Steward Agents uses AI models to help non-profit staff draft and analyze. This page explains what the agents do, what stays with your people, and what AI gets wrong.
1. What agents do
- Draft: board updates, briefing notes, replies to funders, newsletter copy, grant sections, budget notes and similar work, from the facts and documents you give them.
- Analyze: summarize documents, email threads and spreadsheets; explain figures; check a board package against a checklist; point out what's missing.
- Cite: in the Office add-in, drafts built from your files, sheets or emails cite each source, so you can check where every figure came from. Citations the agent can't back up are removed and flagged.
2. What agents never do
- Send email. In Outlook they create a draft or open a reply for you; you send it.
- Publish, post, submit applications, sign or file anything.
- Change existing numbers in Excel. They add rows below your data or a new sheet.
- Make decisions about people: clinical, safety, eligibility, employment or funding decisions stay with qualified staff.
- Give legal, tax, accounting, medical or clinical advice.
3. Your staff stay responsible
- Everything an agent produces is a draft. A person reviews it, edits it and decides whether to use it.
- Your organization is responsible for the final content and for decisions made using it.
- High-stakes external material, such as anything going to a board, funder, government, media or the public, needs approval by the appropriate person in your organization before it goes out, following your own approval process.
- Agents mark facts they couldn't confirm as [to confirm] or leave a [placeholder] rather than guessing. Please resolve these before using a draft.
4. Known limits of AI
- It can be wrong. AI can misread a table, mix up periods, or state something false confidently. Check figures, names, dates and quotes against the source.
- It can miss context. It knows only what's in the request, the sources you give it and its general training. It doesn't know your relationships, history or unwritten rules unless you tell it.
- It can be out of date. Laws, funder programs and policies change. Agents are told not to state funder deadlines, eligibility or legal requirements unless you provide the source, but always confirm them yourself.
- It can reflect bias. Review drafts for fairness and for respectful, person-first language, especially about mental illness and the families you serve.
- Safeguards aren't perfect. The check that stops client health information from reaching the AI can miss things or flag harmless text. Don't rely on it in place of your own care.
5. Which AI models are used
Agents use AI models from Anthropic. Everyday requests use a faster model; long or high-stakes drafts, such as board reports or grant narratives, use a more capable one, and the agent tells you which it used. Where the AI processing happens is described in the Privacy Policy.
6. Designed with the person who uses it
Most agents are designed through an onboarding interview with the staff member who will use them, so the agent knows their role, voice and boundaries. Ready-made agents start from a role template. Agent instructions change only through a reviewed change request.
7. Questions or concerns
If an agent produced something harmful, misleading or inappropriate, tell us at [support email] so we can look into it.