Buying Legal AI Under the EU AI Act: What to Check Before You Sign
Most legal AI tools aren't high-risk under the EU AI Act, but your firm still has duties as a deployer. Here's what to check, what to put in the contract, and what compliance typically costs.
Your Role Under the Act Decides Your Obligations
Before you sign a legal AI contract, work out which role the Act gives you. For most EU AI Act law firms questions, the answer is that the firm is a deployer: an organization using an AI system under its own authority in a professional setting. The vendor that builds and markets the tool is the provider. The heavy obligations, such as conformity assessment, technical documentation, CE marking and post-market monitoring, sit mostly with providers. Deployers carry a shorter list, but it's real and enforceable.
The Act also reaches beyond EU borders. It applies to providers and deployers established outside the EU when the system's output is used in the EU. A firm in New York, Dubai, Singapore or São Paulo that runs AI-assisted contract review for a client in Paris may be in scope for that work. Whether a given matter triggers the Act turns on facts that are still being tested, so take the question to qualified EU counsel rather than assuming you're exempt.
Most legal tools aren't high-risk, but some uses are
Drafting assistants, research tools, e-discovery review and contract analytics used by a law firm for its own work generally fall outside the high-risk categories in Annex III. The legal-sector entry targets AI used by or on behalf of judicial authorities to research and interpret facts and law and to apply the law, plus similar use in alternative dispute resolution. A firm building tools for a court or an arbitral institution should read that entry closely.
Firms are more likely to hit the high-risk tier through the back office. AI that screens trainee or associate applications, ranks candidates, or evaluates and monitors employees is high-risk under the employment category. When you inventory your AI, list HR and recruitment software next to the practice tools.
One more trap. If you put your firm's name on a vendor's system, substantially modify it, or repurpose it for a high-risk use, you can become a provider yourself. That matters for firms launching client-facing tools on top of third-party models.
Deployer Duties and the AI Literacy Obligation in Practice
The AI literacy obligation in Article 4 has applied since February 2025. Providers and deployers must take measures to ensure that staff and others operating AI systems on their behalf have a sufficient level of AI literacy, taking account of technical knowledge, experience, training and the context of use. The Act prescribes no curriculum or certificate. A one-off webinar is unlikely to satisfy it, and an 80-hour course for every secretary is overkill. Aim for training proportionate to role and risk, plus a record that it happened.
A workable structure for a mid-sized firm:
- All staff: what the approved tools do, what they shouldn't be used for, confidentiality rules, and how to spot hallucinated citations.
- Fee earners using AI on client work: verification standards, prompt hygiene with client data, documenting review, and when to tell a client.
- Partners, IT and risk: vendor assessment, incident handling, model-change risk and the firm's AI policy.
In late 2025 the European Commission proposed a digital omnibus package that, among other things, would soften the Article 4 duty and push back some high-risk deadlines. Confirm whether it has been adopted, and in what form, before you build a plan around either version. Even if the legal duty is diluted, professional conduct rules in many jurisdictions already expect lawyers to understand the technology they use. Keep the training.
If you deploy a high-risk system
Under Article 26, deployers of high-risk systems must use them according to the provider's instructions, assign human oversight to people with the competence and authority to exercise it, ensure that input data they control is relevant, monitor operation and report serious incidents, keep automatically generated logs for at least six months, and inform workers and their representatives before using such a system in the workplace. Some public bodies and certain private-sector deployers must also carry out a fundamental rights impact assessment. Most law firms won't, but in-house teams at banks and insurers should check.
Transparency duties
Article 50 matters for client-facing chatbots and intake tools: people must be told they're interacting with an AI system unless that's obvious. Published AI-generated text on matters of public interest carries disclosure duties too, with an exception where a human has reviewed it and taken editorial responsibility. That exception should cover most client alerts a lawyer actually edits, but confirm how your regulator reads it.
