AI Governance for Law Firms: Why Your Legal-Research AI Is Probably Minimal-Risk

Probably minimal. That is the short answer for most AI a private law firm uses — and it surprises lawyers, who assume that because the EU AI Act lists "administration of justice" as high-risk, anything legal AI does must be too. It does not work that way. The high-risk heading is aimed at a judicial authority, not at a private firm's research tool. A law firm's real EU AI Act obligations are lighter and different from what the headline suggests — but they are not zero, and the ones that bite are the ones firms tend to overlook.

"AI governance" is a management practice, not a clause in Regulation (EU) 2024/1689. For a law firm, the governance job is mostly to classify the firm's AI correctly (and resist over-classifying it), then to manage the duties that actually apply: confidentiality and privilege, accuracy and hallucination oversight, the Article 50 transparency rule, and the Article 4 AI-literacy duty. This guide walks that line.

---

The administration-of-justice trap

Annex III, point 8 makes high-risk those AI systems "intended to be used by a judicial authority or on its behalf" to research and interpret facts and law and to apply the law to a concrete set of facts (and AI used in alternative dispute resolution where outcomes have legal effect).

Read the words: by a judicial authority, or on its behalf. The high-risk classification attaches to AI that a court uses — or that someone uses for a court — to help decide cases. A private law firm doing its own legal research, e-discovery or contract review for its clients is not a judicial authority and is not acting on behalf of one. Its AI tools therefore generally fall outside Annex III point 8 — they are minimal-risk under the Act.

The edge case to watch: AI a firm builds or operates for a court — for example, a system supplied to a tribunal to support adjudication, or an ADR platform whose AI shapes outcomes with legal effect — can fall inside point 8. That is the exception, not the rule. The everyday tools below are the rule.

The practical message for a governance model: do not let "administration of justice = high-risk" frighten the firm into a conformity assessment it does not owe. The exposure is real elsewhere.

---

Where the real obligations are

1. Confidentiality and legal professional privilege

This is the lift the Act does not even directly impose — but it is the one that should dominate a law firm's AI governance. Feeding privileged client material, matter documents or personal data into a third-party AI tool raises questions of confidentiality, privilege waiver, and GDPR (lawful basis, processor terms, international transfers). A generative tool that retains or trains on prompts could expose privileged content. The governance model's first control is a rule about what may be put into which tool — approved tools, data-residency and no-training contractual terms, and a hard line around privileged and special-category (GDPR Article 9) data. This is professional-conduct and data-protection risk, layered over the AI Act, and it is where a firm's reputation actually lives.

2. Accuracy and hallucination oversight

Generative legal AI fabricates citations. Courts in several jurisdictions have sanctioned lawyers for filing AI-invented authorities. The Act does not make a research tool high-risk for this, but professional duties of competence and candour do. The governance model needs a verification rule: AI output is a draft to be checked against primary sources by a qualified lawyer, never filed unverified. This is human oversight as professional discipline, not a statutory Article 14 obligation (which applies to high-risk systems) — but the discipline is the same.

3. Article 50 transparency for client-facing tools

If the firm runs a client-facing chatbot — intake triage, an "ask a question" assistant on its website — Article 50(1) requires that the user be told they are interacting with an AI system unless it is obvious. Building that in is the provider's design duty — the chatbot vendor's, or the firm's own if it built or white-labelled the tool — so the firm's job is to confirm the notice is there. It is a disclosure obligation (a clear notice), not a conformity assessment, and it has applied since 2 August 2026. A breach sits in the €15M/3% penalty tier (Article 99(4)). For most firms this is a one-line UI change, easily missed.

4. Article 4 AI literacy

Article 4 requires providers and deployers to take measures to support the AI literacy of staff who operate or use AI on their behalf — the wording since the Digital Omnibus took effect on 27 July 2026, which no longer demands a guaranteed individual level. The duty itself has been in force since 2 February 2025 and applies to all AI, not just high-risk — so it covers a firm's use of generative research and drafting tools today. For a law firm this means a baseline training programme: what the approved tools can and cannot do, the confidentiality rules, the verification rule, and the limits of generative output. It is a present obligation, not future work.

