Ask five legal-tech vendors what document automation means and you will get five products. Strip away the branding and there are three layers, and they build on each other.
Layer 1: Merge-field templates
The client care letter, the terms of business, the standard letters at each stage of a conveyancing matter. These are the same document every time with the names, dates, fees and addresses swapped in. Clio, LEAP, Actionstep and Word itself all do this with merge fields: the data comes from the matter record, the template fills, a person reads and sends.
Most firms have this, partially. The gap is usually that the templates are inconsistent (five versions of the care letter on five desktops), the matter data is incomplete so half the fields are filled by hand anyway, and nobody owns the templates. The automation is cheap; the discipline is the work. Clean the data at intake and give one person ownership of the precedent bank and this layer runs itself.
Layer 2: Structured precedent assembly
A lease, a shareholders' agreement, a will, a settlement agreement. These are not one document with blanks; they are documents assembled from clauses according to answers. Tenant or landlord. Joint or sole. With or without a restrictive covenant. The precedent bank holds the clauses and the rules; a questionnaire drives which clauses appear in which order.
This is the layer where most of the fee-earner time in drafting actually sits, and it is deterministic: the firm's own clauses, the firm's own rules, a document a supervising solicitor has already approved in every permutation. Done well it is the single biggest drafting saving a firm can make, and it involves no AI at all. Done badly (rules nobody maintains, clauses that drift from the approved version) it produces confident wrong documents quickly, which is the same failure mode people fear from AI, arrived at the old-fashioned way.
Layer 3: AI first drafts, from your own precedents
The new layer. A language model takes the matter facts, the instructions and the firm's precedents and produces a first draft of something that is not a pure assembly: a letter of claim with the facts woven in, a bespoke clause, a summary of a long document, a response to the other side. It is useful precisely because it handles the parts that templates and assembly cannot.
It is also where the SRA's August 2026 warning notice lives. The failures it describes are unchecked output and fabricated citations. So the rule for this layer is simple and non-negotiable:
- Draft from the firm's own precedents and the matter file, never from the open internet. The model is given the firm's clauses, the firm's style and the facts, and told to work from those. Anything it adds from elsewhere is a flag, not a feature.
- Nothing cites an authority a person has not checked. If the draft mentions a case, a statute or a rule, the fee earner verifies it before it leaves the firm. Better still, the system marks every citation it produces so they cannot be missed.
- Every AI draft is reviewed by a solicitor before it is sent, filed or relied on. It is a draft in the same sense a trainee's draft is a draft. Faster, tireless, and not the final word.
- Business-tier models only, under contracts that forbid training on the data. No consumer chatbots, ever, for anything from a client file.
The honest summary: layer 3 is a speed gain for competent people, not a replacement for them, and it is only safe once layers 1 and 2 are in order, because they are what the model drafts from.
Where to start
Audit the precedent bank first. Count the versions of each core document, find the owner, pick the approved one, retire the rest. Then wire the matter data so the merge fields fill themselves, which usually means fixing intake. Then build assembly for the two or three document types that eat the most fee-earner hours. Only then add AI drafting, scoped to the tasks where it genuinely saves time, with the four rules above enforced by the system rather than by memory.
That order is slower to sell and faster to pay back. The law firms page has pricing for each stage and the compliance framework the whole thing sits inside; the intake automation guide covers the data-cleaning half.
Founder, Anco AI. Builds the systems this site talks about, and runs his own business on them.
