An associate is finishing a first draft at 11pm. The client’s term sheet goes into a personal chatbot, a clean draft comes back in a minute, and it reads well. 2 weeks later, opposing counsel points out a clause reference that doesn’t exist, and nobody at the firm knows which tool wrote it or what else went into it.
That scene combines the two things a firm needs to know about AI. It saves real time, and it creates real exposure when each person picks their own tool. So the firm decides once how AI may be used, and every tool follows those rules.
This primer is for partners, practice managers and compliance leads at EU firms, with a section on Malta. It covers the terms, the rules, the risks on record, a governed setup and a 90-day plan.
In plain terms
These are the words that come up in every AI conversation at a firm.
| Term | What it means for a firm |
|---|---|
| Provider | The company that builds an AI system or model and offers it under its own name (AI Act, Article 3). Your AI vendor. |
| Deployer | Anyone who uses an AI system under its authority for professional work. Your firm, as soon as staff use AI for client or firm work. |
| High-risk system | An AI system in a use the Act lists in Annex III, such as hiring, promotion or performance evaluation. It carries the heaviest duties. |
| AI literacy | Enough skill and knowledge for staff to use AI with its risks in mind. Required by Article 4. |
| Hallucination | An answer that reads well and is wrong, such as a case that doesn’t exist or a quote a judge never wrote. |
| Privilege | The protection that keeps lawyer and client communications confidential. In Malta it rests on professional secrecy under Article 257 of the Criminal Code. |
| Automated decision | A decision about a person made solely by a system, with legal or similarly significant effects. GDPR Article 22 limits these and gives people rights over them. |
Which rules apply to your firm
Go down the left column. Every line that describes your firm brings the rule beside it.
| If this is you | What applies | From |
|---|---|---|
| Anyone at the firm uses AI for work | You are a deployer. AI literacy under Article 4 | 2 February 2025, supervised from 2 August 2026 |
| AI touches personal data, such as client names, emails or case files | GDPR: a legal basis, a processing contract with the vendor and an impact assessment where risk is high | Now |
| Client documents go into an AI tool | Professional confidentiality and secrecy rules, and your bar’s AI guidance | Now |
| A chatbot on your website talks to the public | People must be told they’re talking to AI (Article 50). If you build it yourself, you may be its provider | 2 August 2026 |
| AI scores, ranks, allocates work to or monitors staff | High-risk under Annex III point 4, with deployer duties under Article 26. GDPR Article 22 today | 2 December 2027 |
| You put your firm’s name on an AI tool you offer clients | You may be its provider, with the provider’s duties (Article 25) | Depends on the system |
| Any tool reads emotions at work | Banned under Article 5(1)(f), except for medical or safety reasons | 2 February 2025 |
| You practise in Malta | The MDIA and the IDPC supervise, and Maltese fines apply | 2 August 2026 |
AI research and drafting tools a firm uses for its own work aren’t on the high-risk list. The justice entry in Annex III covers AI used by or for courts, and in a similar way in alternative dispute resolution.
Where AI already pays in legal work
The strongest evidence comes from a law firm’s own trials. Ashurst ran generative AI with 411 partners, lawyers and staff across 23 offices and measured the time to produce first drafts (Legal Futures, June 2024).
| Task | Time saved on the first draft | What a lawyer still does |
|---|---|---|
| Corporate filings | 80% | Confirms facts and figures against the record |
| Industry research reports | 59% | Verifies sources and adds judgement |
| Legal briefings | 45% | Checks every authority and finishes the argument |
Use is spreading fast:
- 61% of UK lawyers now use AI in their daily work, up from 46% in January 2025. 56% of those in private practice put the time saved into more billable work (LexisNexis, September 2025).
- Document review, research and summaries are the most common uses, at 77%, 74% and 74% of legal AI users (Thomson Reuters survey via LawNext, April 2025).
- 3.0 billable hours is what an average lawyer using Clio records in an 8-hour day, a 38% utilisation rate (Clio 2025 benchmarks). Much of the rest is drafting, searching and admin that AI can shorten.
Risks courts and regulators have recorded
2 kinds of failure keep reaching courts and regulators: made-up material, and client data going where it shouldn’t.
| What happened | Source | What it means for a firm |
|---|---|---|
| 2,190 court decisions involve AI-made-up material, 870 of them involving lawyers | AI Hallucination Cases database, 15 October 2026 | Check every authority and quote before anything is filed |
| Leading legal research tools gave wrong or made-up answers to at least 1 query in 6 | Stanford HAI, May 2024 | Specialist tools need the same checking as general ones |
| The SRA’s warning notice says solicitors stay accountable for AI output, and that client data in tools without safeguards is likely to breach confidentiality and may waive privilege | SRA warning notice, 17 August 2026 | Only approved tools with contractual and technical safeguards get client data |
| The SRA received 42 reports of possible AI misuse between July 2025 and July 2026 | SRA news release | Regulators are already looking for this |
| A US federal court held a defendant’s chats with a public AI tool were neither privileged nor work product, because the provider’s terms allowed retention, training use and disclosure | US v. Heppner, S.D.N.Y., February 2026 | Consumer terms can defeat confidentiality, whatever the user intended |
The SRA regulates solicitors in England and Wales, and Heppner is a US case. But the duties behind them are the same ones EU bars apply, as the CCBE guide below shows.
