A legal draft should come from checked facts, not a chat
AI drafting is safe when it fills the firm's own templates from checked facts, and an attorney approves before anything goes out.
Key takeaways
5 things that decide this
- 01AI drafting is safe in a law firm when it fills the firm's own templates from facts already checked in intake and the matter file, and an attorney approves the draft before it goes anywhere.
- 02A chat window has no matter record behind it. It writes from what you typed and fills the gaps with text that sounds right.
- 03The facts belong in the system the firm already trusts: Clio, MyCase, Filevine or whatever holds the matter. From there, the draft reads them.
- 04Templates the firm owns decide the structure and the wording that matters. The model fills fields and drafts the parts a template leaves open.
- 05Attorney review is a recorded step in the workflow, not a habit someone remembers. Yes, there's a trade-off: templates take partner time up front.
Why does drafting in a chat window invent facts?
Open a general chat tool, paste a few notes about a client, and ask for a demand letter. You get a confident letter back. It has a date, a sum, a deadline and a tone. Some of those came from your notes. Some came from nowhere, because a language model is built to produce likely text, and a likely letter has a deadline in it whether or not you gave one.
That is the core problem. The chat window does not know which facts are checked and which it made up, and it does not tell you. A typo in the client's name, a wrong party, a limitation date from another matter: each one reads exactly like the true facts around it. The person reviewing has to re-check every line against the file, which is the work the tool was meant to save.
There is a second problem. Text pasted into a consumer chat tool leaves the firm's systems. Whether that is a confidentiality problem depends on the tool's terms and your bar's guidance. We cover that in does sending client data to an LLM waive privilege and what ABA Formal Opinion 512 asks of a firm. This post is about the drafting itself.
Where should the facts in a draft come from?
From two places the firm already controls: intake and the matter record. Intake is where the facts first arrive. The matter record, in Clio, MyCase, Filevine or the system you run, is where they get corrected and kept.
Intake matters more than most firms expect. If intake is a free-text email or a phone note, the facts are unstructured from the first minute, and every later step inherits that mess. Structured intake fixes it at the door. On Lexpair, a legal platform we built, clients submit their case through guided, structured forms, and AI-assisted logic sorts each case by practice area before it reaches a lawyer. Lexpair is a lead-matching platform, not a drafting tool, but the lesson carries over. As the case study puts it: "If your intake is unstructured, no amount of volume fixes it."
Once the facts are structured and checked, the draft should read them from the matter record, field by field. Party names, addresses, dates, amounts and the matter number come from stored fields a person has already confirmed. At no point does the model retype them from a conversation.
- 01Party names, roles and addresses from the matter record, not from a prompt.
- 02Key dates (incident, filing, deadlines) from fields a person has confirmed.
- 03Amounts and terms from the engagement record or the source documents.
- 04Anything missing is flagged as missing, so the draft stops instead of guessing.
Why should the firm own the templates?
Because the template is where the firm's judgment lives. The clauses a partner insists on, the wording a court expects, the order of sections in a demand letter: none of that should be decided fresh by a model each time.
A firm template splits the draft into two kinds of content. Fixed text the firm wrote and approved, and slots. Most slots are facts pulled straight from the matter record. A few are open sections, like a short summary of events, where a model can draft from the checked facts and the source documents. The model works inside the template, not instead of it.
This is the trade-off to plan for. Good templates take partner time up front. Someone senior has to decide the standard wording and mark which parts may vary. Firms that skip this step end up with a tool that drafts in a different voice every time, and the review burden comes straight back. For the step-by-step build, see how to automate legal document drafting with AI.
How should attorney review work?
As a step in the workflow that a draft cannot skip. Not as a habit someone is trusted to remember on a busy afternoon.
In practice that means the draft lands in a review state. An attorney opens it next to the facts it was built from, edits what needs editing, and approves it. Only an approved draft can be sent, filed or saved as final. The approval is recorded with a name and a time, so the firm can later show who signed off on what.
We built the same pattern into TrialTriage in a different field: the AI ranks clinical trials, and a nurse reviews and finalizes every match. Nothing is finalized by the AI on its own. Legal drafting needs the same rule, with an attorney in the nurse's seat.
What does PremiumAudit.io's report drafting show?
The clearest example of this pattern we have shipped is PremiumAudit.io, an AI platform for insurance premium audits. Its reports are insurance documents, not legal ones, but the drafting pattern is the one this post argues for.
An audit moves through six steps on one shared record. It is scheduled. Documents are collected. The Claude API reads each document and pulls the figures into the record. Sums and classifications are checked against the policy terms. The system writes the report from the validated data. Then the carrier signs off.
The order is the point. Extraction and checking come before drafting, and a person signs off after it. A conversation never becomes the report. It comes from a record whose figures were checked first, and it fits the same format every time because it is built from a template. A law firm's version swaps payroll files for the matter record and the carrier for an attorney.
- 01Schedule: the work is opened against one record.
- 02Collect: source documents are uploaded to that record.
- 03Extract: the model reads documents and fills fields.
- 04Validate: figures are checked against the rules.
- 05Draft: the report is written from validated data.
- 06Review: a named person signs off before it is final.
Where should a firm start?
Pick one document type the firm drafts often, such as an engagement letter or a standard demand letter. Map which facts it needs and where each one lives today. If they are not in the matter record yet, that is the first job, and it pays off for every document after. Our page on legal document automation shows what we build around Clio, MyCase, Filevine or whatever you run.
If you want us to look at your setup, start with a 45-minute call. If it is a fit, we write a fixed-fee Blueprint from your real systems, and its fee is credited in full against the build. Here is how we work.
Questions this raises
01Can AI draft legal documents safely?+
Yes, when it fills the firm's own templates from facts already checked in intake and the matter record, and an attorney approves the draft before it is sent or filed. A chat window is the unsafe version, because the model fills gaps with plausible text and does not mark which facts it made up.
02Does sending client data to an LLM waive privilege?+
Not automatically, but the answer turns on the tool's terms, how it handles the data and your jurisdiction's guidance, so put it to your ethics counsel. Our answer page on privilege and LLMs walks through what to check before any client data reaches a model.
03What does ABA Formal Opinion 512 require?+
It sets out how lawyers' existing duties, including competence, confidentiality, communication and reasonable fees, apply to generative AI. Read our answer page on Opinion 512 for what it means for disclosure, and check your own state bar's guidance as well.
Related
- Legal document automation →What we build for drafting around Clio, MyCase and Filevine.
- How to automate legal document drafting with AI →The step-by-step build this post argues for.
- Does sending client data to an LLM waive privilege? →What to check before client data reaches a model.
- ABA Formal Opinion 512 →What the opinion changes about disclosure.
- PremiumAudit.io case study →Reports drafted from validated data, with a sign-off step.
