AI for Barristers: A Chambers Checklist for Casework and BSB Rules

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI for Barristers: A Chambers Checklist for Casework and BSB Rules.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI for Barristers: A Chambers Checklist for Casework and BSB Rules.

A barrister can use AI for casework only after checking that the tool and the proposed input protect confidentiality, then independently verifying every fact, authority and drafted argument. Chambers should approve use cases, access and record keeping under the BSB's current guidance. A case summary or skeleton argument remains the barrister's professional work, with a human review before it leaves chambers.

The BSB's guidance, valid from 18 May 2026, is organised by stage: competence, checks before adopting a tool, management and use procedures, use on client work, and what to do when others use AI. It also asks barristers to weigh three sources of risk: the application, the manner of use and the tool itself. That makes a one-page “AI allowed” policy too vague. A practical checklist should say which tool is approved for which material, what evidence the barrister keeps and who stops a draft when a source or confidentiality question is unresolved.

Follow me on Instagram@sagnikteaches

Set the chambers boundary before opening a brief

List permissible tasks by risk. Low-risk work can include learning with invented facts or editing a general administrative notice. Higher-risk work includes summarising a live bundle, reviewing witness evidence, researching law and drafting a skeleton argument. For each task, record the approved tool, allowed input type, prohibited input type, reviewer and output location. Do not let a broad procurement approval imply that every client document may be uploaded. A tool suitable for formatting non-confidential text may be unsuitable for privileged instructions.

Connect on LinkedInSagnik Bhattacharya

The BSB guidance says using AI tools counts as outsourcing a support service under rule rC86, whether the work is legal or administrative, and that free, publicly available systems are unlikely to meet the outsourcing conditions. Chambers should examine contractual promises, security, storage, access, deletion and provider use of prompts or files. Ask what happens if the provider stops operating or changes its terms. Check access inside chambers too: barristers with opposing instructions must not be able to see or infer another's confidential prompts. A shared workspace with open search across all matters is a particular concern. The legal AI buying checklist can inform vendor questions, but chambers still needs its own matter access design.

Subscribe on YouTube@sagnikteaches

An illustrative policy line reads: “Approved task: format a case chronology from documents stored in the approved matter workspace; access limited to instructed barrister and named support staff; output marked draft; every event checked against page and paragraph; no shared training collection.” That is operational. “AI may assist casework where sensible” gives staff no test for whether a bundle, pupil exercise or client instruction belongs in a tool.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Classify the material before upload

At the start of each matter, separate public authorities, chambers administration, client instructions, privileged communications, witness material and sensitive third-party information. A case name removed from a document does not necessarily make it safe: dates, rare facts, quotations or the combination of parties can identify it. For a live matter, decide whether the proposed tool and contract meet the client's confidentiality needs. Record any client-specific restriction or court direction. If those checks are incomplete, work without the tool until they are resolved.

Use the least information necessary for the task. To draft a hearing checklist, the model may need hearing type, generic tasks and dates but not witness names or the whole bundle. To format a chronology, it may need selected documents in a restricted system with a clear source index. Do not upload a complete brief merely because a product accepts large files. Retain a list of documents supplied so that the reviewer knows what the AI did and did not see. This protects against a summary that appears complete but omits an unseen annex.

Suppose a barrister wants a ten-line chronology from 120 pages of correspondence. The approved input is 12 relevant letters and a document index, not every confidential email in the matter. The output should list each event with the letter date and page. If an event lacks a source, it remains a question. A generic assistant that keeps the documents under unclear terms is not a suitable workaround. The BSB specifically warns that free general-purpose systems are usually too high risk for client-specific legal work.

Require a source column in every case summary

Ask an approved tool to produce a two-column draft: statement and source location. For disputes, include a third column for whether the statement is agreed, alleged or disputed. This prevents a witness allegation becoming an established fact in the summary. Have a barrister or authorised reviewer compare each row with the actual bundle page, including the text around a quoted passage. A summary is useful when it compresses reading without changing evidential status; it is harmful when it strips out who said what.

From the documents approved for this matter, draft a chronology.
For each event give date, neutral description, source document and page.
Label contested statements as allegation or dispute, not fact.
Do not infer events between documents or resolve inconsistencies.
Put unreadable or missing source material in a separate question list.
Do not include material outside the approved document set.

Run on that set, the tool might return (illustrative): “12 March: the defendant received the notice, bundle page 36.” If page 36 merely shows an email sent to an address, receipt is not established. The corrected entry is, “12 March: the claimant sent an email headed notice to the defendant's recorded address; receipt disputed.” That change matters to the argument. The barrister checks the document and, if necessary, the surrounding correspondence. The model should not decide an evidential issue by making the sentence sound cleaner.

Check omissions as carefully as errors. If the summary includes nine of ten important documents, the missing one may change the chronology. Compare the output with the approved index and mark each document reviewed or intentionally omitted. For longer bundles, sample across document types and dates rather than only the first pages. The intake-notes chronology tutorial applies the same one-source-per-event discipline to a family matter's first meeting notes.

