Your people are already using these tools. Some have permission and a firm account, others a personal login and a habit nobody has looked at, and the firm answers for both. Training is the step that decides which way that goes, and the one most firms skip, because writing a policy feels like the finished job. A policy states the rule. Training builds the habit that keeps the rule alive at five o'clock on a Friday when a draft has to go out. What follows sets out why the regulator now reads your training record as evidence, what each group in the firm needs taught, and how to tell afterwards whether the work changed.
Why training became a regulatory expectation
On 17 August 2026 the Solicitors Regulation Authority issued a warning notice on the misuse of AI. A warning notice is not guidance. It sets out the conduct the regulator expects and states that a firm failing to have proper regard to it risks disciplinary action. The SRA published it alongside the figure explaining the timing, being 42 reports of possible AI misuse received between July 2025 and July 2026, and it names the parts of the Codes already in play, being competence, effective systems for supervising client matters, governance and controls, and confidentiality.
Three days later came the regulator's annual assessment of continuing competence. It reviewed 123 training records from criminal and civil practitioners across 66 firms, and almost half failed to explain how the learning need had been identified and why it mattered. The assessment also records that the regulator now looks at whether a solicitor is aware of and uses warning notices and guidance when planning their learning, which makes the AI notice the newest item your record is read against. The SRA has since consulted on a rule requiring every solicitor to keep a record showing how learning needs were identified and addressed, intended to start with the 2027/28 practising year.
On 3 September 2026 the Solicitors Disciplinary Tribunal published its judgment against Abhishek Kumar, recording that this was the first time a lawyer's use of artificial intelligence in legal proceedings had been litigated before it. Kumar filed a written Answer on 12 March 2026 containing citations that were wrong and authorities relied on for propositions they did not support. When the regulator put that to him on 8 April and pointed him to Ayinde v Haringey LBC [2025] EWHC 1383 (Admin), his reply the next day was drafted with the same kind of tool and carried fresh errors. He told the regulator he did not have the expertise to verify the output, and when asked for copies of the authorities he was unable to produce them. He was struck off. The defences he ran, that he was unaware of the regulatory warnings and that he lacked the skill to check, are the two a partner in your own firm is most likely to reach for.
What staff need to understand before they touch a tool
Teach the why before the how, because people follow rules they understand. Four things belong in the first session whatever the attendee's role. The first is how these tools work. A large language model generates text by predicting what tends to follow what. It holds no index of real cases and no sense of truth, so having seen a great many citations it will produce something shaped exactly like one with no guarantee the case behind it exists. Confidence in the output is therefore no signal of accuracy, since a model writes a fabricated authority in the same assured tone it uses for a real one.
The second is the scale of the failure. An AI hallucination case tracker recorded about 1,598 cases worldwide as at 9 June 2026, of which 59 were in the United Kingdom, and most involved fabricated case law. Of the 543 UK legal professionals LexisNexis surveyed for a report published on 2 September 2026, 83 per cent remained worried about inaccurate or fabricated information and 77 per cent believed AI is leading clients and the public to misread the law.
The third is where the words go. Text typed into a tool leaves the firm and reaches the provider, and consumer products will often store it and use it to train future models. Business tiers usually stand on different terms, where inputs are not used for training and data sits in a defined region, and that gap is often the gap between a breach and a safe workflow. The warning notice expects contractual, technical and organisational safeguards before client material enters any tool. Staff need to know that assessment has been run and that the answer is the approved list.
The fourth is the firm's own policy, taught as a working document rather than circulated as an attachment. Which tools are approved and for what, what must never go into them, who checks output before it leaves, and what to do when something goes wrong. If you have no policy yet, the guide to writing one sets out the sections it needs.
A curriculum by role
One session covers those four points for everybody. What follows them differs, because the duties differ and so does the work.
Partners and supervisors
Supervisors carry the sharpest new exposure and many have never been told what to look for. The warning notice shuts a door some firms had leant on, because the supervisor of a junior who used the tool answers for the output as fully as the person who wrote the prompt. Teach them to check the work rather than the tool, watching for confident wording that misstates the law, citations that do not exist, and summaries that miss the point.
