White paper · Legal teams
AI agents inside
the legal department
Where the files sit, who approves, and what each statement rests on
In brief
- Legal is the department where a wrong sentence carries a signature. A clause goes to the other side. A citation goes to a court. A document goes into a disclosure. In each case a named lawyer answers for it, whatever tool wrote the draft.
- The suppliers handle that fact differently, and in public. One vendor’s document says: “No output should be relied upon or implemented without review by a qualified legal professional.” Another promises “approved redlines in a matter of minutes, without your GC on hand”. A trade report on a third says its product “runs autonomously by default, without human intervention”. A published workflow template sorts confidentiality agreements into “AUTO_APPROVED versus LEGAL_REVIEW”. Hosting separates them less: one legal AI vendor offers an on-premises option in so many words.
- The court record explains why sources matter. A public database of decisions about invented citations and other AI-made material listed 2,313 cases on 20 November 2026. An English court has said the profession can expect it to ask whether those who lead lawyers did their part. A New York court held in February 2026 that documents a defendant made with a consumer AI tool were not privileged, in the circumstances before it.
- This paper is about the legal department’s own work: the reading, comparing, checking and drafting done at the front door, in contract review, on obligations and renewals, in disputes, in research, in corporate work, on privacy requests, in compliance and in legal operations. It sets out what an agent can read, how it reaches legal systems, what it may change without a person, and what the published rules on professional duty, courts, privilege and decisions about people already say. It ends with seven questions to put to any supplier, and with a place to start.
- The position throughout is simple, and it has three parts. The files sit on infrastructure the company controls. A named lawyer approves before anything leaves the team. Every statement an agent makes is shown with its source. No agent gives legal advice, negotiates, sends anything to a counterparty, signs, files, decides what is privileged or predicts how a matter will end. Whether privilege holds over anything is a question for the company’s own lawyers, and this paper does not answer it.
Where legal departments are today
- 2,313cases in a public database of court decisions on invented citations, at 20 November 2026
- $5,000fine on two lawyers and their firm in New York in 2023, for filing opinions that did not exist
The suppliers have arrived, and most began with the contract. Vendor pages now describe agents for contract review, redlining, intake and research. Microsoft introduced a Legal Agent in Word on 30 April 2026 [3]. Suppliers publish their own figures for time saved. We found no named, dated, independently measured result for a legal agent, and this paper quotes none as proof.
They disagree on how much a lawyer sees. This is the clearest difference in the published material.
- Microsoft’s transparency document for its legal features says: “No output should be relied upon or implemented without review by a qualified legal professional.” Its Word agent asks the user to “review all edits before approving” [3].
- Harvey’s guide to contract review says: “The attorney still decides what matters, what needs escalation, and how the organization should proceed” [4].
- Juro describes the difference between an assistant and an agent in its own words: “instead of recommending actions, it takes them”. It offers “approved redlines in a matter of minutes, without your GC on hand” [5].
- Luminance sells a product it calls “Autonomous Negotiation”. Its page says the product will “route non-standard terms and exceptions back to Legal” [1]. A trade report of March 2026 said that the system “sends revised drafts to counterparties”, and that it “runs autonomously by default, without human intervention”, with people added at any stage if the company chooses [2].
None of these is wrong for every company. They are different answers, and a general counsel should know which one the business is buying.
Hosting separates them less than buyers expect. LEGALFLY’s own comparison page lists “optional on-premise or single tenant deployment for regulated industries” [6]. Harvey’s security page says: “Harvey hosts its cloud environment in Microsoft Azure”, with processing offered in named regions [4]. Microsoft says its agent “runs within Microsoft 365 security, compliance, and governance controls” the customer already uses [3]. Several of the other pages we read do not say where the agent runs. So where it runs is still a question to ask. It is not the only question that tells suppliers apart.