---

Provider or deployer? Almost always deployer

A law firm using commercial legal-AI tools (the established legal-research platforms, contract-review tools, e-discovery suites) is a deployer — though for a minimal-risk tool the deployer obligations of Article 26 (which are written for high-risk systems) largely do not apply. The firm's duties come instead from Article 4 literacy, Article 50 disclosure where relevant, and professional/GDPR rules.

A firm crosses into provider territory (Article 16) only if it builds and places an AI system on the market — for instance, productising an internal contract-analysis tool and selling it. And under Article 25, putting the firm's own brand on a third-party tool or substantially modifying it can also flip the role. For the overwhelming majority of firms, none of this applies: they are deployers of minimal-risk tools, and their governance is about confidentiality, verification, literacy and disclosure.

---

Worked example: a mid-size commercial firm

Lindqvist & Sørensen Advokatfirma — ~140 fee-earners and staff, ~€32M turnover, a commercial firm in Denmark and Sweden — uses three AI tools: a commercial legal-research platform with generative answers, an e-discovery/TAR suite for litigation, and a website intake chatbot for new enquiries.

Its governance model, lightweight and centralised:

• Accountable owner: the firm's general counsel / risk partner, who already owns professional-conduct and confidentiality risk.
• Classification: the research platform and the e-discovery suite are minimal-risk — the firm uses them for its own client work, not as a judicial authority or on a court's behalf, so Annex III point 8 does not apply. The intake chatbot is Article 50 limited.
• The duties that actually apply: (1) a confidentiality/privilege policy on what may be entered into each tool, with no-training contractual terms and a privileged-data exclusion; (2) a verification rule — no AI-generated citation or analysis goes to a client or a court unchecked; (3) Article 50 disclosure on the chatbot, confirmed with the vendor and live since 2 August 2026; (4) Article 4 AI-literacy training for fee-earners, run now.
• What it does not owe: no Article 11/Annex IV documentation, no Article 43 conformity assessment, no Article 49 registration, no Article 27 FRIA — because none of its tools is high-risk.

The lesson: Lindqvist & Sørensen's AI governance is real and worth doing, but it is a confidentiality, accuracy and literacy programme — not a high-risk compliance project. Over-classifying would have wasted the firm's time and obscured the privilege risk that genuinely matters.

---

The deadlines that apply to a law firm

For a firm whose AI is minimal-risk, the high-risk dates are largely irrelevant — but it is worth stating them so a firm can confirm it is out of scope. Stand-alone high-risk Annex III systems apply from 2 December 2027 under the Digital Omnibus — Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026 — deferred from the original 2 August 2026. The dates that actually bind a typical firm are earlier and already running: Article 4 literacy and Article 5 prohibitions (in force since 2 February 2025) and Article 50 transparency (since 2 August 2026).

---

Penalties

Under Article 99: prohibited Article 5 practices — €35,000,000 or 7% (99(3)); high-risk obligations and Article 50 disclosure failures — €15,000,000 or 3% (99(4)); incorrect/misleading information to an authority — €7,500,000 or 1% (99(5)). The third tier is 1%, never 1.5%. Most firms meet the SME definition, so the Article 99(6) cap (the lower of the percentage or the fixed amount) applies — a genuine proportionality benefit, not a get-out. For a minimal-risk firm the realistic exposure is the Article 50 tier on a client chatbot, not the high-risk stack.

---

How Confir helps

Confir steps a firm's intake through the Annex III points in plain-English checklists, which is exactly what stops the "administration-of-justice" over-classification: it confirms that legal-research, e-discovery and contract-review tools used for the firm's own work are minimal-risk, flags a client chatbot as Article 50, and surfaces the Article 4 literacy duty. The logic is deterministic and rule-based — same intake, same finding, with the rule that fired recorded — so a risk partner can show why a tool was classified as it was. For the rare firm that builds a tool for a court, it flags the Annex III point 8 route.

Related guides

• AI governance operating model
• Legal-sector AI compliance
• Administration-of-justice AI (Annex III pt 8)
• Article 4 AI-literacy requirements
• Article 50 chatbot transparency
• Minimal-risk AI explained
• GDPR and the EU AI Act intersection