Resource guide
Legal AI, without the hype
A practical guide for independent law firms in England and Wales.
The working rule: start with one repeated workflow. Keep client data out until the tool, purpose and controls are approved. Verify every output before anyone relies on it.
Use this guide to: choose a sensible use case, set data boundaries, test outputs and decide whether to fix the process, pilot AI or leave the work alone.
You will not need: client names, matter details, contract text, privileged material, a software subscription or a sales call.
Operational guidance, not legal advice. Do not enter confidential information into an unapproved tool. Download the printable guide (PDF).
1. The short answer
AI can be useful legal technology. It does not remove professional responsibility, turn uncertain material into fact or repair a broken workflow by itself.
| Good at bounded assistance | Weak at owning the answer |
|---|---|
| Summarising supplied material, extracting fields, comparing text, classifying items, drafting from approved instructions and reorganising information. | Determining truth, resolving legal uncertainty, understanding unstated context, judging proportionality or deciding what a client should do. |
| Value appears when | Risk grows when |
|---|---|
| The inputs are controlled, the task repeats, the expected output is defined, a reviewer is named and errors can be detected before release. | People paste sensitive material into unapproved tools, accept fluent output without checking it or scale a process before its failure modes are understood. |
A useful first question: what repeated piece of work takes enough time to measure, has a clear finish line and can be checked by a qualified person?
Accountability stays with the firm. The SRA's Risk Outlook on AI in the legal market states that "you will remain responsible and accountable for the outputs from AI you are using" - human review, scrutiny and professional judgement remain necessary (source: sra.org.uk
2. Start with the workflow
Do not begin with a vendor demonstration. Describe one current process well enough that the firm can recognise improvement or harm. (The Contract-Workflow Baseline Kit, /resources/contract-workflow-baseline-kit/, is the open workbook for exactly this.)
| Gate | Pass when | Stop or fix first when |
|---|---|---|
| Repeated | The work occurs often enough to measure | It is rare, unique or selected only because it was memorable |
| Bounded | The trigger, finish line and expected output are clear | The task quietly expands into judgement or advice |
| Reviewable | A qualified reviewer can detect material errors | There is no reliable comparison or reviewer |
| Ownable | One named person can change the process | Responsibility is shared but ownership is absent |
| Measurable | Time, turnaround, rework and quality can be compared | The only success measure is that a demo looked fast |
Worked example: first-pass review of supplier paper.
- Trigger: a customer contract arrives through the agreed intake channel.
- Task: identify departures from the approved playbook and prepare a first-pass issue list.
- Finish line: a named lawyer receives a traceable issue list linked to the source clauses.
- Human control: the lawyer verifies every issue, decides materiality and owns the advice.
- Measures: touch time, elapsed time, missed issues, false positives, rework and partner input.
Process before product: if the team has no playbook, inconsistent intake or unclear approval ownership, fix that first. AI will reproduce ambiguity faster.
3. Decide what data can enter
A tool being easy to access does not make it approved for client or personal data. Set the rule before anyone prompts.
| Status | Typical material | Working rule |
|---|---|---|
| GREEN | Public, fictional or synthetic examples with no client or personal data | Suitable for early capability testing, subject to the tool's approved-use policy |
| AMBER | Internal know-how or non-public business material | Use only after the purpose, access, retention, supplier terms and security controls are approved |
| RED | Client-identifying, personal, privileged, confidential, special-category or live matter material | Do not enter into a public or unapproved tool |
Before processing personal data:
- Define the purpose and lawful basis before processing begins.
- Identify who is controller, processor and sub-processor.
- Apply data minimisation: use only what the task genuinely requires.
- Decide whether a data protection impact assessment is required and record the decision.
- Explain material uses of personal data and AI where transparency obligations apply.
- Set retention, deletion, access, incident and individual-rights procedures.
Do not rely on anonymisation by instinct. Removing names may not remove identity, privilege or commercial sensitivity. Treat de-identification as a control that needs its own review.
Safer early test: create a fictional contract excerpt that contains the same clause patterns and formatting problems as the real work. Keep the expected answer separately, then compare the tool's output against it.
4. Prompt for a bounded task
A prompt is an instruction, not a control system. Make the task narrow, state the limits and require output that a reviewer can check.
| Part | What to specify |
|---|---|
| Context | The role of the material and the intended user of the output |
| Limit | The supplied text only; no invented law, facts, clauses or assumptions |
| Expected task | One verb: extract, compare, classify, summarise or draft |
| Answer format | A table, issue list or marked comparison with source references |
| Review trigger | State uncertainty, missing information and every point requiring human judgement |
FICTIONAL EXAMPLE - NO CLIENT DATA. Prompt: "Using only the fictional clause below, list each departure from the supplied playbook. Quote the exact clause words, name the playbook rule and mark uncertainty. Do not give legal advice or invent missing terms. Return a three-column table: clause, issue, reviewer check."
Then challenge the result:
- Which conclusion is not supported by an exact source quote?
- Which playbook rule did you apply to each issue?
- What did you omit because the evidence was missing?
- What would change the answer?
Prompt quality cannot cure unsuitable use. A well-written prompt does not make an unapproved tool safe for confidential data, and it does not prove the output is correct.
