In partnership with
Last week the SRA told the profession that supervision is the real AI risk. This week two of the largest vendors in the market launched products which put another layer between a lawyer and the work: Google Cloud has taken Gemini Enterprise into legal, and LexisNexis has rebuilt Protégé around a system which chooses models and agents for you. Elsewhere, Avvoka has plugged its drafting engine into Harvey, and the Online Procedure Rules are now a fortnight from coming into force.
AI in Practice
Google enters legal, and sells the orchestration layer rather than the tool
Google Cloud launched Gemini Enterprise for Legal on 25 August, a version of its Gemini Enterprise platform configured for law firms and in-house teams. Cleary Gottlieb, Freshfields, Weil and Williams & Connolly are named as launch firms, and the product is in preview rather than general availability. Google says it is the first in a series of industry-specific offerings built on the Gemini Enterprise base, with a companion product for financial services announced the same day.
The description Google gives is worth a close look, because the product is not what a lawyer would expect. It is organised around four components:
skills which guide agents through legal tasks such as contract review, legal research, regulatory monitoring and preparing court filings;
connectors into the systems a firm already runs;
third-party agents supplied by legal technology vendors;
and a governance layer giving IT and risk teams visibility over outputs and access controls.
The named integrations are interesting, covering iManage, NetDocuments, Thomson Reuters, RelativityOne and Harvey. On the data question, Google says client data, prompts and outputs stay within the organisation's own cloud perimeter, that base models are not trained on customer data, and that existing document permissions are inherited through the connectors.
What Google is selling, then, is the layer that sits above the document management system and directs other people's agents, Harvey's included. It competes with Lexis and Harvey only in the sense that a motorway competes with a car.
This author would note two things about that. The first is that it puts the platform decision in the hands of whoever owns cloud infrastructure in your firm, and for the many firms sitting on Microsoft 365 the question of whether to add a Google layer is a considerable one that has very little to do with the quality of the legal output. The second is that "governance layer" is doing a great deal of work in the announcement without much detail behind it. This author is conscious that it arrives eight days after the SRA published a warning notice which put supervision at the centre of AI risk. Vendors have read that document too.
The honest position is that nobody outside the four launch firms can yet say whether this works. It is a preview, and the launch names are heavily weighted towards US firms with Freshfields as the one obviously UK-facing entry. The interesting claims is that a risk team can see what agents did and that permissions genuinely carry through. This does mean that two hyperscalers and the largest legal publisher are now building the same thing, and the choice a firm makes in the next year is less about which assistant to buy than about whose orchestration layer everything else plugs into.
Read: Google Cloud, the press release, and the coverage at Artificial Lawyer
On your radar
LexisNexis has rebuilt Protégé so the system chooses which model runs your task: Lexis announced on 24 August that it is rolling out the agentic capabilities of its Legal Intelligence Engine, described as a harness which dynamically selects and coordinates models, agents, skills and content sources according to the task a user describes, rather than routing the request through a workflow defined in advance. The practical claim is that a lawyer describes what they want, receives review-ready output in Word, Excel or PowerPoint, and the system carries context forward from one step of a matter to the next. Why it matters for UK lawyers: this is the same architectural move as Google's, from a vendor most UK firms already pay, which means it may arrive in your practice without a procurement decision being taken. Ask your Lexis account manager what the engine records about which model and which sources it selected, and whether that record can be exported onto a file. (LexisNexis release, LawSites, Artificial Lawyer)
Avvoka has plugged its drafting engine into Harvey, and can now build templates from your closed deals: The partnership, announced on 20 August, lets a lawyer working in Harvey start from the firm's approved template, draft with matter details in place, pull from the firm's approved clauses and carry the document through negotiation without leaving the platform. Alongside it, Avvoka launched Curate, in beta, which analyses uploaded transaction documents and produces customisable templates in hours rather than the months of knowledge management and partner time template projects normally consume. Curate runs on OpenAI and Anthropic models. Why it matters for UK lawyers: template creation has been the knowledge management bottleneck in most firms for a decade, and if it becomes cheap the constraint moves to whether anyone checks that the "gold standard" the tool extracted is the position the firm actually wants to take. If your firm pilots this, decide now who signs off a generated template before it goes into the precedent bank. (Artificial Lawyer, The Global Legal Post)
