AI tools can save a small web team real time: drafting a news post from meeting minutes, suggesting alt text, rewriting a page in plain language, translating to Spanish. They can also publish confident mistakes. For a government, a wrong date or fee isn’t just embarrassing — residents act on it.
These are the guardrails we built into Sentra CMS, and the ones we suggest for any AI use on a public website. For a broader framework, see NIST’s AI Risk Management Framework.
1. AI drafts; people publish
Nothing an AI writes should reach residents without a person reviewing it. In Sentra CMS, the assistant can only create drafts, shown as before-and-after changes, and the normal approval path still applies.
2. Keep the same permissions
The assistant shouldn’t be able to do anything the person asking can’t. If an editor can only change Utilities pages, so can the assistant working for them.
3. Check the output like any other content
AI-written pages go through the same accessibility checks, reading-level check and approval workflow as everything else. Facts like fees and dates should come from a single source of truth, not from the model’s memory.
4. Protect personal information
Don’t paste residents’ personal information into AI tools. Use providers whose terms exclude training on your data, and document what is sent where. Let administrators turn AI features off.
5. Be transparent with residents
When AI answers residents directly — for example in site search — show the sources it used, link to the full page, and make it easy to reach a person. Record what the AI was asked and what it answered, so you can review it.
Where AI helps most
- Alt text suggestions that an editor confirms
- Plain-language rewrites of dense pages
- First-draft translations for human review
- Turning approved minutes into a news post and a list of pages to update
- Finding every page that mentions an outdated fee or phone number