bundle, producing a chronology from correspondence, running first-pass disclosure review or drafting a clause set is doing work the client is paying a solicitor to do. Say so, and say who checked it. Litigation is the sharp end here, because the courts have already dealt with fabricated citations reaching the bench and the professional consequences that followed. We wrote about that in the High Court hallucination referral.

The second is an egress test. Cloud services stay a sensible choice for work that carries no privilege and no client confidence. Where it does, if the tool sends client documents to a third party's infrastructure, that is a confidentiality question before it is a transparency question, and it needs answering in terms of where the data physically sits rather than in terms of vendor assurances. That is why our Legal studio runs on the firm's own hardware, offline capable, with no data egress. Having nothing to disclose about onward transfer is a better position than a paragraph explaining it.

One honest note on documents, since this is where firms push hardest. A local OCR runtime has read scanned PDFs in controlled tests. Extraction and ingestion integration into SIOS is still being completed, so if your plan depends on scanned bundles today, ask us where that work actually stands rather than reading a capability into a brochure.

Does the client have to consent?

Not as a general proposition, and not for AI as a category. Consent becomes the live issue in two narrower places: data protection, and the client's own contractual terms.

On data protection, consent is only one lawful basis under the UK GDPR, and it is rarely the one a firm relies on to process client personal data in the course of a retainer. The ICO's guidance on AI and data protection covers transparency, lawful basis and the impact assessment work that sits alongside a deployment. Read it before you decide that a consent tick box solves anything, because in most retainers the duty is to inform rather than to obtain permission.

Contract is the part firms underestimate. Outside counsel guidelines and panel terms increasingly restrict or prohibit AI use, or require notice and prior approval. That is not regulation, and it binds you just as tightly. Keep a register of which clients have restricted AI use, and put it in front of fee earners at the point they open a file. If you cannot answer the question of which clients have said no, you cannot honour the ones who did.

How do you evidence what the AI actually did?

You need a record written at the time of the action, naming the person who approved it, that somebody outside your firm can check without taking your word for anything. This is the part most deployments have not solved, and it decides whether your disclosure is a statement or a demonstration.

What usually exists is a supplier dashboard. Retention sits at the vendor's discretion, the format is theirs, and you cannot hand a regulator or an opposing party something they can independently test. Mickai works the other way round. Consequential actions wait for a named person to approve them, so the record captures a human decision rather than a model output drifting into a file. Each one is then sealed into the Open Audit Record under ML-DSA-65, the post-quantum signature scheme NIST published as FIPS 204 in 2024.

The distinction that matters is tamper-evident, not tamper-proof. Nothing can stop a person editing a file on a disk they control, and any vendor claiming otherwise is selling you a story. Tamper-evident means that if the record is altered, verification fails, and the failure is visible to whoever checks. An auditor exports the record and verifies the signature offline with a public key, using any standard FIPS 204 implementation rather than software of ours. Our own audit verifier runs the same check in a browser, and it is a convenience rather than the proof. If your evidence can only be validated inside the supplier's own software, it is not evidence. It is a claim.

What should a disclosure clause cover?

Six things, in plain language: the categories of task where AI is used, the fact that a named solicitor reviews and remains responsible for the output, where client data is processed and whether it leaves the firm, the client's right to ask that AI is not used and how to exercise it, what record of AI involvement is kept, and how long that record is retained.

Get those six right and disclosure stops being a compliance chore. It becomes something you can show. A client asks what the model did on their file, and you export a signed record instead of describing a process. That is the difference between a firm that has adopted AI and a firm that can prove how it did.

Mickai is a UK company, held privately by me, and the record and approval mechanics described here sit inside 104 filed UK patent applications carrying 2,340 claims. Our closed beta is open. If you would rather see the record verified than described, the sovereign AI page is the place to start.

Frequently asked questions

Do I have to tell my client that AI helped with their matter?

No rule compels a notice in every case. The SRA expects it to be clear to clients where they are interfacing with AI, and the Code of Conduct requires information they can understand. The workable answer is a standing clause in the client care letter, plus a specific note where AI did work a client would expect a solicitor to do.

Should AI use go in our standard engagement letter?

Yes. It is the cheapest place to satisfy the transparency expectation, and it saves repeating the conversation on every file. Keep it factual: which categories of task are assisted, that a named solicitor reviews and remains responsible, where data is processed, and how a client asks for AI not to be used.

Does the client have to give consent to AI use?

Usually not. Consent is only one lawful basis under the UK GDPR and firms rarely rely on it for retainer work, so the duty is normally to inform rather than to ask. Contract is different. Outside counsel guidelines and panel terms can require notice or prior approval, and those bind you just as tightly.

What do we do if a client says no to AI?

Honour it, then prove you honoured it. Record the restriction against the client and the matter, make it visible to fee earners when they open the file, and route the work to a lane with AI switched off. A signed record of actions on that matter lets you demonstrate the absence rather than assert it.

How do we prove what the AI did on a matter?

You need a contemporaneous record, a named approver, and independent verification. Mickai seals every consequential action into the Open Audit Record under ML-DSA-65, the scheme NIST published as FIPS 204. An auditor exports it and checks the signature offline with a public key, using tooling that is not ours. Alteration makes verification fail.


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)