For structured, repeated documents such as court forms and consent orders, a rules-based tool like Gavel Workflows is the safer choice, because the same answers always produce the same output. Generative AI suits narrative sections and first drafts, with checking. Gavel plans start at $83 a month billed annually; business assistant seats cost $20 to $25.
Gavel Workflows is the former Documate, and Gavel's site says it has since been acquired by Relativity. Gavel Exec, the company's AI contract review tool for Word, is a different product. The Workflows list prices assume annual billing, and paying month to month costs about a fifth more. The better answer for most firms is not either tool but a split, where rules assemble the documents and AI drafts the narrative from the same intake answers.
How rules-based assembly works in a family practice
Gavel describes Workflows as turning intake answers into finished Word documents and PDFs automatically, in three steps: build an intake form, collect the client's information, and generate the documents. Between the form and the output sit rules. Gavel's page calls these conditional logic and calculations: branching so the workflow changes with the answers, and sums done for you. You can start from Word templates you already use, or from the template library it describes, which includes pre-built templates and court forms.
In family work the logic is concrete. Here is how a few answers might steer a divorce set, as an illustration of the kind of rules you would build.
| Intake answer | What the workflow does |
|---|---|
| Children under 18: yes, two | Adds a parenting plan section for each child and asks for each child's date of birth and living arrangement |
| Children under 18: no | Removes every parenting clause and the child-related fields on the financial form |
| Either party self-employed | Adds a business income schedule and a prompt for the last three years' accounts |
| Family home owned jointly | Adds the property schedule and pulls the address, value and mortgage into the settlement clauses |
| Weekly income entered | Calculates the annual figure and carries it into the maintenance and support sections |
| Parties agree on everything | Switches the document from a contested application to a consent order draft |
Three properties make this attractive for official documents. The output is deterministic, so you can explain exactly why a clause appeared. A figure typed once appears everywhere it belongs, which removes the classic error of a name or amount differing between documents. And the build is yours: a firm that wants a particular clause order or house style can encode it. The price is the build and the upkeep, covered below.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat generative AI does well on the same documents
A generative assistant has no intake form and no rules. You give it facts and instructions, and it writes. That makes it strong exactly where rules are weak: free text, tone, summaries of messy source documents, and a first draft of an unusual clause. It is weak where rules are strong: the same facts can produce different wording on different runs, a figure can change between two documents in one matter, and a mandatory field can be left out without any warning. Sorting the document set by those strengths gives most of the answer.
| Document | Better route | Why |
|---|---|---|
| Official financial statement | Rules-based | Fixed fields and sums that must match the court's layout every time |
| Application or petition with standard clauses | Rules-based | Conditional clauses and party details that must be identical throughout |
| Hearing notices and routine covering forms | Rules-based | Repetitive and low in judgement |
| Parenting plan | Rules for structure, AI for narrative | Standard sections plus arrangements described in plain words |
| Consent order | Rules for structure, a lawyer for custom terms | Clause logic matters, and custom terms need judgement |
| Background or statement of facts | Generative AI with checking | Free text built from facts already captured |
| Client letters and progress updates | Generative AI | Tone and variety matter more than exact wording |
| Summaries of bank statements and disclosure | Generative AI with checking | Unstructured sources; see the tutorial on summarising financial disclosure safely |
Sensitive family matters add a confidentiality layer: children's details, addresses and safety information. Whichever route you use, keep a written rule about what may be pasted into any tool and use business plans that state they do not train on your data by default. ChatGPT Business and Claude Team both say so on their pricing pages, at $20 a seat a month on annual billing or $25 monthly, with a two-seat minimum.
Costs side by side for a four-fee-earner firm
Gavel Workflows lists four plans. As of November 2026 the page shows Lite at $83 a month (one builder seat, Word and fillable PDF automation, Clio Manage integration), Standard at $210 and Pro at $290 (more seats, a legal template library, Stripe payments, DocuSign, white-label and embedding), all billed annually; billed monthly they are $99, $250 and $350. Scale starts at $417 a month and is annual only. It offers a seven-day trial with no card, and says annual plans billed upfront get two months free. The plan cards and the comparison tables list different seat, template and workflow limits for the same plans, so get the limits for your plan in writing. For reference, Knackly, another rules-based tool, lists Starter at $250 a month ($209 on yearly billing), Done For You at $500 ($417) and Professional at $1,000 ($834), all month to month.