Where the GDPR Overlap Bites Harder Than the AI Act
For most firms, the GDPR creates more day-to-day work on legal AI tools than the AI Act does. Client documents are full of personal data, often including special category data in employment, health, immigration and criminal matters. Every review of a legal AI tool should cover these points:
- Controller and processor roles. The vendor is usually a processor and needs an Article 28 data processing agreement. Watch for vendors that reserve the right to use your data for their own purposes. That can make them a controller for that processing.
- DPIA. Large-scale processing of sensitive client data with new technology will often trigger a data protection impact assessment under Article 35. Some national supervisory authorities list AI-related processing among the activities that require one.
- International transfers. If prompts or documents go to model infrastructure outside the EEA, you need a valid transfer mechanism: an adequacy decision (including the EU-US Data Privacy Framework for certified US recipients), standard contractual clauses with a transfer impact assessment, or EU-only hosting.
- Automated decisions. Article 22 restricts solely automated decisions with legal or similarly significant effects, which is relevant to AI recruitment screening.
Professional secrecy rules sit on top of the GDPR. In several countries, including Germany, France and Belgium, lawyers' confidentiality duties limit which service providers can access client material and on what terms. Get your local bar's guidance on cloud and AI use.
Outside the EU, the same diligence maps onto other regimes. The UK isn't covered by the AI Act but applies UK GDPR. Switzerland has its revised Federal Act on Data Protection. India's Digital Personal Data Protection Act, Brazil's LGPD, and the personal data laws of Singapore and Thailand all touch on processor and transfer questions. A non-EU firm serving EU clients will often have to satisfy both its home regime and the GDPR.
Contract Terms to Demand in Legal AI Vendor Due Diligence
Standard SaaS paper rarely covers what the AI Act and the GDPR now expect. Use this table during legal AI vendor due diligence to decide what to push for and when to walk away.
| Term | What to ask for | Red flag |
|---|---|---|
| AI Act role and classification | Written statement of the vendor's role, the system's risk classification and its intended purpose | "We're not subject to the AI Act" with no reasoning |
| Instructions for use | Documented capabilities, limitations and known failure modes | Marketing material offered in place of documentation |
| Training on your data | Contractual bar on using prompts, outputs or documents to train or improve models | Opt-out buried in settings, or exceptions for "service improvement" |
| Underlying model | Name of the foundation model provider and its general-purpose AI compliance status, including any Code of Practice commitments | Refusal to disclose which model sits underneath |
| Model changes | Advance notice of material model or behavior changes; right to test before rollout | Silent model swaps |
| Data location and sub-processors | Hosting region options, full sub-processor list, notice of changes with a right to object | Sub-processors "available on request" only |
| Logs and audit | Access to usage logs for at least six months; audit or third-party assurance rights (ISO/IEC 27001, SOC 2, ISO/IEC 42001) | No log export |
| Incidents and regulators | Defined incident notification windows; cooperation with supervisory authorities | Notification only "without undue delay" with no outer limit |
| Liability and IP | Higher or uncapped liability for data breach and confidentiality; IP indemnity for outputs | Liability capped at one month's fees |
| Exit | Data return and certified deletion, including from caches and backups, within a fixed period | Deletion "in accordance with our retention policy" |
Large vendors such as Thomson Reuters, LexisNexis and Microsoft typically offer mature data processing terms and, in many cases, EU hosting options, though availability varies by product and plan. Their standard paper often omits model-change notice or log access, so ask. Smaller startups may move faster on bespoke terms but carry more continuity risk. For a firm under 50 lawyers, prioritize the training bar, data location and exit terms. Those are the clauses that hurt most if you get them wrong.