The court record is about citations nobody checked. In 2023 a New York court fined two lawyers and their firm 5,000 dollars for filing opinions that did not exist. The judge wrote that “existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings” [11]. Two years later an Alabama court went further than a fine. “Fabricating legal authority is serious misconduct that demands a serious sanction,” it said, and it disqualified the lawyers from the case and referred them to the state bar [11].
The count keeps rising. A public database of such decisions said “2313 cases identified so far” when last updated on 20 November 2026 [9].
Tools built for lawyers are not exempt. Researchers at Stanford tested legal research products in 2024. Their summary says two systems “produced incorrect information more than 17% of the time, while Westlaw’s AI-Assisted Research hallucinated more than 34% of the time” [13]. A statement with no source shown cannot be checked quickly.
One ruling shows how privilege can turn on the tool. On 10 February 2026 a judge in the Southern District of New York ruled in United States v. Heppner, with a written opinion a week later. A law firm’s note on the ruling says documents made with a consumer version of an AI tool “were not protected by the attorney-client privilege or the work-product doctrine under the circumstances presented” [14]. The defendant had used the tool by himself, without his lawyers’ direction. The same note says the decision leaves open how a tool under contract, used at a lawyer’s direction, would be treated. We read two law firms’ notes, not the opinion.
What the published accounts leave out. We found no public account of a company’s legal team leaking contract data through a public AI tool, and we do not claim one. We found no vendor page that says, in its own words, that its agent sends to a counterparty with no person involved.
The work nobody has packaged
- One of threeworkflow templates we re-opened waits for a person to approve or decline
- Hostedservices are what all three are built from: model, file store, spreadsheet, mailbox or chat channel
We listed the jobs an in-house legal team does and looked for each in the template libraries and prebuilt agents of the agent and workflow platforms. They are thick around the incoming contract. They thin out wherever the work is a check before a lawyer signs.
Where templates or prebuilt agents exist
- A contract read, scored for risk and sent by email for approval
- A confidentiality agreement scored against a playbook, then filed as approved or sent to legal
- A contract’s clauses rated high, medium or low, with an alert above a set score
- A summary of a confidentiality agreement posted to a team channel with a recommendation
- Risky clauses found and alternative wording suggested
- A legal request given a predicted category, with the transfer confirmed by a person
Where none was found
- A signed contract turned into a list of obligations, each with its clause
- The final version checked against the approved one before signature
- Each citation in a draft checked against the library the team holds
- Documents due for disclosure marked for a lawyer’s look at privilege
- A notice to preserve documents drafted from the description of a dispute
- A law firm’s invoice read against the billing guidelines
- The group’s company records compared with the public register
Two findings matter more than the list. First, the templates differ on whether a person approves. Of the three workflow templates we re-opened, one “Waits for the recipient to approve or decline”. One decides “AUTO_APPROVED versus LEGAL_REVIEW” by rule, files the low-risk agreements as approved, and emails the legal team only “when legal review is required”. One has no review step on its page and is described as “Ideal for small teams without dedicated legal counsel” [7]. The score that separates the two paths is a setting. Whoever sets it has written a legal sign-off rule, whether or not the general counsel knows.
Second, all three are built from hosted services: a hosted model, a hosted file store, a hosted spreadsheet, a hosted mailbox or chat channel. That settles where the contracts go before anyone has asked [7]. One prebuilt workflow shows the other way: it predicts the category of a request and then shows “options to confirm or reject the transfer” [8].
The gaps sit where the reading is slow and the signature already belongs to a named lawyer. That makes them the safer place to begin.
Three doors, one gate
An agent is only as useful as what it can reach. In a legal department that means three kinds of thing.
Documents. The playbook, the template library, the clause library, the company’s policies, past advice, signed contracts, the delegation of authority. These are indexed once as knowledge sources, with each reader’s permissions kept, so an agent answers with the passage it relied on and says when the documents are silent.