Three ways staff end up using AI
The risk mostly depends on the route a request takes. These are the three routes staff use.
| Consumer chatbot | Firm-approved tool | Governed route | |
|---|---|---|---|
| Data handling | Personal account. Terms may allow retention and training | Business contract bars training. Data stays in the vendor’s tenant | Business contracts, and client identifiers are removed before a request leaves the firm |
| Privilege risk | High, as Heppner shows | Lower, and depends on contract and settings | Lower, with less client data in each request |
| Logging | None the firm can see | Per tool, in each vendor’s admin console | Every request, across every tool, in one log |
| Sources | Often none | Depends on the tool | Answers carry their sources for checking |
| Cost control | Personal card or free tier, invisible | Seat licences per tool | Budgets per team and per person |
What the EU AI Act asks of a law firm
The AI Act applies to firms that use AI as well as to the companies that build it. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published on 24 July 2026 and moved the high-risk dates (European Parliament, EUR-Lex).
| Date | What applies | What a firm does |
|---|---|---|
| 2 February 2025 | AI literacy (Article 4). Banned practices, including emotion recognition at work | Train staff for the tools and risks they face, and keep a record |
| 2 August 2025 | Duties on general-purpose model providers. The Act’s penalty rules | Ask vendors for their model documentation |
| 2 August 2026 | National authorities supervise and enforce. Transparency duties apply | Have the training record ready. Label public chatbots |
| 2 December 2027 | High-risk rules for stand-alone systems, including AI used in employment | Meet deployer duties for any high-risk system you use |
| 2 August 2028 | High-risk rules for AI built into regulated products | Rarely relevant to a firm’s own use |
AI literacy is already supervised. Tools that score or monitor employees have until 2 December 2027 to meet the high-risk rules.
AI literacy
The AI literacy duty is lighter than it sounds. The Omnibus kept it as a duty on providers and deployers, and added that it doesn’t require a specific literacy level of anyone (Cuatrecasas summary of the Omnibus).
What the Commission’s guidance says, updated 27 July 2026 (European Commission AI literacy Q&A):
- No certificate is needed. An internal record of training and other guidance is the expected evidence.
- Training matches the tool and the role. Its own example is telling staff who write with a chatbot about hallucination.
- High-risk deployers still train for oversight. The people who supervise a high-risk system need the competence and authority to do it.
Penalties
Breaching most operator duties under the Act can cost up to €15 million or 3% of worldwide turnover, whichever is higher. For smaller firms the lower of the 2 figures applies (Article 99).
What GDPR adds
GDPR applied to AI long before the AI Act. Any AI tool that reads a client file processes personal data, so GDPR is usually the first rule a firm’s AI use meets.
- A legal basis for each use. Drafting from a client’s file and training a vendor’s model on it are two different purposes.
- A processing contract with the vendor. It sets where data is stored, who can see it and that it isn’t used for training.
- An impact assessment where risk is high. Article 35 names systematic, automated evaluation of people as one case that needs it.
- Limits on automated decisions. The Court of Justice held in SCHUFA that a score a third party draws strongly on can itself be an automated decision under Article 22.
- An explanation people can act on. In Dun & Bradstreet Austria, the Court held that a person is owed an explanation of the procedure and principles applied, so they can understand and contest the decision.
The European Data Protection Board’s Opinion 28/2024 on AI models adds 3 points that matter when a firm picks a vendor:
- A model trained on personal data isn’t automatically anonymous. That’s decided case by case.
- Legitimate interest needs a 3-step test. A real interest, a need with no less intrusive option, and a balance against people’s reasonable expectations.
- A deployer should check how the model was built. If a model was developed unlawfully, a firm deploying it is expected to have assessed that.