Price per month is the small part. Take a firm with four fee earners and 120 new matters a year, each needing the same six documents. Every hour figure below is an assumption to replace with your own, valued at $75 an hour.
| Item | Word templates by hand | Rules-based workflows | Generative AI prompts |
|---|---|---|---|
| Build time | None | 72 hours (6 workflows at 12 hours) | 16 hours (a prompt library) |
| Yearly upkeep | None | 18 hours (3 per workflow) | Included in build |
| Time per matter | 4.0 hours | 1.5 hours | 2.5 hours, including full checking |
| Hours for 120 matters | 480 | 270 | 316 |
| Software for the year | None | $3,480 (Pro at $290 a month, billed annually) | $1,200 (4 seats at $25 monthly) |
| Total, time plus software | $36,000 | $23,730 | $24,900 |
The totals for the two automated routes are close, which is the lesson: they have opposite shapes. A workflow costs a lot to build and little per matter. A prompt costs little to start and more per matter, because every output needs checking. Where they cross depends on volume. Setting the two totals equal gives about 104 matters a year: above that the workflow route is cheaper, below it prompting is. The result is very sensitive to the checking time you assume for generative output. At 2.0 hours a matter the break-even rises to about 209 matters; at 3.0 hours it falls to about 70. So time your own checking on five real drafts before you decide.
Where rules-based automation fails, and how to catch it
A workflow does exactly what it was told, which is both its strength and its risk. Five failures recur.
- A stale template. The court issues a new version of a form and your workflow keeps producing the old one. Nothing warns you. Give each workflow an owner, a version number on the first page, and a standing monthly check against the court's current forms.
- A logic error that copies itself. A rule that drops the wrong paragraph when a client answers "no children" will do so in every matter until someone notices. Test each workflow against a set of dummy matters before it goes live, and re-run the set after every change.
- Wrong answers in. The workflow is only as good as the intake. A client who types a joint account balance as their own produces a clean, confident, wrong form. Add validation: dates in order, shares that add up to 100%, required fields that cannot be skipped.
- Edge cases nobody coded. Two children with different arrangements, a business owned through a company, assets held abroad. If the workflow has no branch for it, the document is silently incomplete. Keep an exceptions flag that stops the output and tells a lawyer to draft by hand.
- Free text it cannot write. Narrative sections and unusual terms are outside rules. The temptation is to bolt on generic wording; resist it.
The cheapest defence is a set of eight dummy matters kept in a folder: a simple one, one with children, one self-employed, one with a jointly owned home, one with all three, and three awkward cases you have met in practice. Run them through after every change, and compare the output with a lawyer-drafted answer that has been checked line by line. If you build more than a handful of workflows, the same discipline applies to your precedent library; building a precedent library with AI covers keeping that current.
Where generative AI fails on legal documents, and how to catch it
A generative tool fails in quieter ways, because its output always reads well. These are the failures to design around.
- Drift between documents. A child's name, a date of separation or an amount differs between the application and the consent order. Fix it by pasting the same structured fact block into every prompt, not by describing the facts afresh each time.
- Silent omissions. A mandatory section or field is left out. Fix it with a checklist of required elements for each document type, applied after every draft.
- Invented authority. A case or provision cited that does not exist, or exists and says something different. Ban citations in the prompt, or require them from a list you supply, and verify every one that appears.
- Arithmetic slips. Totals and percentages are not reliable from a language model. Calculate in a spreadsheet and paste the results in.
- Run-to-run variation. Two runs of the same prompt give different wording, so you cannot rely on a draft having been "checked" if it is regenerated. Freeze the approved text.
- Tone in sensitive messages. An email to a distressed client can be technically right and badly judged. A person reads every client-facing message before it goes.
A short test makes these visible. Give the same fact block to the assistant three times and ask for the same background section. Compare the three drafts line by line and note every difference in a fact, an order or a clause. A firm that sees more than a couple of differences in a 250-word section has learned how much checking it will really need. The wider question of which matters suit this at all is covered in which matter types suit generative AI and which do not.