Systems of record. The contract store, the matter list, the signature system, the entity register. Most are reached through an interface the company already runs. The company gives the platform the description of that interface, chooses which operations become tools, and decides which of them need approval.
Closed figures for reporting. Legal spend, budgets and the count of requests, held in a reporting database, and read only.
All three pass one gate: every tool call is checked against the company’s policy, which can allow it, record it, hold it for a person, or block it. Where the company’s interface checks the caller, the agent acts as the person using it and cannot claim another.
Six ways into a legal department’s systems
| Route | What it reaches | What to know | |
|---|---|---|---|
| 01 | A file the user uploads | A contract, a counterparty’s markup, a pleading, a law firm’s invoice, a data subject’s request | Nothing is connected. The safest start |
| 02 | A knowledge source | The playbook, templates, policies, past advice, signed contracts | Permissions are kept per reader |
| 03 | An imported description of an interface | The contract store, the matter list, the signature system, the entity register | You choose the operations. Changes wait for approval |
| 04 | A reporting database | Legal spend, budgets, request counts | Read only |
| 05 | A server from the tool catalogue, run in your cluster | Document stores, web search, document conversion | It runs inside; check what each tool does |
| 06 | Public sources | Regulators’ publications, legislation, company registers | Reads only, from the publisher |
- 103public interface descriptions read, from 15 publishers; 101 import
- 2,363of the 4,432 operations in the 101 that import are sent as changes
What is published. We read 103 public interface descriptions from 15 publishers and ran each through the platform’s importer. 101 import. 2 are refused, because each holds a value the importer cannot read. None is over the size limit. The 101 carry 4,432 operations, of which 2,363 are sent as changes, and the method alone cannot tell which only read [10].
What that exercise taught us. The open interfaces around legal work are built to do things: send an agreement for signature, void an envelope, change a record, delete a file. More than half of the operations are sent as changes, and some of those only read. The one publisher whose operations all read is a national legislative library. So choose the operations at import, read what each one does, start with the ones that only read, and add nothing that sends, signs or deletes until the approval step has been seen working.
What is not published. The 15 publishers are e-signature, contract, e-discovery, privacy-management and document-storage systems, one practice-management system, a company registry and a legislative library. The document management systems most legal teams file in publish no description a file address reaches. Nor do the matter and legal-spend systems, the entity management systems or the trade mark and patent offices we looked for. For those, the description comes from the company’s own system, or the agent reads a report or an uploaded file.
Signing and sending. The safe design keeps the agent on the preparing side. It holds no signing authority and no outbound mailbox, and it sends nothing to the other side.
What an agent may do
One rule runs through every agent in this paper: it prepares, and a named lawyer decides.
- It reads what the person using it could read, and no more.
- It drafts the markup, the summary, the chronology, the reply, with each statement tied to its document and page.
- It lists what departs from the playbook, what is missing, what changed between versions, and what falls due.
- It says when it found nothing. It does not fill a gap from memory.
- It stops before anything that leaves the team, changes a record or binds the company. A tool that changes something waits for approval.
- It decides nothing about a person, and nothing about privilege.
That gives 35 agents in nine areas: the legal front door; contract review; obligations and renewals; disputes; research and knowledge; corporate and transactions; privacy and data requests; compliance; legal operations. The full list is a separate document. Four lines are drawn tighter than the rest.
What goes to the other side. An agent may compare a contract with the playbook, mark each departure, read a counterparty’s markup and propose wording for each open clause. It may check the final version against the approved one and the signatories against the delegation of authority. It negotiates nothing, sends nothing and signs nothing. A lawyer decides each point, conducts the negotiation and releases the contract for signature.
What goes to a court or a regulator. An agent may summarise a case file, build a chronology with each event tied to its page, and check each citation in a draft against the library the team holds. It lists every citation it could not find, and every one that does not say what the draft says. It files nothing and predicts no outcome. The lawyer who signs the document answers for every citation in it.