AI law and supervision in Malta
Malta set up its AI Act supervision with two legal notices in October 2025. Both were fully in force by 2 August 2026.
| Body | Role under the AI Act | Source |
|---|---|---|
| Malta Digital Innovation Authority (MDIA) | Market surveillance authority by default, single point of contact and notifying authority. Responsible for the national AI regulatory sandbox | L.N. 226 of 2025 |
| Information and Data Protection Commissioner (IDPC) | Market surveillance for high-risk AI in biometrics, emergency calls, law enforcement, migration and the administration of justice. Protects the right to data protection under Article 77 | L.N. 227 of 2025 |
| Malta Financial Services Authority (MFSA) | Coordinates with the MDIA where regulated financial institutions use high-risk AI | L.N. 226 of 2025 |
| National Accreditation Board | Assesses and monitors the bodies that certify high-risk systems | L.N. 226 of 2025 |
What this means for a Maltese firm:
- The MDIA is the main regulator for a firm’s own AI use. Anyone can complain to it, and it can fine up to €350,000 per breach or 1% of worldwide turnover, whichever is higher, plus €12,000 a day while a breach continues. These sit alongside the Act’s own penalties.
- The IDPC is still the GDPR regulator under the Data Protection Act, Cap. 586. Any AI tool that processes client data is also its business.
- Professional secrecy is criminal law. Disclosing a professional secret can bring a fine of up to €46,587.47, up to 2 years in prison, or both (Criminal Code, Article 257). The Professional Secrecy Act names advocates and legal procurators among the professions it covers.
- There’s no Maltese bar guidance on generative AI in the CCBE’s list. The annex of national guidance in the CCBE guide of October 2025 has no entry for Malta. The Chamber of Advocates’ new council said in July 2026 that it will promote the responsible, ethical use of digital tools, including AI (Chamber of Advocates).
Malta also has a national AI strategy. The 2019 strategy, “Malta: The Ultimate AI Launchpad”, set a vision to 2030 with a focus on public-sector adoption (OECD.AI).
In November 2025 the government opened a public consultation on an updated strategy (Business Now, 20 November 2025). The Malta Chamber of Commerce’s response to the MDIA’s draft in February 2026 asked for clear owners and timelines for each of its measures (Malta Chamber feedback). Until the Chamber of Advocates publishes its own guidance, the CCBE guide below is the closest professional reference for Maltese advocates.
What European bars say
The CCBE guide on the use of generative AI by lawyers, published on 2 October 2025, is the reference for bars across the EU. It maps generative AI onto 5 professional duties.
- Confidentiality. Don’t enter client data into a tool unless appropriate safeguards are in place.
- Competence. Verify what the tool produces before relying on it.
- Independence. Watch for bias and for a tool that agrees too readily, so the advice stays impartial.
- Transparency with the client. Be open about AI use where an informed client would object or have reservations.
- Conflicts of interest. A tool with access to many clients’ files can carry one client’s data into a matter against them.
National bars say much the same. Here are 3 examples.
| Bar | Guidance | Point worth noting |
|---|---|---|
| Germany, BRAK | Notes on the use of AI, December 2024 | With tools like ChatGPT, client secrets don’t need to be entered and must not be |
| Ireland, Law Society | Generative AI guidelines, December 2025 | Suggests mentioning AI use in the client terms letter. No general duty to disclose it unless asked or the AI Act requires it |
| United States, ABA | Formal Opinion 512, July 2024 | Adds candour to the court and reasonable fees to the same core duties |
The governed setup
Each professional duty maps to a control the firm can run and show.
| Duty | Control | How it works |
|---|---|---|
| Confidentiality | Client data stays in | Card, bank and ID details are removed before a request leaves the firm, names and contact details can be added, and only providers that don’t train on your data are allowed |
| Competence | Sources with every answer | Drafts and research answers show where each statement came from, so checking is fast |
| Supervision | Every use logged | Who asked what, with which tool and when, for partners and for the regulator |
| Access and conflicts | File-level permissions | The knowledge assistant only answers from documents the person can already open, so one client’s file can’t reach another’s matter |
| Fees and cost | Budgets per team | Spend per person and practice group, with limits before it runs over |
All of it sits behind a single governed route, signed in with the firm’s own single sign-on. We describe the full pattern in our AI governance framework guide.
Before any client material goes into AI, staff can run a short check.
We run this setup in production for a regulated enterprise client. Its privacy guardrail removed card numbers, passwords, keys or ID numbers from 2,180 AI requests before they left the company.
Client intake that never gives advice
Intake is a good second step. An intake assistant takes the routine part of every enquiry, and a lawyer handles everything that needs judgement.
The assistant handles the routine part of every enquiry. Step 4 is the line it never crosses: anything that needs judgement leaves the assistant and reaches a person.
- Answer routine questions at any hour. Office hours, fees for a first meeting, what documents to bring.
- Collect the facts and documents a lawyer needs, in a set order for each matter type.
- Run a conflict pre-check against existing matters and record the result.
- Hand any advice question to a lawyer. The assistant never gives legal advice.
- Book the consultation and attach a file note with the matter type, urgency and conflict result.
