In partnership with
In the space of a few weeks, three AI businesses have launched agents which keep working when nobody is asking them to: Meta's Muse on 8 September, Manus's Cue on Monday, and OpenAI's Dots on Tuesday. Only the last of these is likely to reach a UK law firm soon, through business ChatGPT plans, and it is the lead this week. Elsewhere, a legal AI vendor has given clients' own agents a route to ask their law firm questions, Clio has bought a company which builds AI for judges, and a UK-hosted research assistant has launched with a claim this writer would test before believing.
AI in Practice
OpenAI's always-on agents are now available to UK businesses
OpenAI launched dots at its developer day on 29 September, describing them as "remarkably capable, always-on agents built to handle everything". Each dot runs on GPT-6 Astra, has its own cloud computer and browser, can connect to more than 4,000 apps through OpenAI's plugins, and can be messaged or called in ChatGPT, Slack and Teams. OpenAI says a dot learns a user's preferences over time and will sometimes bring work back "before you even think to ask". The rollout covers Pro and Business Premium plans in eligible markets, and reporting on the launch states that the Pro rollout excludes the UK, the EEA and Switzerland while Business Premium is available across all supported ChatGPT regions. Enterprise customers can try a beta once their workspace administrator turns it on. For UK firms, then, the route in is the business plan rather than an individual subscription.
The feature this author would look at first is what OpenAI calls "proactive research". When a user is not actively working with it, a dot looks for ways to help by going through the apps the user has connected. OpenAI stresses that those tools are restricted to read-only, so the dot cannot send messages or change content while doing this. That is a limit on what the agent can do, and it says nothing about what it can read. If a fee earner has connected their email, calendar or document storage, the agent is reading client material, on a third party's computer, at times nobody chose, in search of something useful to do. Business and Enterprise content is not used to train OpenAI's models by default, according to the announcement, but training is only one of the issues a firm's confidentiality and data protection analysis has to answer.
The controls are more detailed than for most earlier products. Users can set Custom Rules which allow, require approval for, or block specific actions, background work is visible in an Activity View, and OpenAI says certain sensitive tasks always stay with the user. OpenAI is also piloting "specialist dots" with their own identity and credentials inside organisations, naming commercial contracting among the areas tested internally, and is working with Microsoft to manage them through its Agent 365 governance tools. OpenAI's own advice is that dots "can still make mistakes, so always review consequential work". It is too early to say how well any of this works in practice.
The broader point is the trend. When this newsletter covered Chamelio's always-on agents for in-house teams on 11 September, it looked like a niche legal product. Within a month Meta, Manus and OpenAI have all shipped general agents built on the same idea, and it seems reasonable to expect the other major vendors to follow. Most firm AI policies were written for a lawyer who asks a tool a question and checks the answer. An agent which decides for itself what to look at, and which can act once a condition is met, needs a different kind of supervision, and the SRA's August warning notice leaves no doubt that the supervision remains the firm's responsibility whoever, or whatever, did the work.
Takeaways
Act: Find out whether your firm has a ChatGPT Business Premium or Enterprise workspace, and if so ask the administrator whether dots are enabled and which apps staff have connected. Until the firm has decided otherwise, keep client email and the document management system unconnected, and set Custom Rules so that any action which sends, shares or signs anything requires approval.
Watch: Whether the other major AI vendors launch their own always-on agents over the coming months, and whether the ICO's planned work on agentic AI addresses agents which read data in the background without a specific instruction.
Risk: An agent reading client correspondence on a third party's systems without anyone at the firm having decided that it should, and a fee earner on a business plan who switched it on without telling anyone.