Privilege and disclosure. An agent may mark the documents in a set that a lawyer should look at, with the reason: a lawyer is named, advice is asked for or given, a draft was prepared for a dispute. It withholds nothing and releases nothing. A lawyer decides what is privileged and what is disclosed. Whether using any tool affects privilege is for the company’s own lawyers, in each place the company works.
People. An agent may find the records in scope of a person’s request for their data, list what may need to be withheld and why, and draft the reply. It may set out what an internal report alleges and what the documents show. It releases nothing and draws no conclusion about a person. The privacy officer decides what is released. The investigator owns every finding.
What the rules already ask
This part needs a read by the company’s own lawyers before it is relied on. It reports what the published texts say; it is a summary, not legal advice. Nothing in it says that a platform makes a company or a lawyer compliant. Professional duties belong to the lawyer.
The lawyer answers for the work
Competence includes knowing the tool. The American Bar Association’s Formal Opinion 512 says that “lawyers must have a reasonable understanding of the capabilities and limitations” of the tool they use. On supervision it says: “Managerial lawyers must establish clear policies regarding the law firm’s permissible use of GAI” [15]. The opinion is written about law firms. Whether and how the Model Rules bind a given in-house lawyer depends on where that lawyer is licensed.
Whoever signs, certifies. Under Rule 11 of the United States Federal Rules of Civil Procedure, a lawyer who presents a paper to the court certifies, after “an inquiry reasonable under the circumstances”, that its legal contentions are “warranted by existing law” or by a proper argument for changing it [16]. The rule does not mention AI. It does not need to.
English courts look to those who lead. In Ayinde, the Divisional Court said of AI tools: “They may cite sources that do not exist.” It called for practical and effective measures by those with leadership responsibilities, and added that “the profession can expect the court to inquire whether those leadership responsibilities have been fulfilled” [12]. Its examples were heads of chambers and managing partners. A general counsel leads lawyers too.
Judges are told what to assume. Guidance for judges in England and Wales, dated 31 October 2025, says: “All legal representatives are responsible for the material they put before the court/tribunal” [17].
What this means for an agent. A draft is only as good as the lawyer’s means of checking it. An agent that shows the passage behind each statement, and lists what it could not find, gives the lawyer something to check. An agent that returns a confident paragraph gives the lawyer a risk to sign.
Confidentiality and privilege
The regulator in England and Wales has issued a warning. The Solicitors Regulation Authority’s warning notice of 17 August 2026 says: “Both paid for and free-to-use AI tools may not provide the contractual, and technical safeguards needed to maintain client confidentiality.” It tells firms to understand the safeguards that apply [18]. The notice is addressed to those the Authority regulates. Whether it reaches a given in-house team depends on how its lawyers are regulated.
Privilege was refused for a consumer tool. In Heppner the court, as one firm’s note reports it, looked at whether the defendant meant to get legal advice from the tool itself, not at whether he later shared the output with his lawyers [14]. The note describes the holding as limited to its facts. It is one trial court’s decision, and it is not a rule about tools a company runs for its own lawyers.
In one European setting, in-house advice is not privileged at all. In Akzo Nobel the Court of Justice held, in a competition investigation by the European Commission, that “legal professional privilege does not cover exchanges within a company or group with in-house lawyers” [19]. National courts and other countries take different views.
What this means for an agent. No tool makes a document privileged, and this paper does not say that any design keeps privilege. Where the files sit, who can read them and what terms govern the tool are facts the company’s lawyers will want when they answer the question. The answer stays with them.
Decisions about a person, and the AI Act
A person decides about a person. Under the GDPR a person has “the right not to be subject to a decision based solely on automated processing, including profiling” where it produces legal or similarly significant effects, with stated exceptions [20]. A reply to a data subject or a finding in an investigation should be a person’s.