DraftWise is making the case that your own documents, rather than the model, are the asset: DraftWise launched "Legal Ontology" on 24 August, which structures the relationships between data sitting inside a firm's agreements, side letters and amendments, so that a lawyer can trace how a particular right was negotiated and varied across matters and see whether extending it to another counterparty would create obligations elsewhere. The company's example is a funds team tracing which LPs hold a given right. Why it matters for UK lawyers: this is the counterweight to the two stories above, because whichever orchestration layer a firm ends up on, the thing which distinguishes its output is its own document history. Before evaluating any of this, find out what state your document metadata and filing discipline are in, since none of these products work well on a badly kept system. (Artificial Lawyer)
The ICO's statutory code on AI and automated decision-making is due to move this month: The Data Protection Act 2018 (Code of Practice on Artificial Intelligence and Automated Decision-Making) Regulations 2026 require the Information Commissioner to produce a statutory code covering transparency and explainability, bias and discrimination, and rights and redress. The ICO's published guidance plans put a consultation in this period, with a final version due in the winter, though this timetable has slipped before and readers should check rather than assume. Why it matters for UK lawyers: a statutory code is not guidance, and a court or regulator must take it into account where it is relevant, so this will end up being cited at you rather than merely read by you. Check the ICO's guidance plans page, and if your firm advises on automated decision-making, diarise the consultation while there is still time to respond to it. (ICO, plans for new and updated guidance)
Reminder: the Online Procedure Rules come into force on 7 September: Covered here on 7 August, and repeated only because the date is now a fortnight away. The Online Procedure (Rules and Practice Directions) Rules 2026 take effect on 7 September 2026, with the first procedural rules, for possession proceedings, to follow later in the year. Why it matters for UK lawyers: if you conduct civil litigation, put the date in the diary this week and, if you do possession work, get the forthcoming practice direction onto your team's agenda before it lands. (legislation.gov.uk)
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.
Stop making AI decisions in the dark.
Leadership is asking: are we getting value from AI? Which tools are worth the spend? Where are we exposed? Right now, most teams have no idea.
Harmonic Security Usage Explorer changes that.
You get a complete picture of how your organization uses AI, automatically categorized into custom tasks and use cases.
You’ll see the projects being worked on, who’s using what tools, where AI investments are driving value, and where employees are engaging in risky behavior.
CIOs can rationalize spending and cut wasted licenses. CISOs can pinpoint where risk exists and neutralize it. AI committees can show exactly how their efforts are paying off.
For Review
Introducing Gemini Enterprise for Legal (Google Cloud)
The announcement behind this week's lead, and short enough to read in ten minutes. Go to the section describing the four components and the list of named integrations, which together tell you more about Google's strategy than the customer quotes do.
Read or listen: Google Cloud
LexisNexis Unveils "Legal Intelligence Engine", Rebuilding Protégé Around Dynamic Agentic Orchestration (Bob Ambrogi, LawSites)
The most detailed write-up of the Lexis announcement, and the one which explains what a harness is and why moving from predefined workflows to dynamic selection is an architectural change rather than a feature. Useful if you have to explain to a management board why this is not simply a new version of a familiar product.
Read or listen: LawSites
Why General AI Alone Is Not Enough for Legal Work (Senne Mennes, Artificial Lawyer)
Published on 19 August by Senne Mennes of LawVu, a former practising lawyer who co-founded ClauseBase, and the clearest short statement of the argument running underneath all three of this week's product stories: general models are competent at drafting and review but hold no knowledge of a particular organisation's positions, so the systems worth buying are the ones grounded in a firm's own material. It is a vendor byline and should be read as such, though the underlying point survives that.
