What Singapore's generative-AI guidance actually asks of a boutique practice

On 6 March 2026, the Ministry of Law launched its Guide for Using Generative AI in the Legal Sector. That same week, the High Court published a decision ordering counsel to personally pay S$5,000 each for citing cases that did not exist. Singapore's message to professional practices could not be clearer: use the tools, and own the output.

The MinLaw Guide, developed through a public consultation with more than twenty local and international stakeholders, is formally non-binding. That word offers less comfort than it appears to. As one contributor to the Guide put it: "The guide is non-binding. But it is not optional in any practical sense. It references binding professional conduct rules. It sets the standard against which legal professionals will be assessed if something goes wrong."

The disruption is not hypothetical

Step back from the guidance for a moment. The work professional services firms sell is being repriced in real time. First-pass legal research, contract review, due diligence, audit sampling, tax memo drafting: the tasks that once filled a junior's first three years are precisely the tasks the current generation of tools does quickly and cheaply. The Law Society-commissioned sustainability study released in June made the same observation from the other direction, warning that as AI absorbs the routine, document-heavy work on which juniors have traditionally learned the craft, the profession will need to rethink how lawyers are formed at all. When the Legal Year opened in January 2026, the announcement was a committee co-led by the Minister for Law and the Chief Justice, with lawyer attrition and the impact of artificial intelligence as its twin subjects. Those two topics did not land on the same agenda by accident.

The state is not standing still either. A national AI council chaired by the Prime Minister was convened in February 2026. MinLaw has moved beyond publishing guidance and into subsidised adoption: the PSG-Legal grant defrays half of a firm's first-year cost of pre-approved legaltech and AI tools, capped at S$45,000 a year, and the LIFT initiative goes further, deploying legaltech consultants into law firms to run the change process end to end. Read that policy design closely, because it contains an admission. The obstacle was never access to tools. It was that adoption is a change-management problem, and most firms have nobody whose job that is.

The courts moved first

Before any guide existed, the Supreme Court's Registrar's Circular No. 1 of 2024 (effective 1 October 2024) set the tone. Generative AI is not prohibited in court documents, but users remain fully responsible for everything filed, and must be prepared to identify AI use and explain how output was verified if the court asks.

The case law since reads like an escalating tariff. In 2025, a lawyer was ordered to personally bear S$800 in costs after citing a fictitious AI-generated authority. In March 2026, the High Court ordered S$5,000 in personal costs per counsel across related suits, noting it was "plainly evident" that an AI tool had hallucinated plausible-sounding cases. By May 2026, in a third decision, the court made the sharpest point of all. Counsel denied using AI, and the court held that it did not matter: the sanction attaches to the failure to verify what you put before the court, whatever produced it.

The regulators did not write these rules to slow you down. They wrote them because the profession was already tripping over its own tools.

The accountants got their lesson too

Lawyers are not the only profession holding a cautionary tale. In October 2025, Deloitte's Australian firm agreed to partially refund a government department for a A$440,000 report found to contain AI-generated errors, including citations to academic papers that did not exist and a fabricated quote from a federal court judgment. The revised report disclosed — after the fact — that a generative AI system had been used; the refund ran to roughly A$97,000. Substitute a valuation report, an audit workpaper or a due diligence memo for that policy review and the lesson transfers without modification. The signature on the report owns everything beneath it, however it was produced.

The disruption map looks much the same across the advisory professions. Audit teams are moving from sampling to full-population testing, with AI flagging anomalous journal entries for human follow-up. Valuation teams use it to screen comparables and assemble first drafts. Tax teams run research assistants that read legislation and rulings in seconds. Deal teams compress information memoranda and due diligence summaries into afternoons. The Big Four have committed billions to AI partnerships precisely because this is where the leverage is. And in every case the pattern repeats: the routine layer compresses, the review layer becomes the job, and the professional standard — scepticism for the auditor, reasonable basis for the valuer, verification for the lawyer — does not bend to accommodate the tool.

What the MinLaw Guide asks, in practice

Read the Guide itself and the demands on a boutique practice are concrete and manageable:

• Verify before you rely. Output used in legal work must be correct, accurate and fit for purpose. High-stakes work such as submissions, contracts and advice needs a human reviewing everything; only low-risk, reversible tasks can run on lighter supervisory checks.
• Match the tool to the data. The Guide tiers tools from free consumer products to enterprise and proprietary systems. With free tools: disable training and retention, anonymise with placeholders, and never input client-confidential material. With enterprise tools: contractual bans on training use, data-residency checks, recognised security certifications, deletion and audit rights.
• Tell clients when it matters. Consider disclosure where AI use is substantial, affects cost, or involves data handling a client might not accept, and offer clients the ability to opt out. The Guide includes template engagement-letter clauses; one boutique's version runs to a single sentence.
• Give AI an owner. Small practices are asked to appoint an AI lead (larger ones, a committee), classify their data, prepare an incident-response protocol aligned to data-breach notification duties, and train staff annually. The Guide ships with a sample internal policy.
• Supervise everyone. Responsibility extends to paralegals, secretaries and support staff, and juniors are to be developed into competent professionals rather than prompt operators.