A company’s own legal work is not on the high-risk list as we read it. The AI Act lists as high-risk “AI systems intended to be used by a judicial authority or on their behalf to assist a judicial authority” in researching and interpreting facts and the law [21]. A company’s legal team is not a judicial authority. Whether another entry catches a particular use is for counsel.
The dates have moved. Regulation (EU) 2026/1744 says the high-risk rules apply from “2 December 2027 as regards AI systems classified as high-risk pursuant to Article 6(2) and Annex III” and from 2 August 2028 for those under Article 6(1) and Annex I [22].
| Place | What we read | What it asks | |
|---|---|---|---|
| 01 | United States, federal courts | Rule 11(b) | The certificate follows “an inquiry reasonable under the circumstances” [16] |
| 02 | United States, one Texas federal court | Standing order of 9 April 2025 | Filings carry a “Certificate of Generative Artificial Intelligence Usage” [16] |
| 03 | New South Wales | Supreme Court Practice Note SC Gen 23 | “Such verification must not be solely carried out by using a Gen AI tool or program” [23] |
| 04 | Singapore | Registrar’s Circular No. 1 of 2024 | “the Court User assumes full responsibility for the output” [23] |
| 05 | Dubai, DIFC Courts | Practical Guidance Note No. 2 of 2023 | Parties should “declare at the earliest possible opportunity if they have used or intend to use AI-generated content” [23] |
We did not open the Heppner opinion, any state bar’s opinion, the Law Society’s guidance, court rules in Canada, India, Japan or elsewhere in the Gulf, or any country’s rules on who may practise law. Court practice notes change often. A company should check each with its own lawyers.
What the rules have in common
- The lawyer who signs answers for the work, whatever tool wrote the draft.
- Those who lead lawyers are expected to have rules for how the tools are used.
- Verification is done by a person, from the source.
- Confidentiality and privilege depend on the facts: which tool, on what terms, at whose direction.
A worked example: one mid-sized group’s legal team
Take a made-up industrial group: nine legal entities, about 2,600 staff, and a legal team of eleven. Nothing here is a customer’s story, and every figure is illustrative; it is section 4 put to work.
- At the front door, a sales manager sends a customer’s confidentiality agreement. The agent compares it with the standard positions, marks six departures and drafts the reply with the reason for each. A lawyer changes two points, approves and sends.
- In contract review, a supplier returns a 48-page agreement with its markup. The agent lists 31 changes: 19 it can match to wording the company has agreed before, and 12 that are new. A lawyer decides each one and conducts the negotiation.
- Before signature, the agent compares the final version with the approved one and finds one clause that differs. It checks the two signatories against the delegation of authority. A lawyer releases the contract.
- In a dispute, the agent reads 420 documents and builds a chronology of 96 events, each tied to its page, with three dates that conflict. The lawyer running the case checks each entry.
- Before a filing, the agent checks 27 citations in a draft against the library the team holds. It lists two it could not find and one that does not say what the draft says. The lawyer who signs corrects the draft.
- In legal operations, the agent reads a law firm’s invoice of 140 lines against the billing guidelines and lists 11 that do not fit, each with the rule. The matter owner queries them.
Every one of these reads documents or records the company already has, changes no record by itself, sends nothing outside the team, and leaves a record of who asked and what was shown.
Seven questions before an agent touches a contract
- Where does it run? On our infrastructure, in our private cloud, or in the supplier’s service? Where is the model? Does any step send contracts, matter files or advice elsewhere, and which?
- What can it read? Does it act as the person using it, with their permissions, or as one all-seeing account? Can a colleague outside a matter reach it through the agent?
- What can it send or change? Which tools send to a counterparty, start a signature, file, or change a record, and does each wait for a lawyer?
- Who approves, and is there a score below which nobody does? Show us every threshold, who set it, and who can change it.
- Does every statement show its source? Show us the document and the page behind a summary, and what the agent does when it finds nothing.