5. Verify and supervise the output
Fluent output can still be false. Build review around the source, the expected standard and the consequences of error.
| Check | Reviewer action | Evidence to retain |
|---|---|---|
| Source fidelity | Trace every factual statement and quotation to supplied material | Source reference and quoted span |
| Legal accuracy | Verify every authority, citation, date and proposition independently | Authoritative source and access date |
| Completeness | Compare against an approved checklist or expected-answer set | Missed and extra items |
| Bias and consistency | Test comparable inputs and material edge cases | Test set and observed differences |
| Confidentiality | Check inputs, outputs, logs, access and unintended disclosure | Data record and incident path |
| Professional judgement | Name the authorised person who approves use of the output | Approval, corrections and final version |
Stop conditions:
- A fabricated source, authority, quotation or clause.
- A material issue that the reviewer cannot trace to source evidence.
- Sensitive information entered outside the approved boundary.
- Error rates or review time worse than the current process.
- The output cannot be supervised at the speed the system produces it.
Do not ask the model to grade itself. Self-critique can help surface questions, but it is not independent verification. A qualified person remains responsible for the final work.
6. Question the supplier
Do not buy a confident demonstration. Ask for the operating facts needed to approve the real use case. (The standing version of this question set, kept current, is at /insights/legal-ai-vendor-data-security-questions/.)
| Area | Questions to answer before approval |
|---|---|
| Purpose | What exact task is the system designed to perform, and what is outside scope? |
| Data use | Are prompts, uploads or outputs used for training, evaluation or product improvement? Can this be disabled contractually? |
| Storage | Where is data stored, for how long, and how is deletion verified? |
| Access | Who can access content and logs? What role, authentication and audit controls exist? |
| Processors | Which sub-processors are involved, where do they operate and how are changes notified? |
| Security | What assurance, encryption, vulnerability, incident and recovery evidence is available? |
| Output | How are sources exposed? What known limitations and tested failure modes apply? |
| Change | How are model, feature and policy changes communicated and revalidated? |
| Exit | Can the firm export its data, logs and configurations and confirm deletion? |
| Responsibility | What indemnity, liability, insurance and support terms apply to the intended use? |
Evidence, not a questionnaire alone: retain the contract terms, data-flow diagram, processor list, security evidence, approved use case and named internal owner together.
7. Run a bounded pilot
A pilot should answer a decision, not merely prove that the software can produce output.
- Freeze the current baseline before using the tool.
- Choose a representative test set, including ordinary work and known edge cases.
- Write the expected output and material-error definition before viewing results.
- Run the current and proposed processes on the same safe test set.
- Have a qualified reviewer record corrections, uncertainty and review time.
- Apply the pre-agreed stop conditions and write the decision.
| Measure | Baseline | Pilot | What changed? |
|---|---|---|---|
| Hands-on lawyer time | |||
| Elapsed turnaround | |||
| Material issues missed | |||
| False positives | |||
| Correction or rework cycles | |||
| Partner or senior input | |||
| Write-off or fixed-fee overrun |
The review burden counts. A system that drafts in seconds but adds twenty minutes of checking may not improve the workflow. Measure the completed, supervised result.
8. Write the decision
A valid outcome may be to fix the process, run a narrower test, buy a tool or leave the workflow alone. (The one-page decision sheet is at /resources/fix-buy-pilot-leave-it-alone-guide/.)
| Decision test | Answer and evidence |
|---|---|
| What problem were we trying to reduce? | |
| Did completed quality improve? | |
| Did total review time fall? | |
| Were confidentiality and data controls satisfied? | |
| Can the output be supervised at expected volume? | |
| What new failure modes appeared? | |
| Who owns the process after the pilot? | |
| What is the next review date? |
- FIX THE PROCESS: use when intake, playbooks, ownership or approval are the real constraint.
- TEST AI: use when the task is bounded, measurable, reviewable and safe to test. Buy only after the evidence supports it.
- LEAVE IT ALONE: use when recurrence, value, evidence or supervision cannot justify change.
Decision: ___ Owner: ___ Review date: ___
9. Official sources and next step
This guide is a practical starting point. Recheck current regulatory, professional, data-protection and security guidance for the intended use.
- Solicitors Regulation Authority - AI Risk Outlook: sra.org.uk
- Solicitors Regulation Authority - Effective supervision: sra.org.uk
- The Law Society - Generative AI, the essentials: lawsociety.org.uk
- The Law Society - How AI tools hallucinate and why it matters in law: lawsociety.org.uk
- Information Commissioner's Office - Guidance on AI and data protection: ico.org.uk
- National Cyber Security Centre - Guidelines for secure AI system development: ncsc.gov.uk
Important limitation: this resource is general operational guidance, not legal, regulatory, information-security or data-protection advice. Requirements depend on the firm, tool, data, client terms, jurisdiction and intended use. ICO AI guidance is currently under review and should be rechecked before reliance.
Start with one workflow: use the open Contract-Workflow Baseline Kit (/resources/contract-workflow-baseline-kit/) to record the current process before changing it, then the Readiness Score at /readiness/. If you want the measurement and decision facilitated, the Workflow Value Workshop is GBP 2,500 fixed - one workflow, one half-day, one written decision - preceded by a 20-minute fit call (/fit-call/); both described at /consultancy/.
MARGO LEGAL LTD, company number 17322603. Registered in England and Wales. Registered office: 66 Paul Street, London, England, EC2A 4NA. Originally published as a PDF (version 1.0, 18 August 2026); converted to HTML 15 September 2026.