Read: OpenAI, with regional availability reported at Unite.AI, and the launch coverage at TechCrunch
On your radar
Your client's AI agents may soon send legal questions straight to your firm: August, a legal AI platform aimed at midsize firms and in-house teams, launched August for Agents on 29 September. When an agent working inside a company (a procurement agent reviewing a supplier's terms, for example) reaches a legal question it cannot resolve, it sends the question to the company's law firm, whose system either answers immediately where it has been authorised to or routes the question to a lawyer for review and approval. August says the firm keeps supervision of, and responsibility for, all advice given, and its co-founder Thomas Bueler-Faudree put the case simply: "Companies are handing real work to agents, and those agents keep hitting questions only a lawyer should answer." It is a US business and the announcement is light on detail. Why it matters for UK lawyers: a question from a client's agent is still a request for advice, and an answer given without a lawyer looking at it is still advice given by the firm, with the usual questions about scope of retainer, who the client actually is, and what goes on the file. If a client proposes anything like this, settle in the engagement letter which categories of question can be answered without a lawyer's review, and make sure every answer is recorded on a matter file. (Artificial Lawyer)
Clio has bought a company which builds AI for judges: Clio announced on 30 September that it has acquired Learned Hand, whose AI workspace connects to court case management systems and helps judges and clerks work through filings, research the law, and prepare bench memoranda and draft orders. It is used across US courts, including the Superior Court of Los Angeles County, and no UK courts are mentioned. Terms were not disclosed. Clio's chief executive Jack Newton said that if lawyers and litigants move dramatically faster while courts stay constrained, "the bottleneck simply shifts to the courts". Why it matters for UK lawyers: this is US news, but Clio sells practice management software to a large number of UK firms, and a vendor serving both the bar and the bench raises an obvious question about how data from each side is kept apart. If your firm uses Clio, ask in writing how matter data held in its practice products is separated from its judiciary business, and keep the answer on your supplier file. (PR Newswire, Artificial Lawyer, Legal IT Insider)
A UK-hosted research assistant says it "cannot fabricate citations": Alesis, built by L25 Group (whose co-founders Luke Roberts and Alex Clayton previously worked in the public sector, including at the Ministry of Justice), went live on 28 September. It is grounded in the law of England and Wales, Scotland and Northern Ireland, draws on UK case law, Companies House filings, ombudsman decisions and regulator material, holds data in the UK with processing in the UK and EU, and is priced on credits rather than seats (around £25 lasting about 17 days of typical use, according to the company). It reports 25 firms and 150 users, mostly SME firms and in-house teams. Roberts says of it: "It doesn't lie. It cannot fabricate citations, it cannot make things up." Why it matters for UK lawyers: a UK-specific, low-cost tool is useful for the smaller firms this newsletter is written for, but the claim that a tool cannot fabricate is a sales claim, and the EAT's judgment in Hancox, covered here last week, is a reminder that the checking duty sits with whoever files the document. If you trial it, run it first on questions where you already know the answer and check every citation against the source before relying on it for anything else. (Legal IT Insider)
Ad Break
In order to help cover the running costs of this newsletter, please check out the advert below. In line with my promises from the start, adverts will always be declared.
How much does a $70k hire actually cost?
A $70K salary might look straightforward on paper. But once you factor in employer taxes, statutory benefits, payroll, and country-specific employment costs, that number can skyrocket.
And those costs can vary significantly depending on where your employee is based. So before you build your hiring budget, check out Oyster's Global Hiring Cost Calculator.
The calculator can help you see the full picture by letting you compare employment costs across 120 countries and estimate what a hire could really cost your business. And unlike that new hire, it’s actually 100% free.
For Review
AI can do the work but it cannot inherit the lawyer's duty (Rocky Scopelliti, Legal Futures)
A short blog from 30 September on what the author calls the "human rubber-stamp problem": as AI output becomes more reliable, lawyers are more likely to approve it without reasoning through it themselves, and a human somewhere in the process is not the same as human judgement having been exercised. It is the right piece to hand to anyone drafting a supervision rule for agents, because it explains why "a lawyer approved it" is not, on its own, an answer.
Read or listen: Legal Futures
ICO tech futures: agentic AI (Information Commissioner's Office)
Not this week's news: the ICO published this in January. It is the closest thing to a UK regulator's view of always-on agents, and its list of risks (controllers and processors becoming unclear along the supply chain, purposes set too broadly for open-ended agents, and agents processing more personal data than the task needs) reads almost as a checklist for the dots launch. Use it as the starting point for any data protection assessment before an agent is connected to client systems.
Read or listen: Information Commissioner's Office
UKJT Legal Statement on Liability for AI Harms (Society for Computers & Law)
Also background rather than news: the UK Jurisdiction Taskforce published its statement on 7 July, concluding that existing English private law (contract and negligence above all) can deal with most AI harms without bespoke legislation, that AI has no legal personality of its own, and that negligence can extend to a failure to use AI as well as to over-reliance on it. When an agent acts for a firm, this is the framework which will decide who bears the loss, so it is worth reading before a client asks.