Partners need one thing beyond that. The UK Jurisdiction Taskforce, chaired by Sir Geoffrey Vos within LawtechUK, published its Legal Statement on Liability for AI Harms on 7 July 2026, concluding that English private law as it stands can resolve most disputes about harm caused by AI. For professionals it runs both ways. A solicitor can be negligent for using AI badly, by selecting an unsuitable model or sending work out unchecked, and can equally be liable for failing to use AI where a competent member of the profession would have done so.
Fee earners
Drafting is where the saving is felt first and where the boundary needs drawing precisely. Routine correspondence, first drafts of standard clauses, plain-English summaries and reformatting sit on the useful side. Anything turning on the precise facts, the current state of the law or a client's instructions does not, because a model produces confident wording that misstates the position. Treat the draft as a trainee's first attempt, being quick, useful and never final, then read it against the file and the law, correct what is wrong and record the review.
Research needs its own half hour, and the distinction that matters is what the tool is built on. A general chatbot holds no reliable index of cases or statutes and can invent them, while a tool built on a known legal database works from real sources and links to them so the authority can be opened and read. Even then the summary is a pointer rather than the answer, and one that is ninety per cent right is still wrong where it counts.
Trainees and paralegals
Three quarters of the lawyers LexisNexis surveyed think junior lawyers who use AI heavily will struggle to develop legal judgement, and one in ten said they are dependent on the tools. Keeping juniors away from them is the wrong answer, since it trains them for a job that will not exist. Set the order of work instead. On the tasks that teach most, being a first piece of research, a first advice and a first witness statement, the junior drafts unaided and runs the tool over the draft afterwards to compare. Write the sequence down, because a habit living only in a supervising partner's head goes on holiday when the partner does.
Support and business staff
A policy covering only fee earners leaves the widest gap in the firm, since administrative staff, consultants and temporary hires all touch client information. Teach this group the confidentiality material and the approved list in full, then add the AI note taker, the tool they are most likely to introduce without anyone deciding. Many note takers capture the audio, transcribe it on the provider's servers and store the transcript, which for a client meeting is a disclosure of confidential and often privileged material to a third party. Ask where the recording is processed and stored, whether it trains the provider's models, who can reach it, and whether a data processing agreement meets the UK GDPR.
How to run the sessions
A long seminar is forgotten by Friday, while short sessions built around the firm's own work stay with people. Half a day for up to twenty attendees works, because everyone takes part and nothing has to be cut short.
Build the middle from live exercises using documents off your own files. Put a real output in front of the room and have attendees open the authority behind each citation and read the passage. Take a letter that went out last month and ask its author to explain the advice with the tool closed. Send them away with a workbook and a short action card naming what they will do differently over the next thirty days, which is the shape Adliora's Staff AI Literacy Session runs to.
Three things to avoid. Do not make it a lecture on the technology, because the subject is the work, and attendees should leave knowing which tools they are permitted to use, for what and with which checks. Do not overfill the room, since a session of forty becomes a broadcast. And do not treat a single induction as the end of it.
Verification drills and the check before filing
The habit worth drilling above all others sits between the AI draft and the thing that leaves the building. The SRA's requirement is flat, that every authority put before a judge is checked as genuine and accurately cited before filing. The Irish High Court's practice direction, in operation from 1 September 2026, adds the point English firms should adopt as their own, that running an output through the same or another AI tool does not amount to checking it.
Drill it as a fixed step rather than as an attitude. Name who performs the check, decide how it is recorded, and refuse to let anything go out until it is done. It has to be firm-wide rather than personal, because the lawyer who drafted the passage is the one least able to see the fabrication in it, the confident phrasing having already persuaded them.
Extend the drill beyond court documents while the Kumar judgment is fresh, because his second failure happened in an email to the regulator rather than in a filing. Size is no protection either, since on 10 July 2026 Legal Futures reported that the Crown Prosecution Service had acknowledged putting cases that do not exist before the High Court in an extradition appeal, having not verified the authorities relied on in formal submissions.