A hybrid design: rules assemble, AI drafts the narrative
Most firms get the best of both by splitting each matter's documents at the line between structure and story. The workflow collects the intake once, validates it, and assembles everything that is structured. The same answers are then exported as a fact block that a generative tool turns into the narrative sections, which a person checks against the block. The facts live in one place, so drift between documents has nowhere to start.
Here is the prompt for the narrative step, with an illustrative fact block pasted under it.
Write the "Background" section of the consent order draft in plain, neutral English,
about 150 words. Use only the facts in the block below. Do not add any fact, reason,
adjective about the relationship or legal citation. If a fact needed for a sentence is
missing, write [MISSING: item]. Keep names as Parent A, Parent B, Child 1, Child 2.
FACT BLOCK
Parent A: date of birth 02/06/1985; Parent B: date of birth 19/11/1983
Date of marriage: 12/09/2012; date of separation: 14/03/2025
Child 1: born 04/05/2015; Child 2: born 21/08/2018
Both children live mainly with Parent A; Parent B has them every second weekend
and for half of each school holiday.
Family home: jointly owned; to be sold; proceeds split 60/40 in favour of Parent A.
Parent B pays monthly child maintenance of 640 (currency as per the form).A realistic draft comes back looking right, and the checking is where it earns its place. Imagine it opens: "The parties married on 12 September 2012 and separated amicably on 14 March 2025." Two corrections follow. The word "amicably" is not in the fact block, so it goes: the prompt said no adjectives about the relationship, and the model supplied one anyway. And a sentence later it may write that the children "spend alternate weekends with Parent B", omitting the half of each school holiday. Your checker compares each sentence with the block, deletes what is unsupported and restores what is missing. Do it with a pen on a printed copy for the first month; a tick against each fact is faster than rereading for a feeling.
The same structure works for the other narrative jobs. Client updates use the same fact block plus a short note of what happened this week, and AI-assisted client intake with urgency triage shows how to collect the facts safely in the first place. The generative step never sees anything the lawyer would not paste into any tool, which is a rule worth writing on the wall.
A one-week test on three documents
Do not test on a whole practice. Test on three documents that represent the three kinds of work: a court form, a document with conditional clauses, and a narrative. Use ten dummy matters, including three awkward ones.
- Day one: choose the documents and the matters. Pick the official financial statement, the consent order and a client progress letter. Write ten dummy matters on one page each, including a self-employed party, two children with different arrangements, and a jointly owned home with a mortgage.
- Days two and three: build the rules-based version. Use a seven-day trial of a workflow tool with the financial statement and consent order. Time yourself, and note every point where the tool made you decide something you had not anticipated.
- Day three: build the prompt version. Write the prompt and fact block for the same documents, and run it on all ten matters.
- Day four: score both. Check every field and clause against a lawyer-drafted answer key.
- Day five: decide. Weigh wrong or missing items first, build time second, per-matter time third.
The scorecard below is filled with invented results to show the reading. Neither column describes a real product, so the routes are labelled A and B.
| Measure over 10 dummy matters | Route A: rules-based | Route B: generative prompts |
|---|---|---|
| Fields wrong or missing on the financial statement (of 340) | 4 | 19 |
| Clauses wrong or missing in the consent order (of 90) | 3 | 11 |
| Facts that differ between documents | 0 | 6 |
| Awkward cases handled without a lawyer redrafting (of 3) | 1 | 2 |
| Client letter judged ready to send with light edits | 2 of 10 | 8 of 10 |
| Build time | 26 hours | 5 hours |
The pattern in that illustration is common to the logic of the two routes, not to any product. Rules win on fields, clauses and consistency, and lose on awkward cases and letters. Prompts win on letters and edge cases that need judgement, and lose on the structured documents. So the decision is a split, not a winner.
Which route each kind of document should take
Turn the test into a standing rule your whole team can apply, so nobody argues about it matter by matter. The thresholds are my rules of thumb built on the worked example, not vendor guidance.
| Question | If yes | If no |
|---|---|---|
| Is it an official form or a document whose clauses must be identical every time? | Rules-based workflow | Go to the next question |
| Do you produce it more than about 100 times a year? | Build a workflow even for non-official documents | Use a prompt with a fact block and a checklist |
| Is most of it free text built from known facts? | Generative draft, with every fact checked against the block | A lawyer drafts it, with a template for the standard parts |
| Does it contain a calculation? | Calculate in the workflow or a spreadsheet, never in a prompt | No restriction |
| Would an error be hard to spot on a second reading? | Second-person check before it leaves the firm | Author's check is enough |
Two housekeeping rules sit on top. Every workflow and every approved prompt has a named owner and a review date, and every document that leaves the firm records which route produced it and who checked it. Without those, either route decays quietly.
