“AI-powered” tells a school almost nothing. The useful questions are more specific: What information can the assistant use? Who approves it? Who can ask? What happens when the source is missing? What data reaches a provider?
School AI should begin with boundaries, not a demo prompt.
1. What job is the assistant doing?
A public family assistant might help locate calendars, policies, enrollment steps, forms, and resources. An authenticated staff assistant might work with private role-limited material. These are different risk and access contexts even if the conversational interface looks similar.
Write down the intended questions, users, sources, and prohibited uses before evaluating model quality.
2. What are the approved knowledge sources?
Ask whether answers come from school-approved files and pages, open web search, vendor material, or a mix. Confirm how sources are added, reviewed, updated, removed, and separated by school.
3. Is the context public or authenticated?
A public assistant should not have access to private role-based resources. An authenticated assistant should evaluate school, tenant, role, and permission before retrieving private context. Ask whether a user changing roles or campuses changes available sources.
Auzi’s trust model separates public knowledge from authenticated/private contexts and describes access by school, tenant, and role.
4. What data reaches an AI provider?
Request an AI data-use statement and current subprocessor list. Ask what prompt, account, content, metadata, and logs reach each provider; where processing occurs; how long data is retained; and what contractual restrictions apply.
Do not accept a broad “we never train on your data” statement unless every relevant provider, product configuration, and contract supports it. The statement should be verified, not inferred from a model brand.
5. How are accuracy and failure handled?
- Does the answer cite or link to approved source material?
- Can the assistant say it does not know?
- Can a user report an incorrect or outdated answer?
- Who reviews reports and changes the source?
- Are high-risk topics excluded or routed to staff?
- Are responses and administrative changes logged where appropriate?
6. Who is responsible for human review?
The vendor should maintain the product and document technical responsibilities. The school should appoint owners for approved sources, role assignments, policy decisions, and escalations. A shared-responsibility table is more useful than a “responsible AI” badge.
7. What are the usage and cost boundaries?
Model calls have variable costs and different model tiers. Ask what standard allowance is included, what happens near a limit, whether premium models cost more, and whether the vendor can add surprise overages.
A defined allowance with notice is more credible than “unlimited AI.”
8. What documentation is available?
A school should expect a Data Processing Agreement, privacy policy, subprocessor list, retention and deletion process, incident-response process, security overview, AI data-use statement, accessibility documentation or roadmap, and export and termination process.
The U.S. Department of Education has emphasized that education AI requires attention to trust, safety, privacy, civil rights, and human oversight. Review current official guidance from the Department of Education’s AI resources alongside local counsel and policy.
A decision-ready set of questions
- What exact job and audience are approved?
- Which sources can be used, and who governs them?
- How are public and authenticated contexts separated?
- What reaches each AI provider?
- Which training, retention, and deletion terms are contractual?
- How does the product handle uncertainty and correction?
- Which actions require a human?
- What usage is included and how are limits communicated?
- What logs, exports, and termination processes exist?
- Which claims are verified today versus roadmap items?
Make the comparison concrete.
Auzi publishes its starting price, launch cost, standard scope, and common add-ons before a sales call.