AI Automation Agency for Law Firms: Questions Before You Sign

Before hiring an AI automation agency, a law firm should ask six things: what exactly will be built and measured, where client data goes, how privilege and confidentiality are protected, what every fee covers, who owns the finished workflows, and how you leave. Get each answer in writing before any money moves.

The reason to be strict is that an automation agency sells time and judgement, and neither is easy to inspect from outside. A law firm also carries duties a retailer does not: client confidentiality, supervision of anyone handling matters, and a record of how decisions were made. Each of those duties can be turned into a question you put on the first call and then confirm on paper.

Follow me on Instagram@sagnikteaches

What to settle before the first call

Agencies run better calls with prepared clients, and you ask sharper questions when you know your own problem. Spend an hour with the practice manager and one fee earner and write down three things: the jobs that eat the most unbilled time, the systems those jobs touch (practice management, document store, email, billing, intake forms), and the data that must never leave your control.

Connect on LinkedInSagnik Bhattacharya

Take a four-partner litigation and property practice as an illustration. The practice manager lists new-client intake, chasing documents, drafting routine letters, and month-end billing prep. Intake touches the website form, the conflict check and the case-management system. Billing prep touches time entries and the ledger. Under data, the list is short: anything inside a live matter file, and anything identifying a client. That one page changes the call from "what can you do for us?" to "can you do these three things, around these systems, without touching those files?"

Subscribe on YouTube@sagnikteaches

If you are still deciding whether an outside party is the right route at all, read what an AI consultant does for a law firm and what it costs first. It covers the advisory side; this tutorial is about the contract and the questions.

Scope: ask for the deliverable, not the activity

Vague scope is the commonest source of disappointment. "AI strategy", "automation roadmap" and "workflow optimisation" describe activity. You want nouns: a working intake flow, a draft-letter template library, a billing pre-check report. Put these questions to the agency.

1. What will exist on the day you finish that does not exist today? A good answer names flows, templates or integrations. A weak answer names meetings and reports.
2. Which systems will you connect, and have you connected them before? Ask for the exact product names. Joining a practice-management system to an email account is a different job from joining it to a legacy document store.
3. What is out of scope? A confident agency will list exclusions without being asked. If it cannot, ask it to write down the three things clients most often assume are included and are not.
4. What do you need from us, and by when? Delays on access, sample documents or a decision-maker's time are the real cause of most overruns. Get your obligations in the plan too.
5. How will we know it worked? Insist on a number agreed before the build: intake emails answered within one hour, say, or letters drafted in under ten minutes of fee-earner time.

A scope line that works might read like this. Treat it as an illustration of the level of detail to ask for, not as a template to sign.

Deliverable 3: New-enquiry intake flow
- Web form and email enquiries create a draft record in the case-management system (fields: name, contact, matter type, opposing party, urgency).
- A conflict-check request is prepared for a named person to run; the flow never marks a conflict as clear.
- An acknowledgement email is drafted and queued for human approval before sending.
- Out of scope: sending anything to a client without approval; reading or writing live matter files.
- Acceptance test: 20 sample enquiries supplied by the firm, 18 or more routed correctly, reviewed by the practice manager.
- Handover: written runbook, access list, and a 60-minute walkthrough with two staff.

Notice the acceptance test and the handover. Both are things you can verify, which is the point of the whole exercise.

Client data: where it goes and who can see it

Everything in this section is about the path a piece of client information travels. Ask the agency to draw it, on one page, for each automation. Start at the source system, and mark every service the data touches: the automation platform, the AI model provider, any storage, any logging or monitoring tool, and the agency's own staff.

1. Which AI model provider handles the text, and on which plan? Consumer plans and business or API plans treat data differently, and the terms change. Ask for the provider's current data-handling terms on the plan being used, then read the part on retention and on whether inputs are used for training.
2. Is any client data stored by the agency? Automation platforms keep run histories. Ask how long, who can open them, and whether you can set retention to the minimum.
3. Whose accounts does everything run on? The cleanest arrangement is that every subscription, API key and automation workspace is created in your firm's name, with the agency invited as a user. Then ending the relationship means removing a user rather than begging for an export.
4. Who at the agency will have access, and are they bound by confidentiality? Ask for names or roles, and whether any work is passed to a subcontractor or freelancer. Subcontracting is common and legitimate; hidden subcontracting is not.
5. Can the build use dummy data? Most of it can. Ask the agency to build and test on invented matters, and agree that live data comes only after sign-off.
6. What happens if there is an incident? You want a notification time stated in hours, a named contact, and a duty to give you the facts you need to meet your own reporting obligations.

If you have clients in the EU, add a question about whether the agency acts as a data processor and will sign a data-processing agreement. Your professional body or ethics adviser should confirm what your own rules require; an agency cannot answer that for you.

A worked data-path check

Say the agency proposes an automation that summarises incoming client emails and drafts replies. Ask it to fill in a table like this one. The values below are illustrative.

If the agency cannot fill in the third column, it has not thought about retention. If a row says "agency's own server", ask why, and whether it can move into your accounts instead.

