Resource guide
Context-aware AI for lawyers
A boundary guide for source material, instructions, retrieval and review. Use it when an AI output depends on matter context, internal material or retrieved sources.
Boundary: this guide does not approve an AI product or permit client data to be used. Professional duties and firm controls continue to apply.
Operational guidance, not legal advice. No confidential information is needed. Original PDF editions (all v1, 18 Aug 2026, now combined into this single maintained page): Context-Aware AI for Lawyers, Five AI Workflows for Legal Work, How AI Fits Into Legal Work, Understanding AI in Legal Practice and AI for Lawyers Foundations Guide.
1. Frame the decision
Start with one task, a named owner and a representative test set. Speed is not evidence that legal work is safe or useful.
Specific focus - pick the line that matches your situation:
- Context and boundaries: record the context, permission and source checks rather than relying on fluent output (when an AI output depends on matter context, internal material or retrieved sources).
- Workflow testing: compare five bounded workflow types - summarising, extraction, comparison, drafting support and triage - and choose one for evidence-led testing (when the firm wants concrete AI use cases without treating every task as suitable).
- Plain-language adoption: start with repeated tasks and explicit review rather than an organisation-wide claim (when partners need a practical explanation of where AI may and may not belong).
- Shared vocabulary: distinguish generation, retrieval, classification and automation without replacing judgment (when people need shared vocabulary before discussing adoption).
- Foundations: use an approved example to teach limits, source checking and escalation (when lawyers need an introduction before trying an AI workflow).
Questions to answer:
- What exact task is in scope?
- What material may enter the tool?
- Which errors matter?
- Who checks output against sources?
Avoid:
- starting with client data
- counting output rather than corrected outcome
- letting a model make the final legal decision
2. Control sheet
Treat a blank evidence field as a reason to pause, not a reason to guess. Record sources and owners before changing the workflow.
| Area | Evidence to collect | Pass condition |
|---|---|---|
| Scope | Task, users and exclusions | Boundary is written |
| Data | Inputs, retention and access | Route is approved |
| Test | Examples and error labels | Threshold is frozen |
| Review | Reviewer and source checks | Every output is scrutinised |
| Stop | Error and incident triggers | Named owner can halt use |
Fail closed: if an input, owner or approval is missing, record the gap and keep the affected step outside the proposed change.
3. Working record
Complete this page with counts, dates and named sources. Do not include client names, matter names, contract text or privileged material.
| Field | Evidence or answer |
|---|---|
| Task | |
| Tool and version | |
| Approved test data | |
| Baseline | |
| Pass and stop rule | |
| Reviewer and date |
Evidence quality check:
- The source is named and dated.
- An estimate is labelled as an estimate.
- The owner can reproduce the measure.
- The same method can be used after a change.
4. Decision record
Select the outcome that the evidence supports. A stop or no-change decision is valid.
- Stop: a required source, owner or approval is missing.
- Fix: repair the underlying route before adding technology or cost.
- Test: run one bounded change with a baseline and review date.
- Proceed: adopt only the option supported by the recorded evidence.
| Decision field | Record |
|---|---|
| Chosen outcome | |
| Reason and evidence | |
| Owner and approvers | |
| Review or expiry date |
Related reading on this site: what human review actually means (/insights/what-human-review-means-for-legal-ai/) and the failure modes of legal AI research (/insights/legal-ai-research-failure-modes/).
Primary sources
- SRA - Compliance tips for solicitors on AI and technology: sra.org.uk
- SRA - Effective supervision: sra.org.uk
- SRA - Code of Conduct for Firms (confidentiality of client information, para 6.3): sra.org.uk
- ICO - Guidance on AI and data protection: ico.org.uk
- Law Society - How AI tools hallucinate and why it matters in law: lawsociety.org.uk
Source status: links checked 15 September 2026 (SRA x3 and ICO verified live 200; the Law Society page blocks automated fetch with 403 but is indexed - confirm by hand before relying). Confirm the current version and obtain appropriate professional advice before relying on regulated, contractual, privacy or security conclusions.
MARGO LEGAL LTD, company number 17322603. Registered in England and Wales. Registered office: 66 Paul Street, London, England, EC2A 4NA. Originally published as five PDF guides (v1, 18 August 2026); consolidated and converted to HTML 15 September 2026.