Keep skeleton drafting behind a legal research gate

A skeleton argument has factual propositions, legal propositions and submissions. Check each layer separately. First verify factual statements against the bundle and instructions. Then open every cited case, statute or rule in an authoritative source, check the exact passage and current treatment, and decide whether it supports the proposition. Finally assess whether the argument fairly presents adverse material and fits the client's instructions and court directions. The BSB guidance says AI outputs should not be incorporated without independent verification, that relying on AI without standard legal research processes may breach rule rC9, and that a barrister must not knowingly or recklessly mislead the court through AI use.

In a sample draft, AI writes, “The authority establishes that a late notice is always invalid.” The actual decision may concern one type of notice, a particular time limit or a rejected submission. Replace the sentence with the narrower supported point or remove it. A citation that exists is not enough. A citation ledger should show proposition, source passage, later treatment and reviewer. If a source cannot be located, do not use it. The citation-checking workflow gives a fuller method for that ledger.

Do not ask AI to write a witness's own account from scattered notes and then present it as their words. A witness statement is meant to record the witness's own account, and AI-polished wording can drift from what the witness actually said. If the system helps organise issues for conference, keep its output separate from evidence. A barrister must decide the proper use of each source and follow the relevant procedural rules. A model's polished narrative can conceal a contradiction that cross-examination will expose.

Make an individual-use check beside the chambers policy

A chambers policy is the ceiling for permitted use, not automatic permission in every matter. Before each client-specific use, ask: Is this tool approved for this task? Is the material within the permitted class? Does the client or court impose a restriction? Who can access the input and output? What must be verified? Where will the audit record live? Record a short answer beside the matter ID. If the answer changes when new documents arrive, reassess before another upload.

Consider a public-law matter where an authority bundle is public but the instructions include confidential health information. The barrister might use an approved tool to organise the public authorities while keeping the client-specific instructions out. The same tool may be unsuitable for a combined summary of both. Splitting the task is a practical way to retain useful drafting help without assuming one permission covers every data class. If the model later produces a case-specific conclusion without access to the instructions, treat it as incomplete.

Write a stop rule into the matter checklist: “If the output includes an unsourced fact, unlocated authority, unexpected personal detail or apparent cross-matter information, stop using it, preserve the record and investigate.” A cross-matter detail could signal an access or workspace problem; it deserves more than a sentence edit. The reviewer should know whom to notify in chambers and how to suspend the tool for the affected work. The AI error response tutorial covers customer-facing corrections; casework errors require chambers' own professional and confidentiality route.

Pilot on a small bundle with a timed review

Take an illustrative chambers pilot: two barristers and a clerk test an approved tool on a synthetic 40-page bundle containing 22 dated events and six authorities. They spend 30 minutes preparing the document index, 20 minutes generating a chronology and issue list, 55 minutes checking each event and 35 minutes checking authorities. The total is 140 minutes. They find three date errors, two allegations stated as facts and one authority cited for a proposition it does not support. These are test figures, not claimed productivity gains.

The pilot should score accuracy by type, not give the draft a general “good” rating. A date error may be corrected quickly; an allegation converted into fact shows a deeper instruction problem. The team changes the prompt to require disputed-status labels and source pages, then repeats with a different synthetic bundle. If the same error recurs, keep that task human-led or narrow the tool to document indexing. Compare total preparation plus review time with the current method, including any rework. Speed alone cannot justify a weaker duty-to-court control.

After a successful synthetic test, chambers can consider a limited live task with the approved tool and a named barrister, subject to the matter's confidentiality assessment. Start with an internal chronology, not a court-ready argument. Keep the original source, output, corrections and final document. Audit whether staff followed the checklist without informal shortcuts. A pilot that works only when the technology lead watches every step is not yet a reliable chambers procedure.

Check what a client or instructing solicitor needs to know

Make transparency a matter-level question rather than a generic footer. The barrister should consider whether the proposed use, data handling or effect on the work calls for discussion with the client or instructing solicitor under the current duties, instructions and contract. The chambers policy can provide a short explanation of approved uses and safeguards, but it cannot decide every client's concern in advance. A client may have prohibited a particular provider, or the brief may contain information supplied on a limited basis. Record the answer before use, especially when the tool will process client-specific documents.

As an illustration, take an instructing solicitor who sends a bundle with an instruction that witness records remain in the firm's document system. A barrister should not upload the bundle to a chambers AI workspace simply because chambers has approved that platform in general. Ask whether a confined task on public authorities is possible, or seek a suitable authorised route for the confidential material. Record the agreed scope. The practical outcome may be that AI organises public law while the witness documents are read conventionally.

Do not claim that disclosing AI use solves an accuracy or confidentiality problem. A client can agree to a tool and still expect competent work; a court can require accurate authorities regardless of how a draft was made. Keep the discussion specific: task, information involved, provider controls, human review and any remaining risk. If the answer is unclear, obtain professional advice within chambers and review the relevant BSB material. The decision should be made by the responsible practitioner, not by a default setting in a document template.

