1. Find answers in approved internal documents
Let staff ask questions about a bounded collection of policies, procedures and program guides. Require links to the source passages so a person can check an answer. Start with documents that have an owner and a clear review date.
Test with questions staff already receive, including questions the collection cannot answer. Measure whether the system locates the right source and admits when the answer is missing. Respect document permissions rather than making everything available to everyone.
2. Draft plain-language versions of program information
Use an assistant to produce a draft from an approved source, then have a program owner verify eligibility rules, dates, names and instructions. Keep the source and the final approved version together.
Evaluate clarity with intended readers. A fluent rewrite can still change meaning, so the review should compare the facts as well as the tone. Publish only after a person accepts the content.
3. Help route incoming requests
Suggest a category and responsible staff group for an incoming message. Use a human review step before routing sensitive or uncertain cases. Keep the category list small and give the system an “unsure” option.
Measure incorrect suggestions and how much checking they require. If staff spend more time correcting the suggestions than reading the original messages, adjust the task or stop the pilot.
4. Draft recurring communications from verified inputs
Create first drafts of event reminders or program updates from a structured brief containing the approved facts. Have the communications owner review every draft before sending or publishing.
Use sample or approved information during development. Compare the final editing effort with your current process; generating a long draft is not useful if staff have to remove most of it. Keep sending and publishing outside the pilot's automatic actions.
5. Assist with a content inventory
Suggest topics, duplicate content and potential owners for a website migration. Treat the output as a proposed inventory for editors to review. Do not let a model delete pages or decide retention requirements.
Evaluate a small sample manually, then look for useful patterns. Measure how many suggestions editors accept and whether important items are missed.
Use the same pilot plan for all five
- Describe one task and the current process.
- Name approved data, access limits and the responsible owner.
- Choose test cases, including failures and ambiguous requests.
- Define what a human must check before an output is used.
- Set an acceptable result and a condition for stopping.
- Review operating costs, data handling and the supplier's current terms.
NIST's AI Risk Management Framework provides a broader structure for evaluating and managing risk. For your first project, turn that into concrete ownership, test cases and review responsibilities.
Use the project scope worksheet to bring a candidate task to a conversation. We can help decide whether AI, ordinary automation or a simpler process change fits it.
Prepared by Awesome Cause with AI assistance. These guides provide planning advice; project scope and estimates are agreed individually.
