Stepping Up to the Plate: The Importance of an AI Policy Checklist
November 05, 2026
In our previous post in this series, we discussed why having humans in the loop remains central to defensible AI usage and identified best practices for ensuring proper a defensible approach to leveraging AI in discovery. That principle is increasingly finding its way into formal requirements, including California’s SB 574, which reinforces the need for attorneys to maintain human oversight, verify AI-generated content, protect confidential information, and meet applicable disclosure obligations.
A practical AI policy checklist helps turn broad commitments – such as commitments to protect confidentiality, supervise technology, and communicate with clients – into decisions people can make consistently during everyday work. For law firms and corporate legal departments, an AI policy checklist’s greatest value is requiring users to establish what is permitted, who is responsible, what must be verified, and when to stop and seek guidance. In this post, we’ll discuss why a practical AI policy checklist matters and considerations for ensuring the checklist addresses the needs of the organisation over time.
Why a Practical AI Policy Checklist Matters
An internal policy might state that employees must use AI responsibly. But that statement gives a lawyer little help when deciding whether to upload deposition testimony, activate a meeting assistant, use AI to identify privileged documents, or charge a client for AI-assisted research.
A checklist brings those decisions into the workflow before information is uploaded, results are relied upon, or an invoice is issued. It also creates consistency across attorneys, litigation support personnel, contract reviewers, and outside providers. Without that consistency, the same organisation can apply very different standards to substantially similar work.
The checklist should produce an identifiable result: permission to proceed, permission subject to conditions, or a decision to pause. Each completed item should create a record of what was reviewed and why the requirement was satisfied. For example, “security reviewed” should identify the approved configuration and review date; “client consent obtained” should identify the authorized representative and the scope of consent.
This structure also makes responsible adoption easier. Staff can use previously approved workflows without repeatedly seeking permission, while unusual or consequential uses receive additional review. A useful implementation separates organisation-wide checks, such as vender assessment, from matter-specific checks, such as protective-order restrictions.
Client Consent Requires a Defined Decision Process
The policy should distinguish notifying a client from obtaining informed consent. ABA Formal Opinion 512 treats confidentiality-based consent and client communication as separate inquiries. It requires informed consent before entering representation information into certain self-learning systems presenting disclosure risks and explains that generic engagement-letter language is insufficient when informed consent is necessary. Client communication may also be required because of client instructions, questions, fees, or significant decisions influenced by AI.
Operationally, the checklist should require the responsible lawyer to identify the intended use, information involved, material risks, safeguards, and alternatives before deciding whether consent is needed. It should also establish who can consent on behalf of an organizational client.
Permission should have a defined scope. Consent to use one approved system for document summarization should not automatically authorise uploading the same documents to another provider or enabling an agent to communicate externally.
Data Security and Confidentiality Need Separate Attention
Security asks whether information is adequately protected. Confidentiality asks whether the organisation is authorized to use or disclose it in the proposed way. A technically secure platform can still be inappropriate for information subject to a restrictive protective order or client instruction.
The confidentiality duty under ABA Model Rule 1.6 concerns information relating to representation, extending beyond documents that qualify for attorney-client privilege. The rule also requires reasonable efforts to prevent unauthorized access or disclosure. A checklist limited to “privileged documents” would therefore miss a substantial category of protected information.
It’s important to check the entire information path: prompts, uploaded documents, connected repositories, retrieved material, outputs, logs, and shared links. Removing a client’s name should prompt a further question: could the remaining facts still identify the client or reveal sensitive information?
Similarly, a provider’s promise not to train on customer data answers only one question. Review should also cover retention, support access, subprocessors, sharing permissions, deletion, and the organisation’s actual settings.
Approved Tools Should Mean Approved Uses and Configurations
An approved-tool list works best when approval specifies the product, feature, account type, permitted data, and authorized tasks. Otherwise, staff may reasonably assume that approval of a platform includes every newly released feature.
For example, an organisation might approve an application for summarising public information but require additional review before connecting it to client files. Approval for drafting should also be distinguishable from approval for taking actions.
The state bar of California’s updated 2026 guidance expressly addresses agentic AI and emphasizes that greater autonomy requires appropriate supervision and verification. It also addresses the risks of persistent access to internal systems and autonomous external transmission of client information.
A practical policy should therefore identify which actions require human authorization, including sending substantive advice, filing documents, producing discovery, changing records, and transferring information outside an approved environment.
Billing Guidance Prevents Inconsistent and Misleading Charges
Billing instructions should be concrete enough for both timekeepers and invoice reviewers to apply. Florida Bar Opinion 24-1 rejects falsely inflated billable hours and duplicate charges, including costs already recovered through overhead. It also addresses the distinction between general education and work performed for a particular client.
For hourly work, the checklist should confirm that entries reflect actual time spent. If an attorney spends 15 minutes preparing instructions and 45 minutes checking and revising the result, the recorded work is one hour – not the four hours the task might previously have required.
Alternative fees require a separate assessment. Under ABA Model Rule 1.5, reasonableness remains relevant, and the basis of fees and expenses must be communicated as the rule requires. A fixed fee should not simply conceal hypothetical hours or unexplained technology charges.
Corporate legal departments can apply the same checklist when reviewing outside counsel guidelines, budgets, invoices, and proposed AI surcharges.
Disclosure Needs Its Own Checklist Entry
“Disclose AI use when required” leaves several decisions unresolved: disclose to whom, disclose what, by when, and under which authority?
Client disclosure, court disclosure, and disclosure to opposing counsel should be assessed separately. California’s guidance specifically directs lawyers to check applicable rules, orders, and other jurisdictional requirements concerning AI disclosure.
For eDiscovery, include the governing discovery agreement and any representations already made about review methods. The matter team should document its disclosure determination and avoid assumptions that one recipient’s requirements answer every other disclosure question.
The policy should also require factual accuracy in certifications. A statement that authorities were independently verified should describe work that actually occurred.
Verification and Maintenance Make the Checklist Useful Over Time
A checklist should connect authorization to a defined review method. As we discussed in our analysis of California’s SB 574, human oversight means more than simply reviewing AI output; attorneys remain responsible for verifying the work before relying on it. For an AI-generated brief, that includes checking authorities, quotations, and factual support. For document classification, have a documented validation plan addressing both incorrect inclusions and missed responsive material, with separate attention to privilege and consequential errors.
The organisation should also plan for change. NIST’s voluntary Generative AI Profile approaches risk across the AI lifecycle and highlights governance, testing, provenance, and incident disclosure. That supports treating the checklist as an ongoing control, with reassessment after significant changes, rather than a one-time procurement document.
Conclusion
A practical AI policy checklist guides responsible, consistent use of AI, and periodic review ensures it remains relevant and useful as technologies, legal obligations, and organizational needs evolve.
Next time, we’ll provide a sample checklist to consider when creating the checklist for your own organisation.
For more regarding Cimplifi specialised expertise regarding AI & machine learning, click here.