If the assistant is on your website, tell people they’re talking to AI. It’s the same pattern we use to check support AI conversations, where every answer is graded against the record and sensitive topics always reach a person.
Employee scoring for employment clients
Employment firms are being asked to recommend, or offer, tools that score employees. They can, if the tool is built for the rules.
- The AI Act: AI used to decide on promotion or termination, allocate tasks from behaviour or personal traits, or monitor and evaluate performance is high-risk from 2 December 2027 (Annex III, point 4).
- GDPR, today: a score that drives the decision can itself be an automated decision, and the person is owed an explanation they can act on.
A tool built for both sets of rules has 10 parts.
- A clear purpose and legal basis. Consent rarely works between employer and employee.
- An impact assessment before launch. Scoring people in a weaker position calls for one.
- Bias testing, before launch and at intervals. The amended Act allows special category data for this where strictly necessary, with safeguards.
- An explanation for every score. Which factors drove it and what would change it.
- A person who can override the score. And every override recorded.
- Logs kept at least 6 months. Inputs, scores and overrides (Article 26).
- Notice to employees and their representatives. Before use, with works council rights where they apply.
- A way to contest a score. With human review and a recorded answer.
- No emotion or biometric inference. Emotion recognition at work has been banned since February 2025.
- Technical documentation. Ready for December 2027.
A 90-day plan
The plan picks one task, proves it on live work and decides on the numbers.
Most of the 13 weeks go to live use and measurement, so the decision at the end rests on the firm's own numbers.
- Weeks 1 and 2: choose. Pick a single task with high volume and easy checking. Measure today’s time and risk.
- Weeks 3 to 6: build. Set it up on the firm’s sign-in and EU hosting, with AI suggesting and lawyers deciding.
- Weeks 7 to 10: pilot. A small group uses it on live matters, and the numbers are reviewed weekly.
- Weeks 11 to 13: decide. Train staff by role, keep the record, and choose the next task on evidence.
What to measure
| KPI | How it is measured | Target |
|---|---|---|
| Time to a first draft | Matter time records before and after | Down by a third or more |
| Citations checked before filing | Review log | 100% |
| Requests with client data removed | Guardrail log | Every request scanned |
| Staff trained for their role | Training record | 100% of AI users |
| AI spend per practice group | Monthly spend report | Inside budget each month |
If your firm wants to put this in place, talk to us. We start with a single task and the numbers that show it working.
Frequently asked questions
Does the EU AI Act apply to a law firm that only uses AI tools?
Yes. A firm that uses AI for its work is a deployer under the Act. The AI literacy duty in Article 4 has applied since 2 February 2025, and national authorities supervise it from 2 August 2026. If the firm deploys a high-risk system, for example one that evaluates its own staff, the deployer duties in Article 26 apply from 2 December 2027.
Who supervises the AI Act in Malta?
The Malta Digital Innovation Authority is the default market surveillance authority, the single point of contact and the notifying authority, under Legal Notice 226 of 2025. The Information and Data Protection Commissioner supervises high-risk AI used in law enforcement, migration and the administration of justice, and protects the right to data protection, under Legal Notice 227 of 2025.
Do our lawyers need an AI certificate?
No. The European Commission's AI literacy guidance says no certificate is needed. It suggests an internal record of training and other guidance, matched to each person's role and the tools they use. Since the 2026 amendments, the Act also says it does not require a specific literacy level of anyone.
Can we use ChatGPT or another public chatbot for client work?
Only with care. The CCBE tells lawyers not to enter client data into these tools unless appropriate safeguards are in place. The SRA warned in August 2026 that doing so without safeguards is likely to breach confidentiality and may waive privilege, and a US court held in February 2026 that chats with a consumer AI tool were not privileged. Use tools that keep data under the firm's control and are contractually barred from training on it.
Is AI legal research a high-risk use under the AI Act?
Not on its own. The justice category in Annex III covers AI used by or on behalf of a judicial authority, or in a similar way in alternative dispute resolution. A firm's own research and drafting tools fall under AI literacy, GDPR and professional rules, and the firm still checks every output.
Is an employee scoring tool legal in the EU?
It can be, if it is built for the rules. Scoring employees for promotion, task allocation or performance is high-risk under Annex III of the AI Act from 2 December 2027, and GDPR rules on automated decisions apply today. The tool needs a legal basis, an impact assessment, bias testing, explanations, human override, logs and a way to contest a score. Emotion recognition at work is banned outright.
Where should a firm start?
With a single task that has high volume and clear checking, such as first drafts of routine letters or summaries of document bundles. Measure today's time, run it with a small group on live matters for a few weeks, and decide on evidence before adding the next task.
How we implement AI governance Talk to us about your project