Module 7.3 · Topic 2

Managing Risk for Protection

Bottom Line Up Front: Risk management for AI is systematic, not aspirational. Operationalize the Four Pillars—policies, training, quality control, and technology controls—into your daily workflow so every AI use is…

2.1 Comprehensive Risk Assessment Processes

Every AI-assisted task carries different risks. A risk assessment process is your first operational step: a checklist that forces you to think through what could go wrong before you deploy AI on a matter.

Identify the task and its stakes: Start with the specific task: "Research case law on privilege waiver." Then assess stakes: "This research will appear in a client memo and potentially in court filing." Stakes define risk level. Research for internal brainstorming is lower risk; research for court filing is high risk.
Categorize the risk by type: Ask: "What could go wrong with AI here?" Common risks include hallucinated citations (citing non-existent cases), outdated law (AI trained on data before recent decisions), confidentiality breach (using a consumer AI tool that retains client information), incorrect reasoning (AI misapplies a legal standard), or attribute error (AI applies the wrong jurisdiction's law). Use the NIST AI Risk Management Framework as your categorization guide: identify the harm (inaccurate advice, confidentiality loss), the likelihood (common issue with this AI tool?), and the consequence (client loss, sanctions, malpractice).
Map risk to ABA Model Rule 5.3 supervisory responsibility: Who controls this AI tool and who verifies its output? If a junior attorney is running the AI, your supervision obligation is to establish standards and review output. If a client is using an AI tool, you have a duty to advise them on risks. Map your supervisory role to the risk level.
Document your assessment: Create a simple checklist in your file: Task → Identified Risks → Control Type (policy, training, QC, tech) → Responsible Person → Verification Method. This three-minute documentation becomes your audit trail. If ever questioned about your AI use, this log proves you thought through risks methodically.
Scale your response to the risk: Low-risk tasks (internal research, brainstorming) need lightweight controls (lawyer spot-checks, no special tool restrictions). High-risk tasks (court filings, sensitive client communication) need heavyweight controls (mandatory peer review, enterprise-only tools, confidentiality training). Scale control to risk so you're neither over-cautious nor negligent.

2.2 Implementing the Four Pillars in Practice

The Four Pillars—policies, training, quality control, and technology controls—work as an integrated system. Each pillar supports the others; weakness in one pillar undermines the whole structure. Operationalize all four in your practice.

Pillar 1: Policies (Written Decisions About AI Use)

AI Use Policy: Document which tasks can use AI, which cannot, and the approval process for new tools. Example: "Research and legal writing can use AI-assisted tools. Client direct communication and testimony preparation cannot. Any new tool requires IT and ethics review before deployment." Written policy removes ambiguity and creates accountability.
Tool Approval Process: Establish criteria for approving new AI tools: Does it retain client data? Is it enterprise-licensed or consumer-grade? Does it support audit logging? Does it comply with confidentiality requirements? Create a three-question gate: "Confidentiality safe? Accurate on legal tasks? Auditable?" If all are yes, approve. If any is no, restrict or prohibit use.
Escalation Trigger: Identify what requires partner or managing attorney approval: "Any AI use on high-stakes litigation, regulatory matters, or matters where client has specifically requested no AI use requires written approval from the responsible partner."

Pillar 2: Training (Ensuring Everyone Understands Risks)

Annual Mandatory Training: Every lawyer and relevant staff member completes annual AI training covering your policies, common risks (hallucinations, privilege waiver, confidentiality), and the QC process. Track attendance. This covers your duty under ABA Model Rule 1.1 to ensure competence in your tools.
Targeted Tool Training: When you deploy a new AI tool (e.g., legal research AI), conduct a 30-minute hands-on training: "Here's how to use the tool, what it's good for, and what it hallucinates on. Here's the verification process. Here's when NOT to use it." Document who attended and when.
Incident-Driven Training: If an AI error is discovered (even internally), conduct a brief team debrief: "Here's what went wrong, why the AI failed, how we catch this next time." This reinforces vigilance and shares learning across the team.

Pillar 3: Quality Control (Catching Errors Before Delivery)

Verification Checkpoints: For high-risk AI output (legal research, drafted motions, contract terms), establish a verification step: paralegals or attorneys manually check every citation, cross-reference every legal conclusion against a trusted source, review every factual claim. The National Center for State Courts' guide to AI hallucinations recommends spot-checking: if AI is used to research 50 cases, manually verify at least 10 to calibrate reliability.
Escalation for Uncertainty: If a QC reviewer cannot independently verify AI output (e.g., AI summary of a complex regulatory analysis), escalate to a senior attorney or outside expert. Do not deliver unverified output to a client or court.
Audit Log: Maintain a simple log: "AI tool, task, output quality (verified/flagged/escalated), reviewer, date." Over time, this log shows which tools are reliable, which are risky, and where errors cluster. Use this data to refine your AI strategy.

