The Governance Imperative

Lawyers are already using AI: for a first draft of a letter, a summary of a bundle, or a first pass at a research question. The live question is whether that use is governed or invisible. For a law firm, the gap between those two states is not an IT preference. It is a professional-conduct problem, a data-protection problem, and a problem courts are willing to sanction.

This is the opening paper in a series that sets out a practical blueprint for adopting AI inside a professional practice without giving up client confidentiality, privilege, or the trust that underwrites both. It makes the case for why governance has to come first. The papers that follow set out how: data classification, deployment architecture, human verification, and the professional-conduct map. Here, the stakes.

1The tools are already inside the firm

Adoption has moved faster than many boards realise. Thomson Reuters’ 2025 survey of professional-services organisations found generative-AI use across the sector nearly doubled year on year. Within the legal profession, the share using these tools rose from 14% to 26% in a single year, and more than nine in ten expect it to become central to their workflow within five years.¹

Adoption is not the problem. The problem is what has not kept pace. In the same research, roughly half of legal professionals reported that their organisation has no formal policy governing generative AI at all. The capability arrived through software firms already licence, and on every lawyer’s phone, while the governance that should have accompanied it did not.

A firm without a policy has not opted out. It has left every individual to decide what client information may go into which tool. That is ungoverned risk. “We haven’t really adopted AI yet” is, in almost every such firm, untrue. The firm has adopted it. It has simply not decided how.

2Why a law firm cannot treat AI like ordinary software

Most office software can be evaluated on price, features and reliability. AI that touches confidential client work cannot, because it engages the duties that define legal practice. A single tool, used on a single matter, can put six of them in play at once.

This is not only a theoretical mapping. In July 2024 the American Bar Association issued Formal Opinion 512, its first ethics guidance on generative AI. It read the technology against the existing duties of competence, confidentiality, communication, candour to the tribunal, supervision and reasonable fees.² The same duties bind advocates and solicitors across Malaysia and Singapore under each country’s Legal Profession Act and professional-conduct rules.

The regional direction is clear. Singapore’s Ministry of Law has issued dedicated guidance on generative AI in the legal sector, and its courts require lawyers to remain fully accountable for AI-assisted filings. Malaysia has strengthened its data-protection regime.³ The posture across the region is not prohibition but governed, supervised adoption.

Responsibility does not transfer to the machine. When a lawyer relies on an AI output, they remain the lawyer of record for it. That fact shapes everything that follows.

3Two ways it goes wrong: bad output, and bad input

Ungoverned AI fails in two directions. The first is bad output: the model inventing something untrue and presenting it with confidence. The second is bad input: client information leaving the firm’s control on its way into the tool. Different failures, different fixes.

The output failure has already produced a visible body of case law. An independent database maintained by the legal researcher Damien Charlotin catalogued more than 1,800 court decisions worldwide, as of July 2026, in which AI-fabricated citations or quotations reached a filing.⁴ A large share involved represented parties, not only litigants in person. Consequences have escalated from the original $5,000 sanction to five-figure penalties and, in at least one 2026 matter, suspension.

The input failure is quieter and, for a law firm, often more dangerous, because it leaves little paper trail. Published reporting by the data-security firm Cyberhaven has consistently found that a meaningful share of the material employees paste into public AI tools is confidential, with client data among the categories most often exposed.⁵ In a legal context, a single paste of confidential client work into a public AI tool can be a confidentiality breach. Depending on the tool’s retention terms, it can also be a disclosure that is difficult to undo. An output error is at least visible and correctable. An input leak is often discovered only after the fact, if at all.

4The direction of travel

Firms sometimes wait for a settled rulebook before acting. That is a mistake. The rulebook is arriving in layers, and the professional duties already bind regardless.

The regulatory picture is uneven and still forming. Singapore governs AI through voluntary frameworks, the IMDA Model AI Governance Framework and its generative-AI edition, alongside sector and court guidance. Malaysia has strengthened its Personal Data Protection Act and issued national AI ethics guidelines, with dedicated AI legislation still to come. Further afield, the EU’s AI Act is phasing in. A firm that waits for a single settled rulebook will wait a long time.

Underneath the regulation, two voluntary standards are becoming reference points professional firms are measured against. The US National Institute of Standards and Technology publishes an AI Risk Management Framework, with a dedicated profile for generative AI, that has become a de-facto baseline for structuring AI risk.⁶ ISO/IEC 42001, published in 2023, is the first certifiable standard for an AI management system, the AI equivalent of the ISO 27001 that firms already recognise for information security, and is becoming a procurement and trust signal.⁷ The broader shift is worth naming: the conversation is moving from one-off governance statements toward continuous assurance, being able to show on demand what AI is in use across the firm and that it is under control.

5Governance is the enabler, not the brake

It would be easy to read all of this as a case for caution, or even for a ban. It is the opposite. The firms that write a clear policy, classify their data, choose the right architecture and train their people are the firms that can then adopt AI safely, because they know which uses are allowed and which are not.

Governance converts ad-hoc use of AI into supervised and defensible use. It is the precondition for capturing the benefit, not a tax on it.

The honest framing of the choice is this. Doing nothing is not the safe option; it is the ungoverned option, and it carries the confidentiality and candour risks above with none of the controls. Banning the tools outright, or buying a product on features alone without asking where the data goes, trades one risk for another. Doing it well costs upfront effort in policy, classification and architecture, and returns a capability the firm can stand behind in front of a client, a regulator or a court.

6Where this leaves the firm

The imperative is not to decide whether to use AI. That decision has effectively been made by the people already using it. The task is to bring that use inside a deliberate governance framework before an incident forces the issue.

In practical terms that means four things, each the subject of a paper in this series. A data-classification scheme, deciding what information may touch which system. A deployment architecture that keeps the most sensitive inference under the firm’s control. A verification discipline that stops unchecked output reaching a client or a court. And a map of the firm’s professional and data-protection obligations onto the technology that serves them.

None of that requires the firm to become a technology company. It requires the firm to make, and write down, a small number of decisions it is currently leaving to chance. The remainder of this series is a blueprint for making them well.