Start from what is already happening #
Most firms come to this after the fact. A paralegal has been summarising documents with a free tool for three months because it works, a fee earner has drafted a first attempt at a letter the same way, and nobody has ever been told whether either is permitted. The policy is then written to close a gap rather than to open a door, which is the wrong order but the normal one.
That matters for what the document has to do. A policy that describes an intended future does nothing about the tools in use today.
What the regulator has actually said #
The SRA published a Risk Outlook report on the use of artificial intelligence in the legal market on 20 November 2023. It is worth reading in full, and three of its points do most of the work.
On accountability it is unambiguous.
As with any other technology or system in your firm, you will remain responsible and accountable for the outputs from AI you are using.
On reliability it explains why these tools fail in a particular way, describing systems that work "by anticipating the text that should follow the input they are given" without "a concept of 'reality'", producing what it calls hallucination, "where a system produces highly plausible but incorrect results".
On confidentiality it names specific mechanisms rather than a general worry.
-
A staff member using an online tool to answer a question on a client’s case
-
“Confidential data being revealed when it is transferred to an AI provider for training”
-
“The output from an AI system replicating confidential details from one case in its response to another”
Those are three different failures with three different controls, and a policy that treats them as one covers none of them properly.
What the Law Society has said #
Law Society guidance for solicitors on generative AI is blunter about free tools than most firms expect.
If you are using a free, online generative AI service where you have no operational relationship with the vendor other than use, do not put any confidential data into the tool.
More generally it advises against feeding confidential information into these tools "especially if you lack direct control and oversight over the tool's development and deployment".
On supervision it removes a common escape route. A solicitor's professional duties, in particular the duties to the court and to the client, apply to work carried out by the solicitor regardless of whether AI was used to assist with it, and they apply "whether AI was used by the solicitor personally or by anyone under that solicitor's supervision".
The guidance also separates two kinds of training that firms routinely conflate, noting that technical training in using a tool is distinct from training in the ethical use of it and that both matter.
Confidentiality and privilege are two questions, not one #
Paragraph 6.3 of the Code of Conduct for Solicitors is a professional obligation: "You keep the affairs of current and former clients confidential unless disclosure is required or permitted by law or the client consents." It is owed by the solicitor, and it is the duty most AI policies address.
Legal professional privilege is a different thing. It belongs to the client, not to the firm, and whether a particular disclosure to a third party service affects it is a legal question that depends on the facts. This page will not answer it and no supplier should offer to.
Data protection runs on a separate track #
If client information identifies a living individual, UK GDPR applies and the firm is accountable for how it is processed, independently of anything the SRA says.
The ICO publishes guidance and resources on artificial intelligence, including detailed guidance on applying UK GDPR principles to AI systems, guidance on explaining decisions made with AI, and an AI and data protection risk toolkit for assessing risks to individual rights and freedoms. Those are the documents an assessment should be built on.
An accountancy practice faces the same three-way split, set out on the client data in AI tools page, and the general version is on whether ChatGPT is GDPR compliant.
The five decisions a policy is really recording #
Strip the boilerplate out and a usable policy settles five things. Not how to do any of them, which depends entirely on the firm, but that a decision was taken and by whom.
-
Which tools
Named, with a route for asking about a new one, because an unanswerable request turns into quiet use.
-
What may go in
By category of information, not by good intentions. This is where confidentiality, privilege and personal data all land.
-
Who is accountable
A named person, consistent with paragraph 2.1 of the Code of Conduct for Firms on governance structures and 2.5 on managing material risks.
-
What happens to output before it leaves the firm
Given what the SRA said about non-existent cases, this is a supervision question under paragraph 4.4, not a preference.
-
What is recorded
Paragraph 2.2 requires records to demonstrate compliance, and a decision nobody wrote down is hard to evidence later.
What runs on rules, and what does not #
Runs on rules
- Knowing which systems hold client information, though most firms have never written the list down
- Knowing who has access to what
- Recording that a decision was taken, and when
- Evidencing that training happened
Needs a person
- Deciding whether a category of information may be used at all
- Deciding whether a disclosure affects privilege
- Reviewing output before it reaches a client or a court
The first belongs to the firm, the second is a legal question, and the third is judgement, always.
What this consultancy will not do #
- Tell you that a tool is safe for client matters
- Tell you that a configuration is compliant
- Tell you that a supplier's assurances are enough
Those conclusions are the firm's to reach, and automating a step has never made anybody compliant with anything.
What is useful to say is that a surprising amount of the admin load in a firm never needs to touch privileged material at all. Knowing which files have gone quiet, which letters have not come back, which matters are waiting on the same third party: those are questions about the state of the firm rather than about the substance of a case, and that distinction is where the safe ground is.
It runs through the update nobody has time for, and the task sorter is a quick way to see which of your own jobs fall on which side.