Expect AI material from other participants too

The BSB guidance also addresses situations where clients, solicitors or other participants use AI. A barrister may receive a client chronology, draft argument or research list whose sources were never checked. Treat it as supplied material requiring assessment, not as verified because it arrived through an instructing professional. Ask for original documents and citations. Mark unknown provenance in the working note, and avoid incorporating a confident passage into a skeleton until its source and context have been checked.

An illustrative client sends a six-page “case law summary” with four decisions and no source links. The first two are real but discuss a different issue; the third appears to be a commentary page; the fourth cannot be located. The barrister should tell the instructing team what was checked, what remains unverified and what cannot be relied on. A useful response might request the original cases rather than asking AI to rewrite the summary. The source of the material matters less than whether the barrister can stand behind the final proposition.

A clerk may also receive an AI-written hearing timetable from another participant. Check dates against the actual order, court notice and chambers diary. If the timetable differs, escalate rather than choosing the version that looks newest. This is a simple example of a general rule: incoming AI output can influence case preparation even if chambers did not generate it. The checklist should cover intake of external AI material, with an owner for verifying facts and authorities before they enter the matter record.

Keep review effort visible in the case budget

AI-assisted drafting has a review cost. The chambers pilot should record time for document preparation, generation, source checks, legal analysis, editing and final proofing. If a 20-minute draft takes 90 minutes to verify and still leaves source gaps, the task may not be worth automating. If a chronology consistently saves clerical sorting while retaining reliable page references, it may be useful. Use the same quality standard as work prepared without AI and agree any client-facing billing treatment through the normal professional route.

Over a sample week (illustrative figures), a barrister spends 25 minutes generating two internal summaries, 50 minutes checking 30 source references and 15 minutes correcting three disputed-status errors. The 90-minute total is the figure to compare with the existing process, not the 25-minute generation time. The errors also indicate where the prompt or document index needs work. A good policy should allow the barrister to conclude that a tool is unsuitable for a particular task without being judged only on how often it is used.

Write the policy as roles, records and review triggers

Assign procurement review to a chambers manager or designated group, matter-use decisions to the instructed barrister and technical access administration to a named person. Decide how pupils and clerks may use approved tools and who supervises their work. The policy should specify where prompts, inputs and corrections are recorded, when they are deleted and how an output is marked as draft. The BSB guidance lists governance procedures to consider: an explicit AI governance policy (particularly for chambers-level systems used by several barristers and staff), settings tailored to the use, prompt and data deletion policies, and AI audit logs showing how the outputs you rely on were produced.

A filled illustrative checklist could read: “Matter 24; use: chronology from 12 approved letters; tool: chambers-approved restricted workspace; data class: confidential; access: instructed barrister and clerk; output: 18 events; source checks: 18 complete; disputed-status corrections: two; legal citations: none; approved for internal conference note by [barrister] on [date].” That record shows what was done. A generic “AI used responsibly” note would not help a later reviewer reconstruct the decision.

Set review triggers: a vendor changes terms, a new feature shares prompts, chambers changes membership, a client imposes a restriction, a court direction changes, or an output error suggests systemic weakness. Review the BSB and Bar Council guidance periodically. The barrister still owns the professional judgement in each case. Chambers' role is to make safe choices possible and visible, rather than relying on each person to remember every contract setting and record step under deadline pressure.

Questions chambers should settle before use

Can a barrister put a brief into a free general AI account?

Do not assume so. The BSB's 2026 guidance says free general-purpose systems are usually too high risk for client-specific legal work and calls for scrutiny of contractual terms, access, retention and confidentiality. Chambers should approve a tool and permitted use case before any live brief is uploaded; a barrister remains responsible for the client's information and the output.

Does anonymising a case summary make an upload safe?

Only if the remaining facts cannot identify the matter or reveal confidential information in context. A rare fact pattern, date, role or quoted document can make a supposedly anonymous summary recognisable. Use synthetic facts for learning exercises. For live work, apply the chambers policy, tool approval and client-specific confidentiality assessment rather than relying on names being removed.

Who checks an AI-assisted skeleton argument?

The barrister responsible for the work should check every factual proposition against the bundle, every legal proposition and citation against primary sources and later treatment, and the final document for fairness and accuracy. Chambers can provide process and approved tools, but it cannot transfer the barrister's professional judgement to the system or a generic sign-off tick.

Further reads

Sources: Bar Standards Board Guidance on the use of Artificial Intelligence and Other Technologies, valid 18 May 2026; BSB regulatory expectations of chambers; Bar Council, Considerations when using ChatGPT and generative AI software based on large language models (updated November 2025). Verify current BSB Handbook and court directions for each matter.

Want a chambers AI checklist people follow?

On a 1:1 call we can map approved tools, matter-level checks and a review trail around your chambers' casework.

Book a 1:1 call with me