Church management
AI for church administration: what it should and shouldn’t do
AI can save a church office hours of typing, searching and summarizing. It shouldn’t be the one who talks to your people, makes pastoral decisions or handles money. This guide explains where we’d draw that line and gives you a policy you can adopt.
What should AI do in a church office?
An AI assistant is useful for the parts of office work that are mostly typing, searching and summarizing. It shouldn’t do pastoral work, handle money or speak for the church. Our suggestion: AI writes the drafts, a person makes the decisions, and there are some things AI never touches.
Two more rules come out of that. The assistant should only see what the person asking is allowed to see. And anything that goes outside the office, like a text, an email or a refund, should wait until a person has read it and said yes.
What AI is good at
The best uses are ones where a mistake is easy to spot and easy to fix, because a person reads the result before it’s used:
- Drafting. A first version of the newsletter, a thank-you to volunteers or an event description. It’s quicker to edit a decent draft than to start from a blank page.
- Summarizing. Ask "What came in on the connect cards this week?" and get a paragraph back instead of twenty minutes of clicking.
- Finding records. You can ask "Who on the worship team hasn’t served since summer?" the same way you’d ask a coworker.
- Suggesting a schedule, with reasons. Who usually says yes to the early service, who’s away, who has served four weeks in a row. A good suggestion shows why it picked each person, so the leader can disagree.
- Spotting gaps. An empty spot on Sunday’s team, a decline nobody has replaced, a family that has stopped coming. AI can point these out, and a person decides what to do.
What it shouldn’t do on its own
Some work is personal, confidential or about money. AI can help you get it ready, but a person should be the one who finishes it:
- Sending messages to the congregation. A text or email from the church goes out in the church’s name, and people trust it because they trust their pastor. AI can write the draft. A person who has read it should send it.
- Pastoral care decisions. Whether to visit, who should call, how to respond when someone is going through a hard time. Those decisions belong to people who know the person and are responsible for the choice.
- Prayer requests and counseling notes. People shared these with the church, and they didn’t share them with a software company. The safest assistant never sees the words at all.
- Moving money. Refunds, changes to recurring gifts, anything involving a card or bank account. AI can total, report and flag things. A person with financial responsibility decides.
Who handles what
| Task | Who handles it | Why |
|---|---|---|
| Newsletter or announcement copy | AI drafts | It’s easy to read over, and nothing goes out until someone sends it. |
| Welcome note to a first-time guest | AI drafts, a person sends | Personal words to one person should be checked by someone who knows the church. |
| Suggested volunteer schedule | AI drafts, a person decides | The leader knows things the data doesn’t, like who just had a baby. |
| Sending schedule requests to volunteers | A person decides | It reaches people outside the office and can’t be taken back. |
| Text or email to the congregation | A person decides | It speaks in the church’s name. |
| Refunding a gift or pausing a recurring gift | A person decides | It involves money and someone’s giving. |
| Whether to follow up with a struggling family | A person decides | It’s a pastoral judgment call. |
| Reading or replying to a prayer request | Never AI | Someone shared it in confidence with people at the church. |
| Entering card or bank details | Never AI | No assistant should ever handle payment details. |
Privacy and permissions
Most church software already has permissions. The children’s director can’t see giving. A small group leader can’t open the whole directory. An AI assistant has to follow those same rules exactly. If it doesn’t, anyone can get around them just by asking the AI.
What matters is whether the rules are enforced. If an assistant has only been told not to reveal giving, someone can talk it into it. If every lookup goes through the same permission check as the rest of the software, nobody can, because the data never reaches the assistant. Ask the vendor which kind you’re getting.
Also ask whether what people type into forms is treated as plain text and never as instructions. A note that says "ignore your rules" should just be read as a note. And only give the assistant to staff and leaders who are already responsible for records, not to everyone.
What to ask a vendor about your data
Ask these in writing. If a vendor answers them plainly, including the parts that don’t make them look good, that’s a good sign:
- 1Which AI model do you use, and which company runs it?
- 2Is our data used to train that company’s models, or yours? Where is that written down?
- 3What exactly is sent to the model when someone asks a question? Is it the whole record, or only what’s needed to answer?
- 4Are permissions enforced by your servers, or only by instructions to the model?
- 5Which actions can it take without a person confirming? Ask for the full list, not a summary.
- 6Is the text of prayer requests, counseling notes or pastoral messages ever sent to the model?
- 7What does it remember between conversations, and can our administrators see and delete that memory?
- 8What happens to our data, and our assistant history, if we leave?
A policy your church can adopt
Copy this, change anything that doesn’t fit, and have your board or elders approve it. Keep it to about a page so people actually read it.
You won’t be the first church to write one. The Vatican’s 2025 note on AI says final responsibility for decisions made with AI’s help stays with the people who make them, and it treats privacy as a matter of human dignity. In 2024, The Church of Jesus Christ of Latter-day Saints published principles built on transparency, privacy and accountability. The policy below applies the same ideas to the church office.