Privilege and confidentiality: questions worth asking aloud

Whether sharing material with a vendor affects privilege is a legal question for your own advisers and professional rules. What you can do on the first call is test whether the agency understands why you are asking. Four questions do that well.

• "Does any step send the content of a communication between lawyer and client outside our accounts, and can we design the flow so that it does not?" Many useful automations only need metadata: sender, date, matter type, deadline.
• "Can the AI step work on a redacted version?" A routine such as 'classify this enquiry' rarely needs names. Stripping them before the model sees the text lowers risk at little cost in quality.
• "Where does a human review sit, and what can they see?" Anything that leaves the firm should pass a named reviewer first. Ask the agency to show the approval screen.
• "What do you do when a model is wrong?" Models can produce confident errors, including invented citations. Ask for the agency's standard checks: a mandatory review step, sample audits, and a rule that nothing citing authority goes out unverified. Our tutorial on which legal tasks a small firm should never hand to AI gives you a ready list of exclusions to put into the scope.

An agency that answers "that's covered by our terms" without engaging has told you something. An agency that asks which matters are the sensitive ones, and offers to build around them, has told you something better.

Fees: what each number buys

Agencies price in three common shapes: a fixed fee per build, a monthly retainer, or time and materials. Your job is to find what is inside each and what is billed on top. The questions are plain.

1. Is the first build a fixed fee, and what triggers a change charge? Ask for the change process in writing: who requests it, who quotes it, and a rule that nothing is billed without an agreed quote.
2. What does the retainer cover each month? Hours, monitoring, fixes, small changes, reporting? Ask what happens to unused hours.
3. Which software costs sit outside the agency's fee? Automation platforms, AI usage and practice-management add-ons are usually paid by you directly, and they grow with volume. Ask for a monthly estimate at your current volume and at double.
4. Is there a minimum term? Three, six and twelve months are all seen. A longer term should come with a lower price or a deliverable guarantee.
5. What is billed if we stop in the middle of a build? You want payment by milestone, so a partly finished project leaves you with finished parts and a fair invoice.

A quick sum to test a proposal

Take a hypothetical proposal: a fixed fee of $6,000 for an intake flow and a draft-letter library, then $1,200 a month retainer, plus about $150 a month of platform and AI usage you pay directly. All figures here are made up for the illustration. Over the first twelve months that is $6,000 + (12 × $1,200) + (12 × $150) = $22,200.

Now compare it with what the work is worth. If intake and letters currently cost the firm eight hours of fee-earner and staff time a week, and you value that time at a blended $60 an hour, the work is worth about $24,960 a year at current effort. Automating even two thirds of it saves around $16,600. So on those numbers, year one does not pay back, and the question becomes whether the retainer is needed after month three. That is exactly the conversation to have before signing, and it is why a retainer with no deliverable list deserves scrutiny. For wider context on tool pricing, see per-seat or pay-as-you-go legal AI pricing.

Ownership, handover and the way out

Most firms ask about price and forget the exit. Put this in the contract before you need it.

• Ownership of work product. Workflows, prompts, templates, documentation and any custom code built for you should belong to the firm or be licensed to it perpetually. Ask the agency to name anything it keeps as its own reusable component, and confirm in writing that you may keep using that component.
• Account control. As above: subscriptions in your name, agency as a user. This single choice makes every other exit term easier.
• Documentation standard. A runbook someone else could follow: what each automation does, its triggers, the accounts involved, how to switch it off, and who to call. Ask to see a sample from a previous job, with client details removed.
• Termination. A notice period you can live with (thirty days is a common ask), no penalty for ending a retainer after the first term, and a stated handover period.
• Data return and deletion. On exit the agency deletes anything it holds, confirms in writing, and returns exports you can read without its tools.
• Key-person risk. If the agency is two people and one leaves, what then? Ask who else knows your build, and whether a named backup exists.

A filled-in exit clause might read: "On termination by either party with 30 days' notice, the agency will within 14 days remove its access, hand over the runbook and exports, and delete any firm data it holds, confirming deletion in writing. The firm retains all workflows and templates created under this agreement." You can lift that structure and let your own lawyer polish it, which, as a law firm, you are well placed to do.

References, demos and the first-call script

Marketing pages cannot tell you whether an agency delivers. Three checks can.

1. Ask for a live demo on dummy data of something similar to what you want. A recorded demo shows the best day. A live one shows the rough edges.
2. Ask for two references from professional-services firms, and a call with each. Questions for the reference: did the project finish on time and price, what broke in the first month, and what did you have to take over yourselves? Don't rely on testimonial quotes.
3. Ask what they have turned down. A good agency can name a request it advised against. If everything is possible, nobody is exercising judgement.

Here is a compact script for the first call. Print it, and leave space for the answers.

1. What exactly will exist when you finish? (nouns, not activities)
2. Which of our systems will you connect, and which have you connected before?
3. What is out of scope?
4. Which AI provider, which plan, what retention?
5. Whose accounts does everything run on?
6. Who at your end sees our data? Any subcontractors?
7. Can you build and test on dummy matters?
8. Where is the human approval step? Show me the screen.
9. Fixed fee or retainer, and what does each include?
10. Monthly software costs we pay directly, now and at double volume?
11. Who owns the workflows, prompts and documentation?
12. What does leaving look like, and how long does it take?
13. Two references from professional firms, please.
14. What have you advised a client not to automate?

