Before buying AI software, a small family law practice should check six things in writing: that client content is not used to train models, where the data is stored and who can see it, that the tool fits your professional-conduct and data-protection duties, that a lawyer checks every output, how children's information is handled, and how you leave.
The order matters. Family files hold the most sensitive material a firm ever sees: finances, allegations of abuse, children's medical and school records, addresses a client may be hiding. A tool that saves ten minutes but leaks one address has cost far more than it earned, so the confidentiality questions come before the feature demo, not after it.
Send every shortlisted vendor the same written questions, score the replies against eight sections, and test the survivors on dummy matters before a live file goes anywhere near them. A vendor that answers in a sales call but will not answer in writing has given you its answer. None of this is legal advice; where a rule depends on your regulator, the item tells you what to ask them.
Score each tool out of 40 before you see a demo
Create one row per tool and give each of the eight sections below a score from 0 to 5. A section scores 0 if the vendor will not answer in writing, 3 if the answer is acceptable but unverified, and 5 if the answer sits in the contract or a published trust page you have read. A total under 25 does not go to a pilot. Any 0 in sections one, two or five ends the conversation, whatever the total.
| Section | What you are scoring | Evidence that earns a 5 |
|---|---|---|
| 1. Training and retention | Is your content used to improve models; how long is it kept? | Contract clause: no training on customer content, stated retention period |
| 2. Data location and access | Where is it stored, which sub-processors, who at the vendor can read it? | Named hosting regions, sub-processor list, access-control statement |
| 3. Legal duties | Does it fit your conduct rules and data-protection law? | Data-processing agreement; a clause on breach notice |
| 4. Accuracy and citations | Can you see sources for every statement it makes? | Every output links to the document line it came from |
| 5. Children and third parties | How are minors' data and other parties' data handled? | Role-based access per matter; ability to exclude documents |
| 6. Fit with your workflow | Does it work with your case system, email and document store? | Native integration you have seen run on a dummy matter |
| 7. Cost and contract | Seat price, usage caps, minimum term, price rises | Written quote with every charge and the cancellation terms |
| 8. Exit | Can you export everything and have it deleted? | Export format named; deletion confirmed in writing within a set time |
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 nowConfidentiality: the questions that decide the purchase
Training and retention
Ask three things and get all three in writing. Is customer content used to train or improve models? How long are prompts, uploads and outputs kept, and can you set the period? Can your admin delete a matter's data on request? Business and enterprise plans from the major chat vendors generally state that business content is not used for training by default. Consumer plans leave it to a privacy setting instead: on ChatGPT's personal plans, for example, OpenAI's help centre says model training is on by default until you switch it off. That difference is the main reason a free personal account is the wrong place for client files. Check the exact wording on the vendor's current terms rather than relying on this paragraph; plans change.
A bad answer sounds like "we don't train on your data" with no contract clause behind it, or "data is anonymised" with no explanation of how. Anonymised by whom, and by what method? If the vendor cannot say, assume it is not.
Who can read the content
Ask whether vendor staff can open your matters for support or abuse monitoring, under what approval, and whether that access is logged. Some vendors keep content for a short period for abuse monitoring even when training is off; OpenAI's API documentation, for instance, gives up to 30 days by default. That can be acceptable, but you need the number of days so that you can describe it honestly to a client who asks.
A filled-in example: the confidentiality answers sheet
Send each vendor the same one-page questionnaire and keep the replies in the purchase file. A realistic answered set looks like this (illustrative):
Q3. Is customer content used to train or improve any model?
A: No. Clause 9.2 of the Master Terms. Content is processed only to provide the service.
Q4. Retention of prompts and outputs?
A: Configurable, 30-365 days; default 90. Admin can delete by matter.
Q5. Can vendor staff view content?
A: Only with admin approval for a support ticket; access is logged and expires in 7 days.Notice what makes this a good answer: it cites a clause, gives numbers and names who can change the setting. A vague reply such as "industry-standard security" scores a 1 at most.
