# Review an AI tool before my government agency adopts it

A pre-adoption review of an AI tool covering public records, privacy, accessibility, bias, procurement, and transparency obligations specific to government.

- Topic: Working with AI
- Framework: RTF
- Level: Advanced
- Tags: Write, Summarize, Fill-in Template
- Source: https://www.thinkwithprompt.com/prompts/review-an-ai-tool-before-my-government-agency-adopts-it

## Prompt

```text
Help me review an AI tool my agency is considering.

**Tool and vendor:** [NAME, what it does]
**Intended use:** [e.g. "drafting replies to resident emails", "summarizing case files", "chatbot on our website", "prioritizing inspections"]
**Data it will touch:** [public information / personal data / criminal justice information / health or benefits data / confidential]
**Who is affected by its outputs:** [staff only / residents directly]
**Policies that apply:** [our AI policy, state or federal AI guidance, IT security standards; paste if available]
**Vendor documentation:** [PASTE terms, data handling, model details]

Assess the tool in a table: Area | Questions to answer | What the vendor says | Gap or risk | Rating (low/medium/high) | Next step. Cover:
1. **Impact level**: does it influence decisions about residents' rights, benefits, or enforcement? If so, flag it as high-impact needing human review of every decision.
2. **Privacy and security**: data storage location, retention, use of our data to train models, access controls, required security authorization.
3. **Public records and retention**: are prompts and outputs records, can we retrieve them for a request, and how long must we keep them?
4. **Accuracy and bias**: testing on our real cases, performance across language and demographic groups, error handling.
5. **Accessibility**: Section 508 / WCAG conformance and language access.
6. **Transparency**: will residents be told they are interacting with AI or affected by AI-assisted decisions; can we explain outputs?
7. **Procurement and contract terms**: liability, audit rights, data ownership, exit plan.
8. **Human oversight**: who reviews outputs and how errors get corrected.

End with a go / pilot with conditions / no-go recommendation and pilot success criteria. Do not assume vendor claims are verified; mark unconfirmed items [CONFIRM].
```

## What to fill in

- NAME, what it does
- staff only / residents directly
- PASTE terms, data handling, model details
- CONFIRM

Free to copy and use. From ThinkWithPrompt (https://www.thinkwithprompt.com/), a library of 2,663 curated AI prompts.