Read or listen: Society for Computers & Law
Practice Prompt
Try the below prompt to decide, before anyone switches one on, what an always-on AI agent would be allowed to read and do inside your firm, which is the question this week's dots launch puts in front of every firm with a business ChatGPT plan. Ensure you fill in context and constraints and other aspects marked with {}. Remember to adhere to the Golden Rules and do not upload confidential or privileged information to public tools.
You are assisting a COLP, COFA, practice manager or partner at a law firm in England and Wales who has to decide whether, and on what terms, staff may use an always-on AI agent: one which works in the background, reads connected apps without a specific instruction, and can take actions once a condition is met. Your task is to produce a permissions and supervision plan for that agent. You are not assessing whether the product is good, and you are not concluding that the firm may lawfully or safely use it.
Context to apply:
- The agent product and plan, from public material only: {e.g. OpenAI dots on ChatGPT Business Premium, or another product; say "unknown" if not yet identified}
- Who administers the workspace and who can switch features on: {name the role, or say "nobody has been assigned"}
- The apps staff have connected or want to connect: {e.g. Outlook email, calendar, Teams, SharePoint, the document management system, the case management system, accounts software}
- What the vendor says about background reading, data retention, training on customer data, where data is processed, and the controls available: {summarise from public material, or say "not yet checked"}
- Who wants to use it and for what: {e.g. a partner wanting inbox triage, a paralegal wanting chasers sent, an in-house client wanting contract review}
- Practice areas and client types whose material would be in scope: {e.g. private client, family, commercial, regulated financial clients, clients with confidentiality clauses or AI restrictions in their terms}
- The firm's current AI policy and supervision arrangements: {summarise, or say "none in writing"}
Produce the plan under these headings:
1. What the agent could see
For each connected app, state what client and personal data the agent could read during background work, whether that reading happens without any instruction from the user, and where the data would be processed. Separate what the vendor has confirmed from what has been assumed.
2. What the agent could do
List every action the agent could take in each app (send, reply, share, file, delete, accept a calendar invitation, sign, pay). For each, recommend whether it should be allowed, require approval each time, or be blocked, and give a one-line reason tied to confidentiality, supervision or client instructions.
3. Apps to keep disconnected for now
Identify the apps or data sources which should not be connected until the firm has completed a data protection assessment and checked client terms, and say what the firm would need to know before reconnecting each one.
4. Supervision
Set out how a supervisor would know what the agent did, on which matter, and on whose behalf, using the activity records the product provides. Identify any action which could happen with no lawyer reviewing it at the time, and say how the firm would detect it afterwards.
5. Client and data protection questions
List the questions the firm must answer before any client material is in scope, covering at least: whether client terms of business or engagement letters permit this processing; whether the firm is acting as controller for the agent's processing and what the vendor's role is; whether clients need to be told; and what happens to the agent's notes and memory when a matter closes or a fee earner leaves. Mark clearly where you are uncertain about the law or ordinary practice rather than asserting a position.
6. Decision note
Draft a short note for the management meeting stating what is recommended (no use, a limited pilot, or wider use), on what conditions, who owns the decision, and when it will be reviewed.
7. What this plan cannot tell you
Flag every point where the output rests on an assumption about the product, the firm's systems, client terms or the law, and say what the firm would need to establish to replace it with a fact.
Constraints:
- {Add firm-specific constraints, for example a client who has prohibited AI use, an insurer requirement, an existing Microsoft 365 configuration the firm cannot change, or a pilot budget.}
- Apply the law of England and Wales, UK GDPR and the Data Protection Act 2018, and the SRA Standards and Regulations. Where you are unsure whether something is a legal requirement, regulatory expectation or ordinary practice, say so rather than asserting it.
- Do not invent product features, vendor commitments, regulatory guidance, case law or statutory provisions. If you do not know what the product does, write the question to put to the vendor rather than an assumption.
- Do not conclude that the arrangement is compliant, safe or adequate. The purpose is to set the boundaries and surface what is not yet known.
- Write in plain English and keep the decision note short enough to read in two minutes.
- This is a planning aid, not legal or compliance advice. The firm, its managers and its COLP remain responsible for any use of the agent and for everything done through it.How did we do?
Hit reply and tell me whether anyone at your firm has already been offered an always-on agent, and what you decided.
Serhan
UK Legal AI Brief
Disclaimer
Guidance and news only. Not legal advice. Always use AI tools safely.
Recommended Newsletters
Below are a few newsletters that I recommend, for various reasons. Check them out!