Read or listen: Artificial Lawyer
Practice Prompt
Try the below prompt to interrogate an agentic or orchestrated AI platform before your firm commits to it, which is the question this week's launches all raise and none of them answers. It is built around the supervision obligations the SRA set out on 17 August, so what comes back is a list of questions to put to the vendor and a record of what you were told, rather than a verdict on the product. 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, risk partner or head of legal technology at a law firm in England and Wales. Your task is to produce a supervision and audit-trail due diligence pack for an agentic or orchestrated AI platform the firm is considering. The output is a set of questions and a scoring framework for the firm to use, not an assessment of the product and not a recommendation.
Context to apply:
- The platform under consideration: {name, and what the vendor says it does, pasted from public material only}
- What it would sit on top of: {document management system, research provider, email and document platform}
- Who would use it, and for what work: {roles, practice areas, and whether any output is court-facing}
- How that work is reviewed today, by role: {describe the actual practice, not the policy}
- What the firm records on a file about how a document was produced: {be honest, or say "nothing"}
- Any tools already in use which the platform would orchestrate: {list}
- Constraints the firm is working within: {existing licences, infrastructure commitments, insurer requirements, client mandates on AI disclosure}
Produce the pack under these headings:
1. What the vendor is claiming
Restate, neutrally and in plain English, what the platform is said to do, separating claims about capability from claims about control. Mark anything that is a marketing formulation rather than a testable statement.
2. Supervision questions
Draft the questions the firm must put to the vendor about how supervision remains possible, covering at least: which model handled a given task and whether that is recorded; which agents ran, in what order, and on what inputs; which documents and sources the system reached; what a supervisor sees when reviewing completed work; and whether any of this can be exported into the firm's own file.
3. Audit trail test
Set out a concrete scenario the firm should ask the vendor to demonstrate: a document goes out with an error in it, and eighteen months later the firm has to reconstruct how it was produced. List exactly what the firm would need to be able to show, and turn each into a question about whether the platform retains it and for how long.
4. Data and confidentiality questions
Draft the questions on where data sits, whether existing permissions carry through the connectors, what is retained after a matter closes, what happens to a departing fee earner's history, whether anything is used for training, and what the vendor does on receiving a third party request for firm data.
5. Where responsibility lands
Map each stage of a task run on the platform to the person in the firm who would be accountable for it, and identify the stages where nobody currently is. Flag any stage where the work would be beyond ordinary human review, since that is where the audit trail has to carry the whole weight.
6. Scoring framework
Produce a simple scoring sheet the firm can complete after the vendor responds, with a short list of answers that should stop the procurement outright and the reasoning for each.
7. What this pack cannot tell you
Flag every point where the output rests on an assumption you have had to make, and say what the firm would need to establish to replace it with a fact.
Constraints:
- {Add firm-specific constraints, for example an existing enterprise agreement, a client who has prohibited AI use on their matters, or a pilot the partnership has already approved.}
- Apply the SRA Standards and Regulations, the SRA's warning notice "Misuse of AI" dated 17 August 2026, and the SRA's effective supervision guidance. Work from the text of those documents where you have it, and mark as unverified anything you believe to be in them but cannot quote.
- Do not invent regulatory requirements, SRA expectations, enforcement outcomes, case law or product capabilities. If you do not know what the platform does, write the question rather than an assumption.
- Do not conclude that the platform is or is not compliant, adequate or safe. The purpose is to produce questions the firm can put and a record of the answers.
- Write in plain English and keep the questions short enough that a vendor cannot answer them with a brochure.
- This is a procurement planning aid, not legal or compliance advice. The COLP, the partnership and the firm's managers remain responsible for the decision and for everything done under it.How did we do?
Hit reply and tell me what you would like covered in future issues or any feedback. We read every email!
Thanks for reading,
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!
