Mastery: build a controlled legal workflow

This lesson is educational and is not legal advice. Use qualified legal professionals and applicable organizational controls for real legal matters.

A mature workflow does not begin with a prompt. It begins with a permitted purpose and ends with an accountable decision. Use the following checklist as a design and review tool.

1. Define purpose and limits

• State the narrow job in observable terms.
• Name the user, affected parties, and intended use.
• Separate extraction, synthesis, drafting, recommendation, and decision.
• List decisions the model must not make.
• Define what success and a severe failure look like.
• Confirm that AI provides a real benefit over a simpler method.

Weak: “Review this matter.” Stronger: “Using the supplied agreement and playbook, extract renewal terms, quote each source location, and flag deviations for counsel; do not assess enforceability or send notices.”

2. Establish authority and scope

• Identify jurisdiction, forum, governing documents, dates, and matter boundaries.
• Confirm who may request, review, approve, and act.
• Record the applicable playbook, policy, or review standard and version.
• State escalation paths for unclear authority.
• Treat predictions and recommendations as analysis, not fact.

If nobody can identify the accountable decision-maker, the workflow is not ready.

3. Protect information

• Classify source data before use.
• Confirm the exact system and configuration are approved for that classification and purpose.
• Evaluate privilege, confidentiality, privacy, contractual restrictions, court orders, and legal holds through authorized processes.
• Minimize inputs and inspect redactions, metadata, comments, filenames, and indirect identifiers.
• Limit access, sharing, retention, export, and downstream reuse.
• Apply appropriate protections to prompts, outputs, logs, and evaluation examples.

Never use an unapproved tool to determine whether content is permitted in that tool.

4. Prepare trustworthy sources

• Inventory every document, exhibit, amendment, attachment, and version.
• Check authenticity, completeness, legibility, dates, and scope.
• Preserve section numbers, defined terms, redlines, tables, and stable locations.
• Use a source hierarchy appropriate to the task.
• Record retrieval date and research coverage.
• Mark unavailable or partial sources visibly.

The model should not repair missing text or complete a citation by guessing.

5. Bound the prompt

• Name the allowed source set.
• Delimit source text as data.
• Require quotations and stable locations for material findings.
• State invariants, permitted transformations, and prohibited changes.
• Require missing information to remain [OPEN] or [UNVERIFIED].
• Ask for contrary evidence, exceptions, dependencies, and limitations.
• Specify a structured output that supports review.

Prompts are controls only when the surrounding system enforces access, evidence, and approval.

6. Verify the output

• Split compound statements into checkable claims.
• Retrieve every citation independently.
• Check quotation, speaker, context, proposition, authority, treatment, and factual fit.
• Compare clauses against complete agreements, definitions, and cross-references.
• Confirm names, dates, numbers, thresholds, actors, triggers, conditions, exceptions, and remedies.
• Distinguish source fact, model inference, legal analysis, and authorized decision.
• Re-run checks after substantive edits or document assembly.

Do not ask the same model to certify its own answer. Model-assisted checking may help locate issues, but accountable people verify against authoritative sources.

7. Apply human gates

• Match reviewer qualifications and authority to consequence and uncertainty.
• Give reviewers sources, mappings, limitations, and task-specific criteria.
• Allow rejection and escalation, not only edit and approve.
• Prevent throughput from overwhelming review capacity.
• Record reviewer, evidence, decision, exception, and next action.
• Place a final gate before sending, signing, filing, disclosing, or relying.

For high-impact work, consider independent source verification and separate approval of the legal position.

8. Monitor and improve

• Test representative cases, edge cases, and deliberate stop cases.
• Track severe errors separately from average quality.
• Measure unsupported claims, missed clauses, altered terms, privacy events, reviewer corrections, disagreement, and post-approval failures.
• Review logs without unnecessarily reproducing protected content.
• Revalidate after tool, model, policy, law, playbook, or workflow changes.
• Retire the workflow if controls cannot keep pace with risk.

Do not promote one accepted exception into a general standard without formal playbook approval.

For each run, create a proportionate evidence packet:

1. request, purpose, and prohibited uses;
2. approved tool and data classification;
3. source inventory and versions;
4. prompt or workflow version;
5. model output with citations or clause locations;
6. verification record and open items;
7. reviewer identity, authority, and decision;
8. final artifact and action approval;
9. retention and deletion instructions.

This packet should make the work reconstructable without becoming a second uncontrolled repository of sensitive material. Use identifiers or controlled links when copying full documents would create unnecessary exposure.

Create an AI-assisted review plan for a fictional vendor agreement. The agreement contains a liability cap, security exhibit, confidentiality terms, and renewal clause. One exhibit is missing, one case citation suggested during research cannot be found, and the draft contains personal data that is unnecessary for clause comparison.

A mastery submission should:

• decline or pause until the missing exhibit is handled;
• remove unnecessary personal data before approved processing;
• compare text only against the named playbook version;
• label the candidate citation unverified and exclude it from conclusions;
• preserve definitions and cross-references;
• map every finding to source language;
• route legal positions to qualified counsel and operational duties to business owners;
• require final approval before negotiation language is sent;
• record severe failures and use them to improve the workflow.

The objective is not maximum automation. It is a process that makes unsupported confidence, hidden data movement, and unauthorized decisions difficult.