Questions to put to Gavel, to Knackly and to any AI vendor
List prices are the opening position. These questions turn them into something you can budget.
- The plan cards and comparison tables give different seat, template and workflow limits. Which limits apply to the plan I am buying, in writing?
- What counts as a session, given the plans cap sessions at 100, 300 and 600 a month? Is a client completing an intake form one session?
- Gavel's site says it has been acquired by Relativity. Will the pricing, support and roadmap for Workflows stay as they are, and for how long can I lock the price?
- Where is intake data stored, how is it encrypted, how long is it kept, and what happens to it if I cancel?
- Can I export my workflows and templates in a form I can use elsewhere?
- For a generative assistant: is business data excluded from training by default, what can an administrator see, and can chat history be deleted on a schedule?
- Who updates the court forms in the template library when a form changes, and how will I be told?
If you want to weigh these tools against the wider field of family law software, the checklist for family law AI software covers supplier questions, and the comparison of Spellbook and Gavel Exec for contract review explains how the AI contract review side of Gavel differs from the Workflows product described here. For an owner who would rather work out the split with someone, that is what my AI implementation consultation covers.
Questions family law firms ask about automating documents
Is Gavel Workflows the same product as Documate?
Yes. Documate rebranded as Gavel in January 2023, and Gavel Workflows is the document automation product. Gavel Exec is a separate AI contract review tool for Word. Gavel's site also says it has been acquired by Relativity, so ask how that affects pricing and support.
Can ChatGPT or Claude fill in a court form reliably?
Not reliably enough to file without a line-by-line check. A generative tool can produce different text from the same facts, miss a required field or change a date between documents. A rules-based workflow gives the same output from the same answers, which is why it suits official forms.
How many matters make a workflow tool worth building?
In the worked example, about 100 matters a year of the same document set, because building and maintaining the workflows costs far more up front than prompting. Below that, a prompt library with a strict checking routine is usually cheaper. Measure your own build and checking hours first.
Can I use both together?
Yes, and many firms should. Let the workflow collect the intake answers and assemble the structured documents, then pass the same answers to a generative tool to draft narrative sections such as a background statement or a client letter, with a person checking every fact against the intake data.
Further reads
- Best Legal AI for Family Law: What to Pay For and What to Skip — What to pay for in family law AI and what to skip.
- Family Law Retainer Agreements: AI Drafts, Lawyer Sign-Off — AI drafts of retainer agreements, with lawyer sign-off.
- Family Law Client Communication With AI: Calm, Consistent Emails — Calm, consistent client emails with AI help.
- How Small Law Firms Use AI to Draft Letters and Routine Documents — How small firms draft letters and routine documents.
- Law Firm Docketing With AI: Deadline Reminders Plus a Human Check — Deadline reminders with a human check.
- AI Contract Review for Small Firms: What It Catches and Misses — What AI contract review catches and misses.
- AI Software for Estate Planning Attorneys: Automation or Generative? — Rules-based document assembly against generative drafting for wills and trusts: how each fails, what each costs and the review step that suits each.
- Best AI for Family Law in Australia: Privacy, Court Rules and Cost — A family law shortlist with prices, what the Family Court's AI practice direction requires, and the privacy rules that keep client data out of free tiers.
- Gavel Pricing: Which Workflows Plan a Small Firm Needs — Gavel Workflows from $83 to Scale at $417 a month, matched to a small firm's seats, templates and integrations, with break-even sums and a 7-day trial plan.
- AI Legal Chronologies for Family Cases: Events Linked to Evidence — One dated table for a family case: extract events from messages, statements and orders, tie each row to a source and page, and keep a gaps list.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Gavel's pricing page, home page and rebrand post, Knackly's pricing page, and the ChatGPT Business and Claude Team pricing pages, read in November 2026; all hours and rates are illustrative assumptions.