Pillar 4: Technology Controls (Restricting Access and Retaining Audit Trail)

Data Handling Restrictions: Consumer AI tools (ChatGPT free tier, Google Gemini free) retain query data and may use it to train future models. Never paste client names, case details, confidential facts, or identifiable information into consumer tools. Use enterprise tools (Claude for Work, Microsoft Copilot Pro with enterprise agreement) or self-hosted models that don't retain data. Set this as a non-negotiable IT policy.
Audit Trail and Logging: If your AI tool supports logging (enterprise tools often do), enable it. Logging creates a verifiable record: "Who used the tool, when, for what task, and what the output was." This is invaluable for compliance audits and error investigations.
Access Restriction: Limit access to AI tools by role. Junior paralegals may have read-only access to research AI but cannot deploy it on new matter types without approval. Partners and QC reviewers have full access. This prevents unauthorized or inappropriate use.

2.3 Quality Control Systems for AI-Assisted Work

Hallucinations—confident, plausible-sounding but false outputs—are the signature risk of generative AI. Your quality control system must be ruthlessly systematic: verify every citation, cross-check reasoning, and treat AI output as a draft, not a deliverable.

Task Type & AI Risk Level	QC Method	Reviewer Role	Verification Standard
Low-Risk (Internal Brainstorming, Outlining)	Lawyer spot-check (10% sampling)	Attorney supervising the task	Plausibility review: does the structure and reasoning make sense? No external verification required.
Medium-Risk (Legal Research for Internal Memo)	Systematic verification of citations and holdings	Paralegal or junior attorney trained on verification process	Every citation manually checked against primary source (case database or court website). AI-summarized holdings must match the actual case language. Stanford research (Feb 2026) shows legal AI models hallucinate in approximately 1 out of 6 queries; assume error rate and verify accordingly.
High-Risk (Legal Research for Court Filing or Client Communication)	Mandatory peer review + spot audit	Senior attorney (minimum 3+ years experience with substantive area)	Independent research on top 5 controlling cases. Senior attorney traces AI reasoning to confirm no gaps or logical fallacies. Every statement of law cross-checked against current jurisdiction rules. NCSC guidance recommends independent re-research of key conclusions, not just spot-checking.
Critical Risk (Discovery, Contract Analysis, Due Diligence)	Multi-stage QC: AI output → paralegal first review → attorney peer review → managing attorney sign-off	Tiered: paralegal (rule-checking), attorney (legal sufficiency), partner (client-facing accountability)	AI output is treated as a preliminary draft only. Attorney conducts independent analysis of sample set (e.g., 10% of contracts reviewed by AI) to calibrate accuracy. Any deviation from AI's assessment triggers re-review of full output. Written certification of review before delivery to client or court.

Consequences of Inadequate QC

Hallucination Failure: Bloomberg Law reports that sanctions for AI hallucinations are becoming routine. Delivering AI-hallucinated citations to a court can result in sanctions, fee forfeiture, and referral to bar discipline. Inadequate QC is professional negligence.

2.4 Technology Controls and Access Management

Technology controls are your enforcement mechanism: they make it harder to violate policy and easier to comply. They also create an audit trail for verification and compliance review.

Confidentiality Safeguards: ABA Model Rule 1.6 (Confidentiality) requires you to protect client information. AI tools that retain or use data for training violate this rule. Select only tools with enterprise agreements that explicitly exclude data retention: your contract with the AI vendor should state, in writing, that client data is never retained, shared, or used for training. Document this contract in your files; reference it in your AI Use Policy. Make this non-negotiable.
Access Controls by Role: Set up your IT systems so different people have different access levels. Lawyers and paralegals can use approved research AI. Only the IT director can approve new tools. Only a designated "AI supervisor" can access the audit logs. This prevents unauthorized use and creates accountability. Access control is a technical implementation of your policy governance.
Logging and Audit Trail: Enable logging on every AI system that supports it. Logging should capture: who used the tool, when, for what document or task, what was input, what was output, who reviewed it. Over time, this audit trail becomes your evidence of compliance and your data source for process improvement. If a client ever claims "you used AI without telling me," your logs prove what tools were used, when, and by whom.
Usage Monitoring: Periodically review usage logs to spot anomalies: Is someone using the tool on a type of work outside their role? Is a single person generating unusually high volume of AI output, suggesting insufficient review? Is the same tool being used for different purposes in inconsistent ways? Flag these patterns and address them in team meetings or individual coaching.
Expiration and Offboarding: When an employee leaves, revoke their access to all AI tools immediately. When an AI vendor is replaced, migrate off their platform cleanly and delete all stored data. Document this process in your IT security procedures. Data retention beyond the employee's tenure is a confidentiality risk.