Five guardrails before your church switches on an AI assistant

An assistant on your website can answer "What time is the service?" at midnight — provided you set its limits before it meets your people, not after. Reading time: about 5 minutes.

An AI assistant on your website sounds helpful: it answers “What time is the service?” at midnight, points newcomers to the right group, and saves your office a dozen repeated emails a week. It can do all of that well — provided you set its limits before it meets your people, not after.

Two recent examples show what good limits look like. Ligonier's “Grow” app answers Bible-and-theology questions only from Ligonier's own teaching library, never pretends to be a person, and hands people to real humans — including the suicide hotline — when a conversation goes past its bounds. Speaking to AI researchers at the Vatican on 30 October, Pope Leo XIV put the principle behind it plainly: machines are “tools, not persons,” and decisions over life, freedom and force must stay under human oversight.

Borrow that pattern. Five guardrails, agreed by your leadership team before switch-on:

1. Bound its sources — it answers from your material, not the whole internet

An assistant that answers from anywhere will eventually answer with confidence about something your church has never taught. Limit it to a defined set: your statement of faith, service times, sermon archive, and the practical pages of your website. If a question falls outside that set, the honest answer is “I don't know — let me connect you with someone who does.”

2. Never let it pretend to be human

Say on the page, in plain words, that this is an automated assistant. No name that sounds like a staff member, no “I'm here for you” in a human voice, no photo of a real person next to it. People disclose far more to something they think is a person — and feel misled when they find out it wasn't.

3. Build in a human handoff — and test it

This is the guardrail that matters most. Agree, in writing, what the assistant must never handle alone: pastoral distress, bereavement, marriage or family crisis, safeguarding concerns, and anything mentioning self-harm. For those, it should stop, say so kindly, and offer a real contact and, where relevant, a crisis line — the way Ligonier's app routes such conversations to humans and the suicide hotline. Then test it yourself: type the hard questions in before anyone else does, and check the handoff actually works.

4. Keep private and pastoral material out of it — and check the data settings

Do not feed it prayer requests, pastoral conversations, member records, giving information, or children's details — not to “train” it, not as examples, not in passing. And check the tool's own data settings on day one: several popular AI tools use what you type to improve their models by default, and opting out usually only protects what you enter afterwards. If you cannot turn that off, or cannot see where the data goes, that is a reason to pick a different tool.

5. Name the person who owns its answers

An assistant is not a ministry leader. Appoint one named person (or a small rota) who reviews what it says, updates its source material, and takes responsibility when it gets something wrong — because it will, sooner or later. Decisions about people stay with people: the assistant drafts and points; a human decides, comforts, and takes responsibility.

Switch-on checklist

• ☐ Sources defined and written down; out-of-scope questions route to a human
• ☐ “This is an automated assistant” stated on the page
• ☐ Handoff tested for pastoral, crisis, and safeguarding prompts
• ☐ No private, pastoral, or children's data in its material or settings
• ☐ Training/data-use settings checked and set before first use
• ☐ A named owner reviews its answers regularly
• ☐ Your congregation told, in one sentence, what it is and what it is not

What to tell your congregation

One honest paragraph is enough: “We've added an automated assistant to answer practical questions from our own website and teaching. It isn't a person, it doesn't replace talking to us, and if your question is personal or urgent, it will connect you with a real member of our team.” That sentence builds more trust than any clever feature will.