How the answers sound when they are wrong

Some replies should slow you down. Each one below is a typical warning sign, paired with a follow-up question that tests it.

A mistake that shows how scope slips

Picture a small family-law practice that hires an agency to "automate client communications". Three months in, the agency has built a flow that drafts updates from matter notes and sends them automatically on a schedule. Nobody wrote down that updates needed approval. One update goes out reporting a hearing date that had been moved the day before, because the note had not yet been updated. The client rings the partner, upset.

The fault is in the contract, not only the build. "Client communications" was a phrase, not a deliverable, and the approval step was an assumption. The fix is cheap: draft-and-queue instead of send, with a named reviewer, and a line in the scope that nothing leaves the firm unapproved. The lesson for your call is question eight on the script: ask where the human approval step sits and what the flow can send without one, and make the answer "nothing".

Checking whether the agency is actually needed

Some firms discover on the first call that their need is smaller than the agency's offer. Run a short test before signing: for each proposed automation, ask whether a feature in a tool you already pay for covers eighty per cent of it. Practice-management systems, document platforms and email suites keep adding drafting, summarising and reminder features. Our tutorial on ChatGPT or a legal AI tool for a small firm sets out the trade-offs. An honest agency will say when your own tools can do it and charge for the joining-up instead.

An agency is most useful in three situations: when a job crosses several systems and nobody on the team has the time to maintain the connections; when you need monitoring and error handling, not just a one-off build; and when the first project is the start of a programme and you want a single owner. If none applies, a part-time internal champion plus a well-chosen tool may be cheaper. The 90-day AI implementation plan for a small law firm is a decent yardstick: if the agency's plan is vaguer than that, ask them to sharpen it.

After you sign: keeping control for the first ninety days

Signing is the middle of the process. Three habits keep a good agency good.

• A fortnightly review of what was delivered against the deliverable list. Fifteen minutes, one page, with the acceptance test results.
• A named internal owner. One person at the firm, usually the practice manager, who answers the agency's questions the same day and holds the access list.
• A sample audit. Each week, a fee earner opens five outputs at random and marks them right, fixable or wrong. If wrong exceeds an agreed threshold, say one in ten, the flow pauses until it is fixed.

Say the audit in week two finds that of twenty draft letters, three cite the wrong matter type. That is fifteen per cent wrong, above the threshold, so the letter flow pauses and the agency fixes the classification step before it resumes. A written threshold turns an argument into a rule. If you want help choosing which jobs to automate first before you contact anyone, that is what my AI implementation consultation covers; you can reach me through the contact page.

Smaller questions firms ask before signing

Should a law firm pay an agency before a written scope exists?

No. Pay for a short, fixed-fee discovery piece if you must, but only against a written deliverable: a process map, a ranked list of candidate automations and a quote for the first one. If the agency wants a retainer before it has written down what it will build, the scope is probably being left vague on purpose.

Can an agency see client files during the build?

Only if the engagement letter says so, names who, and your own confidentiality duties allow it. Most builds can run on dummy or redacted matters. Make the default 'no live client data until a named partner approves in writing', and log every exception.

Who owns the automations after the project?

Agree this before signing. The safest wording gives your firm the workflows, prompts, templates and documentation, and lets you keep running them on accounts you control. Agencies often keep their own reusable components, which is fine if the licence to you is perpetual and written.

Is a monthly retainer or a fixed project fee better?

A fixed fee suits a defined first build with a clear finish line. A retainer suits ongoing monitoring and small changes once things are live. Be wary of a retainer that starts on day one with no deliverable list, because you can't tell what you're buying.

Do I need the agency at all?

Not always. If your needs are drafting help, summaries and intake emails, your practice-management or document tools may already cover them. An agency earns its fee on joining several systems together, or on work your team has no time to maintain.

Further reads

• How Small Law Firms Use AI to Draft Letters and Routine Documents — A routine-drafting job an agency may offer first.
• How to Summarise Bundles and Transcripts With AI Safely — Summarising bundles safely, a common agency request.
• How to Hire a HubSpot or CRM Consultant to Set Up AI Features — The same hiring questions applied to CRM work.
• AI Governance Checklist for Small Financial Advice Firms — A governance checklist from another regulated profession.
• AI Billing Software for Lawyers: Which Features Earn Their Cost — Which billing AI features are worth paying for.
• AI Automation Agency for Real Estate: Scope, Fees and Red Flags — Typical scope, retainer versus project fees, red flags and an in-house alternative for a real estate business weighing an AI automation agency.
• Claude or ChatGPT for a Law Firm: Which One to Pay For — Team seat costs, long-document limits, privacy defaults and Claude for Legal plugins compared, with a weighted three-task trial a small firm can run.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general contracting practice for professional services; no vendor figures used. Check confidentiality and privilege rules with your professional body or ethics adviser.