None of this is exotic. It is essentially the discipline a well-run practice already applies to outsourcing and cloud services. The Law Society's cloud computing guidance has required the data-residency and confidentiality analysis since 2017, and the new Guide cross-references it directly.

A subscription is not a capability

The tools themselves are excellent, and expensively so. Harvey raised US$200 million in March 2026 at an eleven-billion-dollar valuation; its European rival Legora is valued above five billion. Thomson Reuters has CoCounsel, LexisNexis has Lexis+ AI, and a long tail of specialist products covers contracts, disputes, tax and audit. Reported pricing for the enterprise platforms runs from roughly US$300 to US$1,500 per user per month, with minimum-seat floors that can push an annual commitment into six figures before anyone has drafted a thing. For a thirty-professional firm, whether it practises law, audit or corporate finance, one flagship subscription is a serious line item. Three of them is a budget crisis.

And here is the quieter problem. Being able to prompt a copilot to summarise a bundle is not the same as making these systems work inside the way your firm actually runs a matter. The professional who gets real leverage from AI has usually spent unglamorous hours on it: mapping the workflow, building the templates and precedent banks the tool draws on, testing where it fails, writing the guardrails, training the team, and vetting every plug-in and connector before it goes anywhere near client data. That is not prompting. That is systems work, and it consumes the one resource a billing professional does not have. Enthusiasts can happily spend the weekend buying a Mac mini to run OpenClaw for the joy of it — the hobby is popular enough that CNN credits it with making the humble Mac mini the hottest product Apple sells. You have submissions due.

"AI" has become a word like "software": true of everything and useful for nothing. Anyone can use AI for something. Almost nobody has the time to refine it for the specific, regulated, confidentiality-bound way a professional firm works. Subscribing to a famous tool does not close that gap. It moves the gap inside your budget.

Red team: defending your AI attack surface

Think of it the way a security team thinks about a network. The attack surface is bigger than any list — new connectors, skills, agents and models arrive weekly, each with its own way in — so defence never comes from memorising every exploit. It comes from the same discipline, applied every time. Below is a five-round drill against representative threats: a connector that over-reaches, a document that fights back, an agent that does too much, and two more. The second is the whole premise of a corner of the internet given over to poisoning AI systems — adversarial content that needs no malware, only a tool that trusts what it reads. Take the AI lead's chair, and hold the line.

Dependence has a price

There is also a harder-nosed question that adoption enthusiasm tends to skip: what, exactly, are you becoming dependent on? Enterprise data terms deserve the same scrutiny the Guide demands of free tools, because the difference between a platform that trains on your prompts and one that contractually cannot is the difference between a productivity gain and a confidentiality incident. Prices are set by vendors who are themselves hostage to compute costs; a workflow that pencils out at today's token prices may not survive next year's. Models get deprecated, products get acquired, terms get rewritten at renewal. And regulation can reach further than any renewal notice. In June 2026, a US government directive citing national security led Anthropic to suspend access to its newest Claude models worldwide, including all public access. The restrictions were lifted at the end of June, and the most capable tier remains limited to vetted organisations. For two and a half weeks, every firm whose workflow ran on those models found out what its fallback was. Access to frontier capability can narrow overnight, and not by your choice or even the vendor's.

None of this argues for abstinence. It argues for architecture. Keep the workflow yours and the tool replaceable: documented processes, portable data, an exit plan, and more than one way to get the work done. Firms already think this way about every other critical supplier. AI deserves the same treatment, not more romance.

Beyond the legal sector

Advisory and finance-adjacent firms have their own regulators to watch. The national framework, IMDA and PDPC's Model AI Governance Framework for Generative AI (2024), extended in January 2026 with a world-first framework for agentic AI, sets the expectations vocabulary: accountability, bounded risk, meaningful human checkpoints. The PDPC closed a consultation on personal data in generative AI on 1 July 2026. For financial institutions, MAS has proposed comprehensive AI risk-management guidelines, built on its longstanding FEAT principles. That consultation closed in January 2026 and finalisation is expected later this year, with expectations around AI inventories, risk-materiality assessment and board accountability that will reach even small licensed firms, proportionately. Valuation, audit and advisory practices outside the MAS perimeter are not off the hook either. The PDPA applies to everyone, professional-body codes carry their own confidentiality and competence duties, and engagement letters make discretion contractual even where no regulator is watching.

The boutique advantage, quietly

The sentiment in the market is not fear. Industry coverage of the Guide's launch was strikingly upbeat. But there is a gap between enthusiasm and governance, and that gap is where the risk lives. Here a boutique holds an unglamorous advantage. A ten-lawyer firm or a twelve-person advisory practice can appoint its AI lead, choose its tools, write its policy and train everyone in a fortnight; a thousand-person firm cannot. The firms that will wear AI comfortably in this market are not the ones with the biggest innovation budgets. They are the ones that did the small, boring things early, and can say so calmly when a client's general counsel asks.

And they will ask.