A partner at a mid-sized London firm recently described her team's first six months with a generative drafting tool as "two associates' worth of speed, and one associate's worth of new supervision work." That ratio — net positive, but not the productivity miracle the vendor promised — is roughly what we hear across the market. The question for legal leaders is no longer whether to adopt AI legal drafting tools, but how to wire them into a workflow that produces defensible work product under English law.
This piece is written for in-house counsel and law-firm leaders running that evaluation now. We will look at what the current generation of tools genuinely does well, where they fail in ways that matter, and a practical model for safe deployment.
What the tools genuinely do well
Strip away the demos and the category breaks into a few honest strengths.
- First-draft scaffolding. For standard-form documents — NDAs, basic services agreements, employment letters, board minutes, simple Letters Before Action — AI tools produce a credible 70% draft in minutes. That draft is not signature-ready, but it removes the blank page.
- Structural review of long documents. Modern tools are good at extracting obligations, identifying defined terms used inconsistently, and flagging clauses that deviate from a house playbook. This is genuinely faster than human review for the mechanical layer.
- Translation and bilingual drafting. For cross-border work — and particularly UK ↔ China matters — AI translation of legal text has improved sharply. It is not yet a substitute for a bilingual lawyer's eye on operative clauses, but for correspondence, exhibits and internal summaries it is reliable.
- Knowledge retrieval inside a closed corpus. Pointed at a firm's own precedent bank or a client's contract estate, retrieval-augmented tools can surface the right clause faster than a trainee with a search bar.
- Summarisation at scale. Bundles, disclosure sets, regulatory correspondence — AI summarisation is now good enough to triage, provided a human verifies before anything is relied upon.
None of this is revolutionary. It is, however, real productivity, and firms that ignore it will find themselves quoting fees their competitors no longer need to charge.
Where they fail — and the failures that matter
The failure modes break into three categories, and they are not equally dangerous.
The first is fabrication. Tools still invent citations, misstate the holding of real cases, and hallucinate statutory sections that read plausibly. In drafting, this most often appears as confidently wrong references to legislation or invented sub-clauses of well-known statutes. The reputational and regulatory exposure here is obvious.
The second is jurisdictional drift. Many leading models are trained on a corpus dominated by US legal material. Ask for an indemnity clause and you may get something that reads natively to a Delaware lawyer and oddly to an English one — consequential loss carve-outs framed against the wrong common-law backdrop, "reasonable best efforts" language that has no settled meaning in England and Wales, or boilerplate that ignores how the English courts actually construe entire-agreement clauses.
The third, and most underrated, is silent omission. AI tools rarely tell you what they have left out. A draft shareholders' agreement that omits a deadlock mechanism, a Letter Before Action that fails to comply with the relevant pre-action protocol, a settlement agreement missing the statutory wording that makes it binding under English employment law — these errors do not look like errors. They look like clean documents.
For regulated firms, the SRA's guidance on the use of AI is consistent with the broader theme of its conduct rules: technology does not displace the duty of competence, confidentiality or supervision. The output is the solicitor's output. Treat it accordingly.
A workflow that uses AI safely
The firms getting this right are not the ones with the most sophisticated tooling. They are the ones with the clearest workflow rules. A workable model has five layers.
- Classify the matter before the tool touches it. Low-risk, high-volume work (standard NDAs, routine correspondence, internal summaries) can use AI from the first keystroke. High-stakes bespoke work (M&A SPAs, contentious pleadings, regulated advice) should use AI only for discrete, supervised sub-tasks.
- Lock the source material. Use retrieval-augmented tools pointed at your own precedent bank, your client's contract estate, or a vetted knowledge base. General-purpose chat interfaces, drawing on the open web, should not be drafting clauses that will be signed.
- Require a named human owner for every output. The fee-earner who accepts the draft owns every word of it. This is both an SRA-aligned position and a practical antidote to the diffusion of responsibility that AI tools quietly encourage.
- Build a verification pass into the workflow, not the lawyer's discretion. Citations checked against a primary source. Defined terms reconciled. Governing law and jurisdiction clauses confirmed against the deal sheet. Pre-action protocol compliance checked against a checklist, not memory.
- Log what the tool did. For client transparency, for professional indemnity insurers, and for the supervision file. If you cannot say which parts of a document were AI-generated and who reviewed them, you are not yet ready to scale the tool.
Document automation has existed in law firms for thirty years. What is new is the breadth of tasks AI can now plausibly attempt — and the corresponding breadth of supervision required. Law firm AI adoption is, in practice, a supervision problem dressed as a software problem.
A note on cross-border drafting
For work that crosses the UK–China corridor, the jurisdictional drift problem compounds. A model drafting an English-law distribution agreement for a PRC manufacturer needs to handle two legal traditions, two languages, and two sets of commercial expectations. The honest answer is that no general-purpose tool does this well end-to-end. The pragmatic answer is to use AI for the bilingual scaffolding and the mechanical review, and to keep qualified solicitors — on both sides — on the operative clauses and the risk allocation.
That is the design philosophy behind JustiScript: AI handles the volume, the translation and the first draft; UK-qualified solicitors handle what only they can. For firms and in-house teams building their own AI workflow, the principle is the same regardless of vendor — let the tool do what it does well, and supervise what it does not.
FAQ
Q: Does the SRA require us to tell clients when AI has been used in drafting their documents? A: The SRA has not mandated a specific disclosure form, but its guidance points clearly toward transparency with clients about how their matter is handled. Most firms are now addressing this in engagement letters rather than per-document notices.
Q: Can we use a public AI tool for a quick first draft if we anonymise the client information? A: This is risky. Anonymisation is harder than it looks, public tools may retain inputs for training, and confidentiality obligations under the SRA Code do not bend for convenience. Use an enterprise tool with a no-training contractual term, or do not use it at all.
Q: How do we price work where AI has cut drafting time by 60%? A: Most firms are moving the affected work to fixed fees or capped fees rather than defending hourly rates that no longer reflect the input. Clients notice quickly; the firms that lead this conversation tend to keep the relationship.
Serene Jade builds JustiScript for exactly this gap — AI-assisted UK legal drafting with qualified solicitor review on the clauses that matter.