- What is recorded? Who asked, which tool ran, what was decided, for how long is that kept, and who can read it?
- Can we stop it? One agent, one tool, or all of them, at once, by us.
Where to start
Pick work where reading is slow, the decision stays with a lawyer, and nothing leaves the team on the agent’s word.
Legal policy questions and answers
Colleagues ask what the policy says and get the answer with the paragraph cited. No matter is touched.
Confidentiality agreement review
The agent marks up an agreement against the standard positions. A lawyer approves and sends.
Outside counsel invoice review
The agent lists each line that does not fit the billing guidelines. The matter owner decides.
One team, one kind of document, invented data first. Then beside the manual process, measured against it.
About Katonic AI
Katonic AI builds the platform that organisations use to build, run and govern AI agents on their own infrastructure: in their own data centre, in a private or sovereign cloud, or with no connection to the internet at all. The platform is in production with governments, banks, telecom operators and utilities. Katonic is certified to ISO 27001.
Sources
Read on 22 November 2026. Figures are the publishers’ own and were not audited by us.
- Luminance, product page for the product it names “Autonomous Negotiation”, undated, read 22 November 2026. https://www.luminance.com/autonomous-negotiation/
- IT Brief Australia, “Luminance opens AI contract tool with clearer reasoning”, 12 March 2026. https://itbrief.com.au/story/luminance-opens-ai-contract-tool-with-clearer-reasoning
- Microsoft, “Word: Legal Agent in Frontier”, Microsoft Copilot Blog, 30 April 2026. https://techcommunity.microsoft.com/blog/microsoft365copilotblog/word-legal-agent-in-frontier/4516218 ; and Microsoft Support, “Copilot for Legal: Transparency Documentation”, undated. https://support.microsoft.com/en-au/word/copilot-for-legal-transparency-documentation
- Harvey, “AI Contract Review: What to Look for and What to Avoid”, 29 May 2026. https://harvey.ai/blog/ai-contract-review-guide ; and Harvey, security page, undated. https://www.harvey.ai/security
- Juro, AI page, undated, read 22 November 2026. https://www.juro.com/ai
- LEGALFLY, “Harvey AI vs LEGALFLY: which legal AI is right for your team?”, 11 March 2026 (the vendor’s own comparison; it says it was last updated in the first quarter of 2026). https://www.legalfly.com/post/harvey-ai-vs-legalfly
- n8n workflow templates 18196 (“Review contract risks and route approvals with Google Drive, OpenAI, and Gmail”, last updated about two months before reading), 18392 (“Review and approve NDAs with Google Drive, Google Sheets, and OpenAI GPT-4o”, last updated about two months before reading) and 13671 (“Review PDF contract risks with a form, Google Gemini, Sheets, and Slack”, last updated about nine months before reading). https://n8n.io/workflows/18196-review-contract-risks-and-route-approvals-with-google-drive-openai-and-gmail/ ; https://n8n.io/workflows/18392 ; https://n8n.io/workflows/13671
- Zapier, “Generate NDA review summary and notify legal team”. https://zapier.com/automations/legal/contract-management/contract-review/generate-nda-review-summary-and-notify-legal-team ; Microsoft adoption scenario library, “Automated contract review agent” (marked as an example for demonstration). https://adoption.microsoft.com/en-us/scenario-library/legal/automated-contract-review-agent ; ServiceNow Community, article on the Triage Legal Requests agentic workflow (released May 2025). https://www.servicenow.com/community/legal-service-delivery-articles/ai-agent-triage-legal-requests-saves-legal-time-by-routing/ta-p/3352416
- Damien Charlotin, “AI Hallucination Cases Database”, marked “Last updated: 20 November 2026”. https://www.damiencharlotin.com/hallucinations/
- Template libraries and prebuilt agents of agent and workflow platforms, searched 22 November 2026 (Katonic research); and 103 interface descriptions from 15 publishers fetched the same day and run through the platform’s importer (Katonic research).
