Compliance and Regulatory Considerations for AI in Legal

Updated August 2026 · 6 min read · Deal Room Intelligence Series

The timeline, all of it verifiable

Standards bodies and bar associations move slowly and do not produce guidance for practices that are not happening. Four significant instruments in under three years is itself evidence about the state of adoption — better evidence, in fact, than any survey, since every element is checkable and none of it is sponsored.

ABA Formal Opinion 512 — the one that binds first

Issued 29 July 2024 by the ABA Standing Committee on Ethics and Professional Responsibility. It holds that lawyers and firms using generative AI must “fully consider their applicable ethical obligations,” naming duties of competence, confidentiality, communication with clients, candor toward tribunals, supervisory responsibility, and charging reasonable fees consistent with time actually spent.^[2]

Four practical readings:

• Competence attaches to output, not procurement. Choosing a reputable vendor does not discharge it. The duty follows the work product that leaves your desk.
• Supervision applies to work you did not personally produce — the same principle as delegating to a junior, except the junior can explain their reasoning and the model cannot.
• Confidentiality governs what you put in. Feeding client material to a system whose terms permit training or undisclosed onward transfer is a conduct problem before it is a security problem.
• Fees must reflect time actually spent. If a tool compresses twenty hours of triage to two, billing twenty is not defensible. This is genuinely uncomfortable for hourly models and deserves a decision rather than avoidance.

EU AI Act — scope first, penalties second

Applying from 2 August 2025, with administrative fines under Article 99 of €35,000,000 or 7% of total worldwide annual turnover for prohibited practices, €15,000,000 or 3% for breaches of provider and deployer obligations, and €7,500,000 or 1% for supplying false or misleading information to authorities. For SMEs the fine is the lower of the percentage or the fixed amount.^[3]

The penalties get quoted; the scoping question matters more and is genuinely fact-specific:

• Are you a provider or a deployer? Different obligations attach. A firm using a purchased tool is generally a deployer, but building or substantially modifying a system changes that.
• What risk classification applies? Document screening for M&A is not obviously in the high-risk categories, but classification depends on use rather than on the technology, and adjacent uses — employment-related screening, for instance — can land differently.
• Does it reach you at all? Territorial scope is broader than establishment in the EU.

NIST AI RMF — the practical one

Published 26 January 2023 and voluntary, but the most immediately useful of the four because it gives you a structure rather than an obligation. Its four core functions — Govern, Map, Measure, Manage — map cleanly onto what an internal programme needs:^[1]

• Govern — who owns AI risk, who approves tools, what the escalation path is
• Map — where AI is actually used, on what data, in which decisions
• Measure — error rates, override rates, coverage gaps, sampled grounding checks
• Manage — thresholds, verification protocol, retention, response when a measurement goes wrong

The Measure function is where most firms are weakest, and it is the one that makes the rest credible. Without measurement you have a policy, not a programme.

What the rules converge on

Read the four instruments together and they demand the same thing in different registers: the ability to show how a conclusion was reached. Competence and supervision under 512. Documentation and record-keeping under the AI Act. Measure and Manage under the RMF. Documented controls under ISO 42001.

That has a concrete implication for tooling, and it is the reason this is not purely a policy exercise:

A programme that satisfies all four

1. Written policy — approved tools, permitted uses, prohibited uses, named owner.
2. Vendor terms — contractual training prohibition with flow-down, named subprocessors, stated retention, model versioning, run-record export.
3. Verification protocol — every escalated finding traced to source text before it reaches a committee.
4. Pre-committed thresholds, recorded as they stood at run time. A threshold you can describe afterwards but not evidence reads as a rationalisation.
5. Measurement — coverage reconciliation every deal, a fixed sample of low-severity findings, quarterly calibration against overrides.
6. Records — inventory, scores, thresholds, quotes, versions, overrides, read-access log. Retain the metadata after deleting the documents.
7. Client communication — resolve when disclosure is required, with your risk function, before a client asks.
8. Insurance — check your PI policy for an AI exclusion or notification condition now, not after a claim.

The compliance argument that is not really about compliance

Grounded commercial legal AI has been measured hallucinating between 17% and 33% of the time in an adjacent task, with providers' hallucination-free claims judged overstated.^[4] Errors are a permanent design assumption, not a phase.

Which means the record is not paperwork produced for a regulator. It is the thing that determines, when something is missed, whether you are describing a documented judgment or an unmanaged gap. The compliance framework and the operational protection are the same artefact, and firms that build it for the second reason tend to build it better.

Bottom line

Four instruments in under three years, all dated and checkable. ABA 512 binds US lawyers now; the EU AI Act binds in scope with serious penalties; NIST gives you the structure and ISO 42001 the certification path.

They converge on showing your work — which is a tooling decision as much as a policy one, because narrative output cannot evidence what these rules ask for and scored, quote-backed output can.

See output built to be evidenced →