Supervision after the training
Training sets the standard and supervision keeps it. Staff need to know which tools they are permitted to use, for what, and what they must never do, such as putting client identifiers into an unapproved tool. Supervisors need to know which of their people use which tools, which sounds administrative until you notice that a firm unable to say who uses a tool cannot show it supervises the use. A named authorised person stays accountable for AI-assisted work as they are for the work of a junior. The judiciary's own AI guidance models the same discipline, keeping confidential information out of public systems, insisting on checking, and fixing responsibility on the person whose name the work carries.
Measuring whether the training worked
Most firms know who attended and almost none know whether the work changed. Research from BARBRI, reported across the trade press on 6 August 2026 and resting on interviews with ten learning, knowledge and innovation leaders at firms from the largest American partnerships to newer technology-led practices, found firms marking their own technology rollouts at a middling seven out of ten. None had built a framework saying what competent AI use looks like at each stage of a lawyer's development.
A firm of eight has neither an innovation team nor a learning team, so the question belongs to nobody. Adoption is rarely even either. Two fee earners use the tool daily, one uses it for a single task, and the rest opened it in the first week and went back to what they knew, which is a training problem, a workload problem or a tool that does not suit their work.
Measure one task rather than the whole firm. Pick something done often that the tool is meant to help with, and write down how it works today and roughly how long it takes, because a baseline taken after the tool arrives is worth nothing. Then count three things for ninety days, being how many matters used the tool on that task, the time from instruction to a usable first draft, and how often the supervising fee earner sent the draft back. At ninety days, widen it, fix the training, or drop the tool at renewal.
Keeping the competence record the SRA reads
An hour of writing turns the training into evidence. For each person using a tool, record what changed in their work, what that exposed as a gap, what they did about it and whether it worked. It is knowing when confident output is wrong, what is safe to put into a prompt, and who checks before the work leaves the building. Read the warning notice against what your people do each day, take the learning needs it exposes, and put them in the record with the date and the reason they arose.
What the training costs
Money for AI is no longer the adventurous line in a law firm's budget. The annual survey by the accountants Crowe, run with the Institute of Legal Finance and Management and reported on 16 September 2026, found four in five firms had committed specific funds to AI-based solutions, across 39 firms turning over between £1.7m and £69m. With 88 per cent of those firms weighing pay rises of no more than 5 per cent, your AI line competes with your people. Name the number the spending is meant to move before you choose the product. Adliora's session carries an indicative fee of £1,850 plus VAT, fixed in writing before any work starts.
Common questions
Does AI training have to be formal to count?
The SRA does not prescribe a format. What its 2026 assessment of continuing competence marked records down on was the absence of any explanation of how a learning need was identified and why it mattered, which is what turns a list of attended sessions into evidence of reflection. A short session run around your own files, written up with the reason it was needed and what changed afterwards, carries more weight than a certificate from a webinar nobody remembers.
How long should a session be and how many people should be in the room?
Half a day for up to twenty people works well, because everyone in the room takes part and nothing has to be cut short. Beyond twenty, run a second session rather than a larger one. A long seminar is forgotten by Friday, while short sessions built around the firm's own work, with real examples of good and bad use, stay with people.
Should trainees and paralegals be kept away from AI tools until they qualify?
No, because that trains them for a job that will not exist. Set the order of work instead. On the tasks that teach most, being a first piece of research, a first advice and a first witness statement, the junior drafts unaided and runs the tool over the draft afterwards to compare. The comparison is the lesson, and where the model has invented something they learn it in your office rather than in front of a district judge.
How often should the training be repeated?
The tools change and so do the risks, so a single induction is not enough. Refresh the training at least once a year, fold AI into how new staff are brought on, and repeat it sooner when you adopt a new tool or the regulator publishes something that bears on the work. The SRA now assesses whether solicitors are aware of and use warning notices when planning their learning, which ties each refresh to the training record.
Where to go next
If your people already use these tools and nobody has taught them the rules, the Staff AI Literacy Session is the half day that fixes it. Where you do not yet know what is being used and by whom, start with the AI Readiness Audit. The guide to writing an AI use policy covers the document the training puts into practice, and the contact page is where to ask which of the three your firm needs first. Adliora Limited is a management consultancy rather than a law firm, and where a question needs formal legal advice you should take it from your regulated adviser.