Data location, sub-processors and your legal duties
Data location is a legal question before it is a technical one. If you have clients or counterparties in the EU, data-protection law such as the GDPR may restrict transfers outside the region, and you will want a data-processing agreement that names where processing happens. Ask the vendor for its list of sub-processors: the cloud host, the model provider behind the product, any speech-to-text service. A legal AI product is often a thin layer over a general model from another company, so your client's data may pass through two or three firms. Each should be named, with a hosting region.
Then look at your own side. Check your professional body's current guidance on technology and confidentiality, and your insurer's position on AI use: an insurer may ask whether you use generative tools and how outputs are checked. Write down what you found and the date. If the guidance is unclear for your situation, ask a qualified adviser or your regulator's practice helpline before the purchase, not after.
The wider supplier questions, including supervision and staff rules, are covered in the data, supervision and supplier checks for law firms; this tutorial stays with what is particular to family work.
Engagement letters and staff rules to settle first
Software is only half of the answer. Before the first live matter, decide two documents. The first is a short clause for your engagement letter that says you may use AI-assisted tools for drafting and summarising, that a lawyer reviews all output, and that the client can ask what is used. The second is a one-page staff rule. A usable version has five lines (illustrative):
1. Only the approved tool, only on a firm login. No personal accounts.
2. Children: initials and ages only in any prompt unless the tool scored 5 on sections 1 and 2.
3. No addresses, bank numbers or passport numbers in prompts.
4. A lawyer opens every cited source before anything leaves the building.
5. Anything odd goes in the correction log the same day.Staff who have no rule will invent one, and the invented rule is usually "whatever is quickest". Having the rule written also gives you something to show an insurer or regulator who asks how you control the tool.
What is special about family files
Children's information and third parties
A divorce file is not only about your client. It holds the other party's finances, children's school and medical details, and sometimes allegations about third parties. Ask the vendor how access is limited per matter, whether a document can be excluded from AI processing, and whether summaries can include names of minors. A prompt that follows your own rule looks like this:
Summarise the attached chronology for a first meeting. Refer to the children as Child A (age 9) and Child B (age 6) only. Do not state any address. List every date mentioned and flag any gap longer than 60 days.Safety-sensitive addresses
Some clients are hiding their location from the other party. An AI tool that drafts letters, fills in forms or sends reminders can expose an address in a footer, a calendar invite or a cc line. Test for it: create a dummy matter with a fake protected address and ask the tool to draft a client letter, a summary of a court form and a meeting reminder. Check every output for the address. This single test catches more problems than a feature demo does.
Distress and vulnerable clients
Intake chatbots and automated emails reach the client at an unknown time on an unknown device. Ask whether the vendor lets you turn off message previews, send nothing by email unless the client opted in, and keep a human in front of any message that mentions abuse. If a bot cannot recognise distress and hand off to a person within minutes, keep it away from first contact. The comparison in legal client intake software covers the intake side in more detail.
Citation checks: make every statement traceable
General language models can produce confident text that is wrong, and even specialist legal tools can cite something that does not support the point. The checklist answer is a routine, not a promise from the vendor. For every output that goes near a client or court, a lawyer opens the source and confirms three things: it exists, it says what the output says it says, and it is current. If you want the full routine, a small-firm checking routine for invented case law sets it out step by step.
When you assess a product, run the same test on all of them. Pick a point of practice you know well and ask each tool a question you can answer. Record whether it gives a source, whether the source opens, and whether it supports the statement. Do not rely on a vendor's accuracy claim; the claims in marketing are not independently tested on your kind of matter. Score section four on what you observed, not on what you were told.
A before and after: the unsupported summary
Before (illustrative output from a tool with no source links): "The parties agreed in 2021 that the family home would be sold." After you ask for a source line for each sentence: "The parties agreed in 2021 that the family home would be sold (email of 14 June, page 3, paragraph 2)." The second version can be checked in ten seconds; the first can only be checked by rereading the file, which defeats the purpose. Buy the tool that produces the second version by default.
