A partner recently told us she had stopped asking junior associates to produce first drafts of standard commercial letters. The model did it faster, and — on a good day — better. Then she paused. "The problem," she said, "is that I no longer know which day is a good day."
That sentence captures the present moment in AI legal drafting more honestly than most vendor decks. The technology is genuinely useful. It is also unevenly reliable, and the unevenness is the part that should shape how in-house teams and firms adopt it.
Where AI drafting is genuinely strong
Large language models are now competent at a narrow but commercially significant band of legal work. They are good at tasks that are structurally repetitive, linguistically dense, and forgiving of a human second pass.
In practice, that means:
- First drafts of standard correspondence — Letters Before Action, chaser letters, basic demand letters, instruction notes. The structure is well-rehearsed; the model rarely invents the format.
- Clause extraction and comparison — pulling indemnity, limitation or termination provisions across a portfolio of supplier contracts and presenting them side-by-side.
- Plain-English translation of dense drafting — turning a heads-of-terms or a settlement deed into a client-readable summary.
- Schedule and table assembly — building a Schedule of Loss skeleton, a disclosure index, or a chronology from raw documents.
- Multilingual handling — drafting in English under UK law while explaining the document to a non-English-speaking client in their own language. This is where the productivity gain is largest and most defensible.
None of this replaces a qualified solicitor. All of it shortens the distance between a blank page and a reviewable draft, which is where most billable friction sits.
Where AI drafting quietly fails
The failure modes are the part that matters, because they are not visible in a demo.
First, citation fabrication. Models still invent case names, paragraph numbers, and statutory references that look plausible. The risk is highest when the model is asked to support a position it has been nudged towards, rather than to summarise something neutrally. UK practitioners have already faced judicial criticism for filings containing non-existent authorities; the regulatory mood is unforgiving.
Second, jurisdictional drift. A model trained predominantly on US material will, unprompted, reach for US contract conventions — "represents and warrants" stacks, indemnification language, choice-of-law phrasing that does not sit well in an English law contract. The drafting reads fluently. It is also subtly wrong.
Third, silent omission. Models are good at producing what you ask for and poor at flagging what you forgot to ask for. A clause may be missing an entire limb — say, a carve-out for fraud in a limitation clause — and the output will look complete. A junior lawyer, asked the same question, would more often say "should we also consider…".
Fourth, confidentiality posture. Many tools route prompts through third-party infrastructure with retention windows that are not always transparent. For privileged material, the question is not whether the vendor is reputable but whether the data path is one a regulator, client, or insurer would accept on inspection.
Fifth, the calibration problem. The model does not know when it is uncertain. There is no honest "I'm guessing here" signal. Every output arrives with the same level of confidence, which is exactly the wrong UX for legal work.
A UK-law lens on adoption
The SRA has been measured rather than restrictive on AI. Its published thinking — and the broader direction of SRA guidance on AI use in regulated practice — emphasises that existing duties already cover most of the territory: competence, confidentiality, supervision, and not misleading the court. The regulator is not asking firms to invent new ethics. It is asking them to apply the old ones to a new tool.
That framing is useful, because it cuts through the binary "should we use AI or not" debate. The honest answer is that most firms already are, often informally, and the governance question is whether that use is structured or shadow.
A defensible position involves at least:
- A written AI use policy, naming approved tools and prohibited use cases (for example, drafting opinions on contested points of law without partner sign-off).
- A clear data path for any tool touching client material — where prompts go, how long they are retained, whether they train future models.
- A supervision model that treats AI output as if it were produced by a paralegal in their first week: useful, but never filed, sent or relied upon without a qualified review.
- Client transparency appropriate to the engagement. Not every matter requires disclosure that AI assisted with a draft; some clients and some matters do.
- A logging discipline so that, if a draft is later challenged, the firm can reconstruct what the model produced and what the human changed.
Building the workflow
The firms getting value from AI legal drafting are not the ones with the most expensive licences. They are the ones who have decomposed their work into stages and asked, honestly, which stage the model should touch.
A workable pattern looks like this. The model produces a structured first draft from a defined template and a defined fact pattern. A solicitor reviews against a checklist that explicitly tests for the known failure modes — fabricated citations, US drift, missing carve-outs, jurisdictional language. The reviewed draft goes to a second human for substantive sign-off where the matter warrants it. Nothing leaves the firm without a name against it.
This is not glamorous. It is, however, the version of law firm AI adoption that survives a complaint, an audit, or a professional indemnity question. Document automation has always rewarded firms that invest in their templates and their review discipline; AI drafting is the same bargain, with sharper upside and sharper downside.
The tools are ready for serious work. The workflow is what decides whether that work is safe.
FAQ
Should we tell clients when AI has been used in drafting their documents? There is no blanket UK rule requiring disclosure, but the position depends on the engagement letter, the sensitivity of the matter, and what the client would reasonably expect. For high-stakes or bespoke work, transparency is usually the safer default.
Can we use a general-purpose AI tool, or do we need a legal-specific one? General tools can handle low-risk tasks if your data path and policy are clear, but legal-specific platforms typically offer better jurisdictional grounding, citation discipline and confidentiality posture — which materially reduces review burden.
What is the single most common mistake firms make when adopting AI drafting? Treating it as a productivity tool rather than a supervised drafter. The firms that get into trouble are the ones who skip the structured review stage because the output "looked fine".
For UK-law drafting with human solicitor review built into the workflow — Letters Before Action, contracts, Schedules of Loss and document review — see JustiScript.