Claimstack Tech • Updated November 2026
📝 Legal AI • 9 min read

Writing a Generative AI Policy for Your Law Firm: A 12-Step Playbook

Your lawyers are already using AI tools, with or without permission. This 12-step playbook covers approved tools, data tiers, client consent, supervision, billing and how the rules differ by jurisdiction.

SPONSORED CONTEXTUAL PLACEMENT
💼
Evaluating Enterprise Software Deployments in 2026?

Request our confidential total cost of ownership matrix and vendor RFP templates.

Launch Calculator →

Five decisions your AI policy must make

An AI policy for law firms has one job: to turn vague professional duties into rules a second-year associate can follow at 11 p.m. before a filing deadline. It has to settle five things. Which tools are allowed? What data can go into each one? Who checks the output? What do clients get told? How is the work billed?

Everything else is supporting detail. Firms that write a 30-page treatise usually produce something nobody reads. Aim for a core policy of four to eight pages plus two appendices that change more often: the approved-tool list and the data classification matrix. That structure lets you update a vendor entry in a week without reopening the whole document for partner sign-off.

Advertisement

Assume you're starting late. In most firms, lawyers and staff are already pasting text into consumer chatbots on personal accounts. The policy's first practical effect should be to move that activity onto tools the firm controls. A blanket ban rarely achieves that. It pushes use out of sight.

RECOMMENDED SOLUTION

Looking to streamline vendor RFP selection?

Access pre-negotiated tier discounts through our independent institutional network.

Compare Pricing →

Steps 1 and 2: Name an owner and inventory what's already in use

Step 1: Appoint an accountable owner and a small working group. One named partner should own legal AI governance, supported by IT or security, the risk or general counsel function, a knowledge management lead and one fee earner who actually uses these tools. Five people is enough. Bigger committees stall on drafting.

Step 2: Inventory current use before you write a word. Run a short anonymous survey (which tools, for what tasks, on which accounts). Pull network and endpoint logs for traffic to the major AI services. Then check which existing vendors have switched on AI features in your document management, e-discovery, contract review or practice management platforms. Embedded features are often missed, and they can process client data by default.

The inventory tells you what to approve first. If half your litigators use a chatbot to summarize deposition transcripts, your policy needs a sanctioned route for that task on day one, or it'll be ignored.

Steps 3 and 4: Build the approved-tool list and vet each vendor

Step 3: Sort tools into tiers. Don't approve or ban "AI" as a single category. Approve specific products on specific plans, because the same vendor's consumer and enterprise tiers often carry very different retention and model-training terms.

Tool tierExamplesTypical 2026 cost (indicative, USD)Suggested data ceiling
Consumer chatbots on personal or free plansFree and individual plans of general chatbots$0–30 per user/monthPublic information only; no client work
Enterprise general assistantsMicrosoft 365 Copilot, ChatGPT Enterprise, Gemini for Google Workspace$20–70 per user/monthInternal and client-confidential, if contract terms bar training and set retention limits
Legal-specific platformsHarvey, Legora, CoCounsel, Lexis+ AI, Westlaw's AI features$100–500+ per user/month, often with seat minimumsClient-confidential; restricted matters case by case
AI features inside existing legal softwareDMS, e-discovery and contract-review add-onsOften bundled, or an uplift on the base licenceMatches the data the system already holds

These are typical 2026 market ranges. They vary widely by contract size, region and term length, so confirm with vendors. Pricing in India, Latin America or Southeast Asia may be structured differently or sold through resellers.

Step 4: Run a vendor risk assessment for each tool. At minimum, ask:

  • Are prompts and outputs used to train models? Get a contractual "no," not a settings toggle.
  • How long are inputs retained? Can the firm set zero or short retention?
  • Where is data hosted and processed? This matters under GDPR, India's DPDP Act, Brazil's LGPD, South Africa's POPIA and the UAE's data protection law.
  • Who are the sub-processors, and will you be notified of changes?
  • Does the vendor hold security certifications such as ISO 27001 or SOC 2 Type II?
  • Are there audit logs, single sign-on and admin controls?
  • What indemnities cover IP claims arising from outputs?

Red flag: a vendor that can't say in writing which underlying model providers receive your data.

Steps 5 and 6: Classify data and write the acceptable use rules

Step 5: Map data classes to tool tiers. Four classes work for most firms:

  1. Public: published judgments, legislation, marketing copy.
  2. Internal: firm know-how, templates, non-client admin.
  3. Client-confidential: matter documents, correspondence, client personal data.
  4. Restricted: material under protective or sealing orders, privileged strategy in high-stakes disputes, inside information on live deals, special-category personal data, export-controlled technical data, and anything a client's outside counsel guidelines exclude from AI processing.

Restricted data should need matter-partner approval even for top-tier tools. Some of it shouldn't go into any external system at all.

Step 6: Write the acceptable use rules as do's and don'ts. Keep them to one page. Spell out permitted tasks (first drafts, summaries, research leads, translation for internal understanding) and prohibited ones: personal accounts for firm work, uploading restricted data, sending AI output to a client or court without review, and producing evidence or expert content without disclosure. State plainly that the person who signs the work is responsible for it, whatever tool produced the first draft.

RECOMMENDED SOLUTION

Looking to streamline vendor RFP selection?

Access pre-negotiated tier discounts through our independent institutional network.