A realistic mistake and how it showed up
Picture an associate who asks a tool to summarise a 40-page financial disclosure bundle and gets back a summary saying the husband's pension is "not disclosed". A reviewer who checks the source finds the pension statement on page 31, filed under a different heading, so the tool missed it. The summary looked tidy and the error was an omission, the hardest kind to spot. The defence is a checking step that does not trust the summary: the reviewer searches the bundle for the word "pension" before accepting any statement about what is absent. Put that step in your pilot as a planted test.
Fit, cost and exit
Integration with your case system
The tool must live where the work lives. Ask for a demo on a dummy matter inside your own case management system and email platform. Check whether notes and drafts file themselves to the right matter, because a tool that makes staff copy and paste between windows will be abandoned in a month, and copied text ends up in the wrong place. For billing-side features, the AI legal billing comparison lists what to look for in time capture.
Pricing traps
Ask about seat minimums, usage caps for document analysis, annual-only contracts and how far the price can rise at renewal. A small practice with four fee-earners and two support staff should work out the cost for all six seats and for four, because tools often price by user and only some staff will use them. The pricing models are discussed in per-seat or pay-as-you-go legal AI pricing. Add internal time: an hour a week of supervision per fee-earner is a cost even if the vendor does not charge for it.
Leaving the product
Before you sign, ask: what format will my data be exported in, how fast, and will you delete everything, including backups, and confirm that in writing? Also ask what happens to your data if the vendor is bought or closes. If the answers are loose, the tool scores 0 on section eight, and you should treat the purchase as a one-way door.
A worked example: a six-person practice scores two tools
Say a six-person family practice (two partners, two associates, a paralegal and an office manager) wants help with first-meeting chronologies, client letters and document summaries. They shortlist Tool A, a general business chatbot plan, and Tool B, a legal-specific drafting product. The office manager sends the questionnaire to both and the partners score the replies. All figures below are illustrative.
| Section | Tool A (general chatbot, business plan) | Tool B (legal drafting product) |
|---|---|---|
| 1. Training and retention | 5 (clause, retention set by admin) | 3 (statement on a web page, no clause yet) |
| 2. Data location and access | 4 (region choice, support access logged) | 3 (region named, no sub-processor list) |
| 3. Legal duties | 4 (data-processing agreement supplied) | 4 (agreement supplied) |
| 4. Accuracy and citations | 2 (no sources unless the user supplies documents) | 4 (source links on most outputs) |
| 5. Children and third parties | 2 (no per-matter access) | 4 (matter-level permissions) |
| 6. Workflow fit | 2 (copy and paste) | 4 (works with their case system) |
| 7. Cost and contract | 4 (monthly, cancel any time) | 3 (annual contract, price rise clause) |
| 8. Exit | 4 (export and delete tools) | 2 (export on request, 60 days) |
| Total | 27 | 27 |
Both reach 27, which shows why totals are only a first filter. Tool A is cheap and clean on data but weak on citations and matter permissions, and its 2 in section five is a warning that matters more than its total. Tool B suits the workflow but its training statement is not yet a contract clause and exit is slow. The partners agree to go ahead with Tool B only if the vendor puts the training clause and a 30-day deletion into the order form. Otherwise they use Tool A on anonymised text for drafting and accept the copy-and-paste cost. The decision comes from the scores, not from the demo.
Now the sum. If Tool B is $60 a seat a month for six seats on an annual term, that is $4,320 a year. If the pilot shows it saves each of the four fee-earners 90 minutes a week, and their time is worth $150 an hour, the gross figure is 4 x 1.5 x $150 x 46 working weeks, about $41,400. Subtract an hour a week each of checking, which is 4 x 1 x $150 x 46, about $27,600, and the net is roughly $13,800 before the $4,320 cost, so about $9,500 after it. That is still positive, but far less exciting than the gross figure, and it only counts if the freed time is billed or used on other work. Show the partners both numbers.
A four-week pilot that produces evidence
A pilot is how you learn whether the paper answers are true. Run it on dummy or fully anonymised matters until the contract is signed. The plan below fits a small practice.
- Week 0: choose three tasks (first-meeting chronology, client update letter, document summary) and pick three past matters, anonymised, as test material.
- Week 1: run each task with the tool and by hand; record minutes taken each way.
