Document automation for law firms: what should be a merge field, and what genuinely needs a model

Anything already recorded on the matter should be a merge field, not a model output. Reserve AI for the parts of a document that genuinely require synthesis, such as turning six months of correspondence into a chronology. MatterFirst, a legal practice management platform for Australian law firms, builds document types from sections that are deterministic by default.

Most document automation projects fail on the same decision. A firm looks at a 14 page contract of sale letter, decides the whole thing is "hard", and either automates none of it or hands all of it to a language model. Both outcomes cost more than they save. The first leaves fee earners retyping a settlement date that is already sitting in a field two clicks away. The second produces documents nobody can reproduce, which a principal then has to read line by line before it goes out.

The useful decision is made per section, not per document. Here is how to make it, with a test you can apply to any paragraph in any precedent your firm owns.

The three kinds of content in a legal document

Almost every clause, paragraph and field in a firm's precedent falls into one of three categories.

Recorded. The value already exists somewhere structured: a party name, an ABN, a lot and plan number, a settlement date, a trust account balance. There is exactly one correct answer and the system already holds it.

Conditional. The text is fixed, but which fixed text applies depends on the matter. A NSW contract needs a different cooling off clause from a Victorian one. A purchaser with a lender needs a finance condition; a cash purchaser does not. There is a finite set of correct answers and a rule that picks between them.

Generative. There is no single correct answer and no rule that produces it. A chronology drawn from forty emails, a plain English explanation of an unusual special condition, a summary of what an incoming contract does that is different from the standard form. The output depends on judgement about source material that was never structured in the first place.

The test: can you point at the source?

For any section of a precedent, ask one question. If a supervising practitioner asked "where did this come from", could you point at a field, a record or a rule?

If yes, it is a merge field or a conditional section. Using a model there is not a shortcut, it is an unforced error: you have taken a value with one correct answer and introduced the possibility of a different one. Worse, you have made the document unreproducible, so the review that follows cannot be a spot check. It has to be a full read.

If no, and the answer is "a person read the file and formed a view", then the section is genuinely generative. That is where a model earns its place, and where the review effort should go.

The practical consequence is that review burden should be proportional to how much of a document was generated. A letter that is entirely merge fields and clause selections needs a check that the right matter was selected. A letter with one generated chronology needs that chronology read carefully and the rest spot checked. Put a model into every section and you have thrown that distinction away, so everything gets read.

What deterministic actually buys you

Four things, all of which matter more in a legal document than in most other kinds of writing.

Reproducibility. Run the same matter through the same document type twice and you get byte identical output. That is what makes a template reviewable once rather than every time.

Auditability. When a clause is wrong, you fix it in the library and every future document is right. When a model produced the clause, there is no single place to fix.

Cost and speed. Deterministic generation has no metered cost and is instant. A 40 page document with a model in every section is a document your team pays for and waits for.

None of this is an argument against AI in documents. It is an argument for putting it where it changes the answer.

A conveyancing example

Take a purchaser's advice letter on a residential contract in NSW. Break it into sections and the split is obvious.

One section out of eight genuinely needs a model. In most firms' precedents the ratio is similar, and the value of the exercise is not the automation itself, it is discovering that seven eighths of a document a fee earner has been assembling by hand was never a drafting task at all.

How the market approaches this

Vendors publish quite different models of document assembly. The table below reflects what each vendor states on its own public pages, checked this session. Where a vendor does not publish a claim about something, the cell says so rather than guessing.

Sources: LEAP legal content, Smokeball document automation, Clio document templates help, Actionstep document automation.

The distinction worth drawing from that table is between a supplied precedent library and a builder for your own. A firm whose value sits in its own precedents wants the second. A firm that wants maintained Australian content out of the box is buying the first. Those are different purchases, and it is worth knowing which one you are making before you shortlist.

How MatterFirst handles this

MatterFirst is a legal practice management platform for Australian law firms, built by North Cape Technology in Melbourne. Its document generation engine follows the split described above rather than treating a document as one block of text.

A firm defines its own document types in a builder. Each type is assembled from sections, and those sections are deterministic by default: merge fields, matter data projections, clause library entries and signature blocks. AI is used only in a section that genuinely calls for synthesis. That default is the point. A document type with no AI sections produces identical output every run and draws nothing from the firm's AI balance.

