Why an Executive Pastor Uses AI Every Day (and Where I Draw the Line)

This is the opening post in a six-part series on how I use Claude as an executive pastor. It sets the ground rules. The posts that follow cover research, prompt and context engineering, running operations, and building software, then close with the full stack and what I'd tell a church starting today.

The last time anyone updated Mosaic's AI Use Policy was November 27, 2023. I wrote it.

At the time, it was a reasonable document for a reasonable moment. AI meant a chat window. Staff were starting to paste in a paragraph and ask for a cleaner version. The questions were about drafting emails and whether it was acceptable to let a machine help write an announcement.

This fall I opened that policy to revise it, and most of it no longer describes how our staff, or I, actually work. Tonight, while I sleep, an AI-driven briefing will scan my calendar, my inboxes, Slack, and our task systems, and have a summary waiting for me in the morning. The software our staff use for HR, facilities requests, meeting notes, and budget visibility was built with an AI writing most of the code. The question has changed from "may we use this?" to "what is it allowed to touch?"

This series is my attempt to answer that in public.

The 2023 policy got the values right. It protected people's likenesses and voices. It committed to regular review. I shared it with a cohort of executive pastors in northeast Ohio this spring, and I'd still defend those principles.

What it got wrong was the world it assumed.

It assumed a chat window. It said nothing about AI that connects to our systems. Today our Claude Team plan connects to Microsoft 365, Planning Center, and QuickBooks. Agents draft and edit things. Meetings get transcribed. Some of our own software calls AI directly.

It assumed staff train models. Almost nobody does. The AI assets a church actually builds now are projects, skills, plugins, agents, scheduled automations, connectors, saved memory, and API keys. A policy that says "Mosaic owns the models" protects nothing, because there are no models. It needs to name the things that exist.

It contradicted itself. One rule required disclosure for anything AI-assisted that looked human-made. The next allowed work to go out without attribution if it was properly vetted. In 2023 you could live with that tension. In 2026, nearly everything a staff member writes has had AI involved somewhere. The disclosure rule couldn't be followed, and a rule nobody can follow teaches people to ignore the rest.

It had no owner. Phrases like "regular audits should be conducted" never said by whom. They were aspirations without anyone responsible for them.

I'm rewriting it rather than patching it. The rewrite is still in progress, and nothing below is final. The rest of this post is the thinking behind it, because it's also the frame for everything else in this series.

Mosaic Northeast Ohio has about 750 people on a Sunday, about 1,500 who call it home, and 11 full-time staff. I carry executive responsibility for finance and IT, and I've recently stepped into a role helping manage our ministerial staff, Adult Ministries, and small groups. AI hasn't replaced anyone on that team. It has taken over a category of work that used to fall through the cracks or land on my evenings.

Here's the map for the rest of the series:

Each post follows the same pattern: the problem, what I built or changed, how it works, where it failed, and something you can take and use. I'll show the failures, because a series that only shows wins is an advertisement.

The useful question is no longer whether staff may use AI. It's what AI is allowed to touch, what it's allowed to do on its own, and what has to be labeled. Here's what we're considering for the rewrite. These are proposals under discussion, not adopted policy.

The draft would require Mosaic work to happen in Mosaic accounts. For us that's a Claude Team plan, bought with the nonprofit discount, not a staff member's personal account or a free tier. Business and consumer plans come with different data terms, and you can't govern what you can't see. If your staff are using personal accounts for church work today, that's the first thing to fix, before any policy language.

The addition I expect to matter most is a simple data classification. The current draft sorts information into four classes, and the class decides where it can go:

The Restricted row is the one that matters most. Some information is entrusted to a church in a way that rules out any convenience argument. No efficiency gain justifies putting a counseling conversation or an abuse disclosure into an AI tool.

A rule on paper only works if people remember it. Where I can, I build the line into the software, so the AI can't see what it shouldn't.

Two examples. In my own task system, any item tagged as care-related comes back to an AI agent with the title redacted. The agent can tell me a care item exists. It can't tell me what it says. I have to open the app myself. And our in-house HR system runs in its own isolated database, with personal and compensation fields masked by default.

Once AI can act, not just answer, a policy needs rules about action. These are the ones we're weighing:

That last one sounds bureaucratic until you count your automations and realize you can't name them all.

This is where my thinking has changed the most, and the proposed language reflects it. The 2023 policy leaned toward disclosing AI help broadly. I no longer think that's the right line for text and graphics. If a staff member uses AI to help write an announcement, and they've reviewed and approved it, the announcement is theirs. Whoever sends or publishes something owns its accuracy and tone, however it was made. One condition: when content is drafted for someone else, that person approves it before it goes out under their name.

Under the proposal, what must be labeled is anything a reasonable person could mistake for a real photograph, recording, or person. An AI-generated image of a "family" on a flyer. An AI-altered video. A cloned voice. Using a real person's likeness or voice requires their permission in advance, and the result is still labeled. The draft also rules out synthetic likenesses of minors entirely. Routine edits like cropping, color correction, and removing background clutter don't count. Edits that change who was there, what was said, or what happened do.

In August, an audit of the AI tools I'd built turned up a problem in one of my own. A plugin I used to process meeting transcripts had started storing profiles of how individual colleagues behave in meetings, for its own convenience in recognizing who was speaking during non-labeled written transcripts. Nobody had asked for those profiles. Nobody had consented to them.

I deleted the inappropriate files, and I pulled it from our internal plugin library entirely. I didn't mark it unavailable; I removed it. It's the clearest example I have of why the rules above are worth writing down. The risk usually isn't a bad actor. It's a useful tool quietly collecting more than anyone intended, until someone checks.

It would be easy to read this series as "AI does the job now." It doesn't. Four things stay with people, and I don't expect that to change.

Accountability. If something goes out with an error in it, "the AI wrote it" is not an answer anyone at Mosaic gets to give. The person who sent it owns it. That rule does more to keep quality up than any review process I could design.

Judgment. Claude is very good at gathering options, sorting evidence, and pointing out what I missed. It's also capable of being confidently wrong, and the next post shows several times it was. The decision still belongs to someone who will answer for it, to our elders and to our people.

Pastoral presence. Nobody's counseling conversation, prayer request, or crisis goes into an AI tool, and no AI tool sits in for a person when someone needs one. Most of what I use AI for is administrative work, and part of the point is to free up time for the work that has to be done face to face.

Ownership of what gets built. I've built a lot of software for Mosaic with AI writing most of the code. That doesn't make it free to maintain. Right now I'm still the person on staff who can keep most of it running, and that's a real risk. I'll be honest about it in the post on building, along with what I'm doing to reduce it.

If your church has no AI policy, or one written in 2023 like mine, answer these questions before you write a word of policy language. The answers are the policy. The document just writes them down.

One more set of questions, for any staff member or volunteer building tools with AI. You don't need technical knowledge to ask these, only the willingness to ask about their specific project:

If they can't answer the third question, they haven't been checking the work.

Next in this series: how I use Claude for research, and why I treat it as a research assistant instead of a source.