- Week 2: add the protected-address test and a child-data test; log any leak.
- Week 3: a second fee-earner repeats the tasks to check the results are not one person's flair.
- Week 4: count corrections, add up minutes, decide.
Keep a correction log with four columns: output, what was wrong, how it was caught, and how long the fix took. Plant ten known errors in the test material, such as a wrong date or a missing pension statement. If staff catch fewer than nine of the ten, your checking routine, not the tool, needs work first. The tasks themselves are covered in what a small firm should test in AI paralegal software.
Red flags that end the conversation
- The vendor will not give the data-processing agreement before payment.
- The product's answers on training differ between the sales call, the web page and the contract.
- No export route, or export only as a PDF.
- A promise that the tool "never makes mistakes" or "guarantees accurate citations". No tool can promise this.
- A request to upload real client files for a demo.
- Automatic client messaging with no way to hold a message for review.
Print the list and keep it next to the scoring table. A purchase you can explain to a client, a regulator and your insurer in two minutes is a purchase you can defend. If you would rather run this process with someone who has done it before, that is what my AI implementation consultation covers; you can get in touch here.
Questions family practices ask before signing
Can a family practice put client documents into a general chatbot?
Only if the plan's terms keep content out of model training, you control retention, and your regulator's guidance and your engagement letters allow it. Many practices start with business-tier plans and anonymised text. Ask your professional body for its current guidance before any real client file goes in.
Do I need client consent to use AI on their file?
It depends on your professional rules and what the tool does with the data. Many practices add a short clause to the engagement letter and a plain explanation on request. Check your bar or law society guidance, and ask a qualified adviser if the rules are unclear.
Is a free trial safe for a real matter?
Treat a trial like a purchase for data purposes. Use dummy matters until the contract, data location and deletion terms are confirmed in writing. Trial terms are sometimes different from paid terms, so read both.
How long should a pilot run?
Four to six weeks on three to five live-like matters is usually enough to see time saved and errors caught. Keep a log of every correction so the decision at the end rests on counted mistakes rather than impressions.
Further reads
- ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — General chatbot or legal-specific tool, side by side.
- Can Solicitors Use ChatGPT Without Breaching Confidentiality? — The confidentiality question for chat tools in detail.
- AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan once you have chosen a tool.
- CoCounsel Pricing: What a Small Firm Should Ask Before Buying — What to ask about a quote-only legal AI price.
- How Small Law Firms Use AI to Answer Client Status Questions — Answering client status questions without losing control.
- E-Discovery Software for Small Law Firms: Costs, Limits and AI Review — Costs and limits of AI document review.
- Conflict Check Software for Small Law Firms: Where AI Helps — Conflict checks, the other software every family practice needs.
- 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.
- Best Legal AI for Family Law: What to Pay For and What to Skip — General assistants, practice-management AI, research and drafting tools and form automation compared by job, with a budget and trial plan for a small practice.
- Family Law Document Automation: Gavel Workflows or Generative AI? — Rules-based assembly against prompting for family law documents: costs for a four-fee-earner firm, how each fails, and a hybrid that uses both.
- AI Receptionist for Family Law Firms: Handling Callers in Crisis — Safety triggers, handoff rules, scripts, recording consent and a monthly cost range for a family law firm's AI receptionist.
- AI for Family Law Attorneys: Summarise Financial Disclosure Safely — A step-by-step method for turning bank statements into a categorised schedule with AI, and the spreadsheet checks that prove the totals.
- Best AI for Family Law UK: Form E Prep and Client Updates — A safe route to Form E bank statement summaries and calmer client updates: what to redact, which tools to use, what to check and what the SRA has said.
- Family Law Retainer Agreements: AI Drafts, Lawyer Sign-Off — A clause-by-clause walk through an AI-drafted family law retainer: scope, fees, AI-use disclosure, and the checks a lawyer must still make before it goes out.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI business data page, ChatGPT data controls help article and OpenAI API data controls documentation; Anthropic's consumer terms update; the official data-protection page on international data transfers (all checked 5 November 2026). Request each supplier's trust page and contract terms, and check your own regulator's guidance.