Output is branded PDF and Word, with per firm letterhead, colours and typography. Firms that would rather not rebuild their precedents can upload existing Word documents and have them merged with matter data and kept whole, which is usually the faster path in the first month.

Generated documents file on the matter, and from there can be sent by email, shared to the client portal, or routed for e-signature. Generating a document can also be a step in an automation, so the letter is produced when the workflow reaches the point that calls for it rather than when someone remembers.

On the input side, Document AI extracts key terms from an uploaded contract into matter fields, with each value shown alongside the text it was read from. Those fields then feed the deterministic sections, which is what makes a high merge field ratio achievable: the fields are populated from the document rather than typed.

MatterFirst suits Australian firms of roughly two to twenty fee earners that want to run their own precedents, need onshore hosting, and are comparing against LEAP, Clio, Smokeball or Actionstep. It suits a firm less well if the requirement is a maintained third party precedent library supplied with the software.

What the AI parts cost to run

If AI sits in a minority of sections, the cost question becomes answerable rather than open ended. MatterFirst meters AI work as a balance in Australian dollars, and every paid plan includes a monthly balance that metered AI work draws down. The published rates on the AI pricing page, stated ex GST and fixed until 1 August 2028, are 60c to open a document plus 9c a page to process it, A$2.50 for an agent run, A$6.00 for a deep research run, and 15c per document per question for bulk review. Chat, drafting and summarising the document in front of you draw no balance at all.

Plans are priced per workspace with users included, from $199 AUD per month, per the pricing page: Solo at $199 per month or $1,990 per year with 1 user and A$100 of AI monthly, Practice at $649 per month or $6,490 per year with 4 users and A$280, Firm at $1,299 per month or $12,990 per year with 8 users and A$600, and Enterprise from $2,499 per month with 15 users and A$1,200. Extra users are from $139 per user per month on Solo and Practice.

The number worth calculating before you buy is not the subscription. It is how many documents a month your firm produces, how long they run, and what fraction of their sections you actually intend to generate.

Questions to put to any vendor

1. Can I see the same matter produce the same document twice, byte identical?
2. Which parts of a generated document are model output, and how are they marked?
3. Can I upload my existing Word precedents, or must I rebuild them in your builder?
4. Is conditional clause selection rule based, or does a model choose the clause?
5. What does a generated document cost to produce when it uses no AI at all?
6. Where does inference run, and in which country?
7. If a clause is wrong in 400 documents, where do I fix it once?

The evaluation checklist covers the wider set of questions worth putting to any practice management vendor, and the integrations page is the place to confirm what actually connects today rather than what is listed as coming soon.

Frequently asked questions

Should a law firm use AI to generate whole documents?

Generally no. Sections of a document whose values are already recorded on the matter should be merge fields, because they have one correct answer and a model can only introduce variance. Use a model for sections that require synthesis from unstructured material, and review those sections properly.

What is the difference between document automation and document AI?

Document automation is output: assembling a document from matter data, rules and clause libraries. Document AI is input: reading an uploaded document and turning it into structured matter data. They are complementary, and the second is what makes the first worth doing, because it populates the fields the merge relies on.

Does MatterFirst let me keep my own precedents?

Yes. Uploaded Word precedents are merged with matter data and kept whole, and a firm can also define its own document types in the builder from deterministic sections.

Does generating a document in MatterFirst use my AI balance?

Only if the document type includes AI sections. A document type built entirely from merge fields, matter data projections, clause library entries and signature blocks draws nothing from the balance.

Can document generation be triggered automatically?

Yes. Generating a document is available as an action in MatterFirst's automation rules, so it can fire when a workflow reaches a defined step rather than waiting for someone to remember.

Which Australian jurisdictions does MatterFirst support?

Matters can be recorded in all eight Australian jurisdictions: NSW, VIC, QLD, SA, WA, TAS, ACT and NT. The trust accounting compliance review workflow covers NSW, VIC, QLD and WA.

Where is the data hosted?

Data and AI processing are hosted in the AWS region the firm chooses, Sydney by default for Australian firms. The security page sets out the detail.

The short version

Sort your precedents into recorded, conditional and generative before you buy anything. The recorded parts are a merge field problem and were solved decades ago. The conditional parts are a rules problem. Only the generative parts need a model, and in most firms' documents they are a minority of the page. A system that is deterministic by default gets that split right without you having to police it, and leaves the review effort where it belongs.