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y.), Opinion and Order on Sanctions, 22 June 2023. https://www.nhd.uscourts.gov/sites/default/files/pdf/Mata-v-Avianca-sanctions-order.PDF ; and Johnson v. Dunn, 792 F.Supp.3d 1241 (N.D. Ala.), signed 23 July 2025 (a reprint hosted by a university). https://www.law.berkeley.edu/wp-content/uploads/archive/2025/12/Johnson-v-Dunn.pdf
- R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), Divisional Court, 6 June 2025, paragraphs 6 and 9. https://www.judiciary.uk/wp-content/uploads/2025/06/Ayinde-v-London-Borough-of-Haringey-and-Al-Haroun-v-Qatar-National-Bank.pdf
- Stanford HAI, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries”, 23 May 2024 (a summary of Magesh and others, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools”). https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
- Proskauer, client alert on United States v. Heppner (S.D.N.Y.), read 22 November 2026 (ruling 10 February 2026). https://www.proskauer.com/alert/sdny-addresses-privilege-and-work-product-implications-of-using-unsecured-public-AI-tools ; and Venable, note on the written memorandum of 17 February 2026, No. 25 Cr. 503. https://www.venable.com/insights/publications/2026/02/ai-privilege-and-the-heppner-ruling-what-the-court
- American Bar Association, Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512, “Generative Artificial Intelligence Tools”, 29 July 2024. https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
- Federal Rules of Civil Procedure, Rule 11(b) (Cornell Legal Information Institute copy). https://www.law.cornell.edu/rules/frcp/rule_11 ; and United States District Court, Eastern District of Texas, standing order on disclosure and certification of generative AI use, signed 9 April 2025. https://coop.txed.uscourts.gov/sites/default/files/judgeFiles/JDL%20Standing%20Order%20on%20AI%204.9.25.pdf
- Courts and Tribunals Judiciary (England and Wales), “Artificial Intelligence (AI) Guidance for Judicial Office Holders”, 31 October 2025. https://www.judiciary.uk/wp-content/uploads/2025/10/Artificial-Intelligence-AI-Guidance-for-Judicial-Office-Holders-2.pdf
- Solicitors Regulation Authority, warning notice “Misuse of AI”, published 17 August 2026. https://www.sra.org.uk/solicitors/guidance/misuse-ai/
- Court of Justice of the European Union, Akzo Nobel Chemicals and Akcros Chemicals v Commission, Case C-550/07 P, judgment of 14 September 2010, paragraph 44. http://publications.europa.eu/resource/celex/62007CJ0550
- Regulation (EU) 2016/679 (GDPR), Article 22(1). http://publications.europa.eu/resource/celex/32016R0679
- Regulation (EU) 2024/1689 (AI Act), Annex III point 8(a). http://publications.europa.eu/resource/celex/32024R1689
- Regulation (EU) 2026/1744 of 8 July 2026 (Official Journal 24 July 2026), the replacement of Article 113(c) of Regulation (EU) 2024/1689. http://publications.europa.eu/resource/celex/32026R1744
- Supreme Court of New South Wales, Practice Note SC Gen 23, issued 28 January 2025, paragraph 17. https://supremecourt.nsw.gov.au/documents/Practice-and-Procedure/Practice-Notes/general/current/PN_SC_Gen_23.pdf ; Supreme Court of Singapore, Registrar’s Circular No. 1 of 2024, dated 23 September 2024, in effect from 1 October 2024. https://www.judiciary.gov.sg/docs/default-source/circulars/2024/registrar's_circular_no_1_2024_supreme_court.pdf ; DIFC Courts, Practical Guidance Note No. 2 of 2023, 21 December 2023. https://www.difccourts.ae/rules-decisions/practice-directions/practical-guidance-note-no-2-2023-guidelines-use-large-language-models-and-generative-ai-proceedings-difc-courts