Checklist · AI decision systems
Approval and training record for a generative AI tool (United States)
The record a US firm completes before a generative AI tool is put on the permitted list, covering confidentiality settings, retention, reliability, liability, notice of breach, and who has been trained.
- Kind
- Checklist. Steps or checks to work through and tick off.
- Where it applies
- United States
- AI use
- Yellow You and the tool draft together, with you directing. You check every statement you rely on.
- Law stated as at
- 22 October 2026
- Status
- Maintained. Checked on 22 October 2026 against the 7 authorities it cites, for United States.
- Version
- 1
- Source
- Written for this library
When to use it
Use before a tool is added to the firm generative AI policy register, and again when the provider changes training, retention, access or the model. Do not use a blank record as permission to try the tool on a client matter.
Used for these tasks
- Map automated decisions that affect individuals in the firm's or client's system
- Draft a privacy-impact or algorithmic-impact assessment for the system
- Compare the system's outputs on a test set for disparate results
Each task page says how far AI may help with the task and what the lawyer must review.
How AI may be used
A tool may lift clause references from vendor terms the lawyer has collected into the answer column. The lawyer decides whether the answers are good enough to approve the tool.
AI may
- Place short quotations from the attached terms into the answer column, each with the clause name
- Mark a row whose answer is "not stated" or is supported only by a sales page
- List lawyers and staff who are not yet on the training acknowledgment
AI must not
- Must not mark the tool approved
- Must not fill a gap in the terms with an inference
- Must not paste client information into the tool in order to test it
- Must not treat a marketing claim about security as the contractual term
The lawyer checks
- Each answer that matters is tied to a clause, not to a brochure
- Self-learning and any ability to send, file or reach other systems are answered yes or no, not left blank
- The lawyer has decided the information class, and Class B is refused if consent cannot be made informed
- A fresh record is required when the model or the data terms change
Method: how the Green, Yellow and Red levels are decided. This is guidance for planning, not legal advice.
Read the full text
The full text is available after you log in. It is free. When you log in you can:
- read the whole precedent, with its drafting notes beside it;
- copy it, or download it as a Word file or Markdown;
- add it to your firm’s knowledge base, so colleagues can find it;
- let the AI tools you connect read it from that knowledge base.
The summary, the law it rests on and the AI-use guidance above stay public. Privacy.
More in AI decision systems
- Inventory of automated decisions that affect individuals
- Law firm AI use policy (master firm policy)
- Law firm generative AI use policy (United States)
- Approved AI tools register (template with field guide)
- Record of a test for different results across groups
- AI vendor due-diligence questionnaire for law firms (master version)
- Legal risk table for a client's AI system
- How to mark generative AI output in the file (United States)
- AI incident response procedure for law firms (master procedure for all AI incidents)
- Record that a person reviewed an automated output
- File note of a generative AI confidentiality incident (United States)
- AI decision system impact and fairness assessment (firm or client systems)
- Reassessment when a model, purpose or provider changes
- Privacy policy clause on automated decisions (APP 1.7 to 1.9, from 10 December 2026)
All precedents for AI decision systems · Tasks and AI use in this area