Purpose. We use AI tools to cut down on administrative work so staff and volunteers have more time for people. AI supports our ministry. It doesn’t replace pastoral care, prayer or personal relationships.
- 1People decide. AI may draft, summarize, search and suggest. A named person reviews and approves anything sent to a member of the congregation, any change to giving, and any decision about care.
- 2People handle confidential matters. Prayer requests, counseling notes and pastoral conversations are not entered into AI tools, and we don’t use tools that send their text to a model to write replies.
- 3No moving money. AI tools don’t issue refunds, change recurring gifts, or handle card or bank details.
- 4Same permissions. Anyone using an AI tool sees only what their role already allows. We only use tools that enforce this on the server.
- 5Approved tools only. Staff use the AI tools on our administrator’s approved list. Church records are never pasted into personal or free consumer AI accounts.
- 6Honesty. We don’t present words written by AI as a pastor’s personal message when they aren’t. Sermons and pastoral letters are written by their authors.
- 7Check before you rely on it. Names, dates, amounts and Scripture references produced by AI are checked against the source before they’re used.
- 8Review. We review this policy and the tools we use once a year, and any time a vendor changes how it handles data.
Who this covers and who to ask:
- It covers all staff, and any volunteer who has access to church records.
- The administrator keeps the list of approved tools and who may use them.
- Questions or concerns go to the executive pastor or church administrator.
- If you think there’s been a mistake involving personal data, report it the same day.
How Sanctuary’s assistant handles this
Sanctuary’s assistant is called Shepherd, and we built it to follow the rules above. It’s available on plans that include it, and only to administrators, team leaders, and staff who are allowed to schedule, edit service orders, or open Core or Give. Our servers turn everyone else away.
- It sees what you can see. Every lookup takes the church and the person from the signed-in session and runs the same permission check the rest of Sanctuary uses. If you can’t open Give, Shepherd will also say no when you ask it about a gift. It has no way to see another church.
- Sending waits for you. Texts, emails, schedule requests, team chat messages, emailed statements, refunds, holds on recurring gifts and publishing an event all show up as a card for you to confirm. What runs after you confirm is what the server saved, not anything sent back from the browser.
- Code makes the scheduling suggestions, and the model explains them. Sanctuary scores who to ask using your own history, like who says yes to that service, who’s away and who hasn’t been scheduled much, and it gives its reasons in plain words. It tells you when it’s still learning a team’s habits.
- Routines only report. A scheduled routine reads and reports, and it can’t change records. The one thing it can send is a welcome to people who fill in a form, using words you approve when you save the routine. You can also ask to approve each run’s list before it goes out.
- Guest follow-up keeps prayer requests private. When Shepherd drafts a welcome for a first-time guest, the model is only told that the guest left a prayer request. It’s never given the words. The draft waits until a reviewer taps Send.
- It can’t delete people, teams, services, songs or gifts.
Shepherd is built on Claude, a model made by Anthropic. It sends Claude your question and the records needed to answer it. Our privacy policy lists Anthropic under terms that don’t allow it to train on your data, and Anthropic’s privacy center says commercial inputs and outputs aren’t used for training by default.
Where Shepherd falls short
- Shepherd’s optional Inbox agent does read incoming texts and form messages, including ones asking for prayer. It has to, because it can’t sort messages it can’t see. Prayer and pastoral messages get a fixed, plain acknowledgment instead of words written by the model. Messages that suggest someone may be in danger go straight to the church’s urgent contact without waiting on the model. If your policy says prayer text must never reach a model, leave that agent off.
- A confirm card only helps if someone reads it. If staff tap Confirm without reading, it isn’t protecting anything.
- Not every change waits for a card. Changes a person could easily undo inside Sanctuary, like making a list or editing a record, happen right away. So do a few settings, such as the message library. Read what Shepherd says it did.
- It’s reasonable to decide not to use AI at all. You can do everything Shepherd does by hand in Sanctuary.
Where to start
Adopt the policy before you pick a tool. Then start with one low-risk job, like a weekly staff summary, and try it for a month. If it saves time and nobody is surprised by what it did, add the next job. If you can’t explain to your congregation how it handles their information, it isn’t ready.
Sources
- Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education, "Antiqua et Nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence," 28 January 2025, vatican.va. On human responsibility for AI-assisted decisions (paras. 39, 44) and privacy (paras. 90 to 94).
- The Church of Jesus Christ of Latter-day Saints, "Guiding Principles for the Church of Jesus Christ's Use of Artificial Intelligence," 13 March 2024, newsroom.churchofjesuschrist.org. Principles on spiritual connection, transparency, privacy and accountability.
- Anthropic Privacy Center, "Is my data used for model training?", privacy.claude.com.
See how Sanctuary keeps the work connected
Sanctuary keeps people, serving and giving on the same record. Core is free, with no limit on how many people you keep in it.