Typical Compliance Costs and a Phased Timeline
The figures below are typical 2026 market ranges in USD. They vary widely by country, firm size and contract, so confirm with vendors and get local quotes.
| Cost item | Typical range | What drives it |
|---|---|---|
| AI literacy e-learning | $40-$250 per user per year | Role-based modules, languages, reporting |
| Live workshops for fee earners | $3,000-$15,000 per session | Customization, presenter seniority |
| External gap assessment and AI policy | $10,000-$60,000 | Number of tools, jurisdictions, high-risk uses |
| DPIA per significant tool | $3,000-$20,000 if outsourced | Data sensitivity, transfer analysis |
| AI governance or inventory software | $5,000-$50,000 per year | User count, integrations |
| Contract renegotiation | $2,000-$15,000 per vendor in outside counsel time | Vendor leverage, number of redlines |
Hypothetical worked example: say a 60-lawyer firm in Madrid with 140 staff uses three AI tools plus AI-assisted recruitment software. Role-based e-learning at about $100 per head comes to $14,000. Add two fee-earner workshops at $6,000 each, a $25,000 external gap assessment, DPIAs on two tools at $8,000 each, and redlines with four vendors at $5,000 each. First-year spend lands near $87,000, with ongoing costs falling to perhaps $25,000-$40,000 a year. The recruitment tool is the high-risk item, and it deserves the most oversight design.
Phased timeline
Check these dates against the current text and any amendments published in the Official Journal of the EU. The digital omnibus proposal may move some of them.
| Date | What applies | Action for firms |
|---|---|---|
| 1 August 2024 | Act enters into force | Start AI inventory |
| 2 February 2025 | Prohibited practices; AI literacy duty | Training in place and documented |
| 2 August 2025 | General-purpose AI model obligations; penalties regime | Ask vendors about underlying model compliance |
| 2 August 2026 (as originally enacted) | Most remaining rules, including Annex III high-risk and transparency duties | Oversight, logging and worker notice for high-risk uses; check whether a delay was adopted |
| 2 August 2027 (as originally enacted) | High-risk AI embedded in regulated products | Mainly relevant to product-sector clients |
Fines for breaching most operator obligations can reach €15 million or 3% of worldwide annual turnover, whichever is higher. The ceiling for prohibited practices is higher. National authorities enforce, and their readiness varies by member state.
A decision rule: don't sign or renew a legal AI contract until you've seen the vendor's written AI Act classification, a DPA that bars training on your data, and a named hosting region. If a vendor can't supply all three within two weeks, pause the deal and take the file to qualified EU regulatory and data protection counsel.
Quick answers
Does the EU AI Act apply to a law firm based outside the EU?
It can. The Act applies to providers and deployers outside the EU when the AI system's output is used in the EU, so AI-assisted work for EU clients may bring a non-EU firm into scope. How far that reaches in a particular case is still unsettled, so get advice from qualified EU counsel.
Are legal research and contract review tools high-risk under the AI Act?
Generally no, when a law firm or in-house team uses them for its own work. The legal-sector high-risk category covers AI used by or on behalf of judicial authorities and in alternative dispute resolution. Firms are more likely to reach the high-risk tier through AI recruitment or employee-monitoring software.
What does the AI literacy obligation actually require us to do?
It requires measures that give staff working with AI a sufficient level of understanding for their role and context. The Act sets no fixed syllabus or hours. Role-based training with records of who completed what is the sensible approach. Check whether the proposed digital omnibus changes to this duty have been adopted.
Is a GDPR data processing agreement enough for a legal AI vendor?
It covers the data protection basics, but not the AI-specific risks. You'll also want a contractual bar on training with your data, notice of model changes, log access, clarity on the underlying model provider, and firm exit and deletion terms. Professional secrecy rules in your jurisdiction may add further requirements.
How much should a mid-sized firm budget for AI Act compliance?
As a typical 2026 market range, a firm of 50-100 lawyers might spend roughly $50,000-$120,000 in the first year on training, gap assessment, DPIAs and vendor renegotiation. Ongoing costs are usually lower. Actual figures vary by country and by how many tools and high-risk uses you have, so get local quotes.
Written by the Counsel Core Tech editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.