Compare Pricing →

Step 7: Settle client consent and disclosure

This is where jurisdictions diverge most. In the US, ABA Formal Opinion 512 (2024) says informed client consent is needed before confidential information goes into a self-learning tool that could expose it to others, and that boilerplate in an engagement letter isn't enough for that purpose. In England and Wales and much of Europe, there's no single consent rule. Confidentiality duties, data protection law and client expectations still push firms toward transparency.

A workable default for an international firm:

  • Add a plain-language clause to engagement terms explaining that the firm uses vetted AI tools under confidentiality controls, with lawyer review of all output.
  • Get specific, informed consent before using any tool that retains or trains on client data, or before processing restricted data.
  • Keep a client register of AI opt-outs and outside counsel guideline restrictions, and link it to matter intake so fee earners see it.

Corporate clients increasingly set their own rules, from outright bans to notice requirements. Your conflicts and intake team should flag these at matter opening, not after the work is done.

Steps 8 and 9: Supervision, verification and court filings

Step 8: Set verification standards by output type. Every cited case, statute and quotation must be checked against a primary source by a human. Courts in several countries have sanctioned or publicly criticized lawyers over fabricated citations, and senior judges in England and Wales have issued pointed warnings. Spot-check summaries of long documents against the source. Review drafts going to clients as you would a junior lawyer's work. Supervising partners remain responsible for the associates and staff who use these tools, as conduct rules in most jurisdictions already require.

Step 9: Track court and tribunal rules on AI disclosure. Some courts require parties to declare or limit generative AI use in filings. Examples include certain Canadian courts, the Supreme Court of New South Wales for affidavits and expert reports, the DIFC Courts in Dubai, and individual US judges through standing orders. Assign your litigation support or knowledge team to keep a list of the forums you appear in and the rules each sets. Refer to that list in the policy rather than copying the rules into it.

Step 10: Decide how AI-assisted work is billed

Billing is where clients will test your policy first. The defensible principles are simple. On hourly matters, bill actual time spent, including review, never the hours the work "would have taken." Don't bill clients for time spent learning a tool. Treat general AI subscriptions as overhead. If you pass through a matter-specific tool cost, disclose it in advance and charge actual cost.

The bigger decision is commercial. Hypothetically, say a 40-lawyer firm used to spend 30 hours on first-pass review for a mid-market share purchase and now spends 8. Billing hourly cuts that revenue by roughly three-quarters. A fixed fee for the diligence phase, priced on value and risk, keeps margin and gives the client cost certainty. For repeatable, high-volume work, most firms will do better with fixed or capped fees. Clients can see the efficiency gain either way, and they'll push for the saving.

Put the billing rules in the policy and the pricing strategy in a separate partner-level document.

Step 11: Train people and plan for incidents

Rules nobody has been trained on are just paperwork. Require a 60–90 minute onboarding session before anyone gets access to approved tools, covering data classes, verification duties and how to spot fabricated output. Follow it with a shorter annual refresher. Firms with EU operations should check the EU AI Act's AI literacy obligation for deployers, which has been in application since early 2025. Training records are your evidence of compliance.

Add an incident procedure. Anyone who puts restricted or client data into an unapproved tool, or finds an error in output already sent, reports it to the policy owner within 24 hours. The owner then assesses whether breach notification rules apply. Deadlines differ: GDPR generally requires notice to the regulator within 72 hours of becoming aware of a qualifying breach, and other regimes set their own timelines. Make reporting blame-free for first-time, self-reported mistakes, or people will hide them.

Finally, extend your risk assessment to high-risk uses. Where AI processing of personal data is likely to be high risk, run a data protection impact assessment under GDPR and comparable laws.

Step 12: Run an annual review, plus triggers for earlier updates

Review the core policy once a year, and update the appendices whenever a trigger occurs: a new tool request, a vendor changing its terms, new bar or court guidance where you practice, or an incident. Re-run the Step 2 inventory each year to catch new shadow use.

JurisdictionMain sources to monitorPolicy emphasis
United StatesABA Formal Opinion 512; state bar opinions; judges' standing ordersInformed consent for confidential inputs, fees, citation checks
England and WalesSRA, Law Society and Bar Council guidance; judicial AI guidanceCompetence, confidentiality, verification of authorities
European UnionEU AI Act, GDPR, national bar and CCBE guidanceAI literacy, DPIAs, data transfers
Canada and AustraliaLaw society guidance; court practice notesDisclosure and limits in court filings
Singapore and the GulfMinistry of Law and court guidance; DIFC Courts guidanceResponsible use in proceedings, verification
India, Latin America, AfricaGeneral conduct rules; DPDP Act, LGPD, POPIAConfidentiality and data protection, pending specific bar rules

Guidance in every one of these jurisdictions keeps moving. Treat the table as a monitoring list, not a statement of current law, and have conclusions confirmed by regulatory counsel wherever you practice.

Here's a decision rule for this quarter. If your firm has no approved enterprise tool yet, do Steps 1 to 4 first and get one sanctioned tool live within 60 days. A short policy backed by a usable tool will cut more risk than a polished document that leaves lawyers on personal accounts.

Common Questions

How long does it take to write and roll out a law firm AI policy? +
Do we need client consent every time a lawyer uses AI on a matter? +
Is it safer to simply ban generative AI at our firm? +
CT
Claimstack Tech Editorial Team
Independent research desk
Guides on Claimstack Tech are produced by our editorial team. Pricing, features and regulations change often - confirm details with the vendor or a licensed professional before making decisions.
SPONSORED 2026 Vendor Pricing Benchmarks: Model exact seat pricing, setup fees, and ROI across our verified database.
Launch Calculator →