93% of pastors are now either using AI or actively exploring it. Only 5% of churches have a written policy for it.
That gap is the whole problem. Using AI and using AI well are two very different things, and right now most churches are doing the first without any framework for the second. So we sat down and went through the AI use cases we actually see happening in churches — not hypotheticals, real things pastors are doing this week — and made a call on each one: inbounds, inbounds with guardrails, or out of bounds.
Fair warning: we don’t agree on all of it. One of us is AI-positive with a seat belt on. The other is more skeptical. That disagreement is the useful part, and we’d rather show it than smooth it over.
The Numbers Behind the Nervousness
A few worth naming before we get into the calls:
- 64% of churches say an AI policy is important. 5% actually have one.
- 75% of church leaders name theological misalignment as their top ethical concern with AI.
- 67% are worried about AI replacing human interaction in ministry.
Three-quarters of church leaders are worried about theology. Two-thirds are worried about losing the human part. And almost nobody has written anything down. The conversation is happening in everyone’s head and nowhere on paper.
That’s why we made a free AI policy template for churches — it’s linked in the episode description. It doesn’t tell you what to believe about AI. It gives you the categories to decide for yourself and get it in writing so your staff isn’t each guessing separately.
Sermon Prep: Three Very Different Questions
The most common AI conversation in ministry isn’t one question, it’s three. And they deserve three different answers.
Research: 100% inbounds
Using AI to compile what’s already out there — Strong’s concordance references, dictionary definitions, cross-references, commentary material — is a fantastic use case. It’s the same work a pastor would do manually with a shelf of books and a subscription library, just faster.
This one isn’t theoretical for us. I built a Claude skill specifically for pastor sermon research, and I use it every week I’m preaching. You give it a passage and it comes back with roughly a seven-page research document — concordance work, definitions, cross-references, commentary pulls. Plenty of pastors I know are running it now as part of their weekly prep.
(Try it on something you think you already know. Jeremiah 29:11 is the verse everyone quotes. Run the actual context and you may be surprised how much less cheery the surrounding passage is than the poster version.)
One personal caveat: it’s out of bounds for me to start there. If my first interaction with a passage is a research doc, something’s off — the same way starting with commentaries is off for me. Prayer, the text itself, and meditation come first. That’s a personal conviction more than a rule for everyone, but it’s the order we’d recommend.
Outlining: depends what you mean
If you mean taking your research document and saying now give me an outline for this sermon — out of bounds. That’s the part that belongs to you and the Holy Spirit. What God wants to say to your church, in your context, to your people, is your job. Handing that over is delegating the one thing you can’t delegate.
If you mean building your own outline and then reaching for help on language — you’ve got three points and no clean way to make them land, or you want them all to start with the same letter — that’s fine. That’s using AI as a thesaurus. It’s a word tool at that point, not a discernment tool.
Writing the whole sermon: never
Using AI to write a sermon and preaching it word for word is clearly out of bounds. We don’t think there’s a version of that we’d ever endorse, and we don’t expect to change our minds.
Here’s why it matters beyond the ethics of it: when you listen to a message, you’re not just there for the information. You want to hear the person’s heart. Their conviction. Their edge. The anecdote that only they could tell because they lived it. A made-up story that AI generated to fit the point perfectly isn’t the same thing, and people can feel the difference even when they can’t name it.
Website Chatbots: Two Levels, Two Answers
Support chatbots have been on websites for years. Churches are now putting them on theirs. But there are really two different products hiding under the same word.
Level one: the practical bot — inbounds
A bot that answers service times, parking, what to wear, what happens with your kids, where the entrance is. It’s pulling information that’s already on your website and handing it back. We’re fine with that.
The honest practical problem is that these bots tend to fail in one of two directions. Either they’re so locked down that anything remotely interesting gets “I’m sorry, I can’t answer that question,” which is worse than no bot at all — or they’re wide open, which is the next section.
Level two: the discipleship bot — this is where we disagree
A bot that answers spiritual questions and suggests next steps for new believers is a much harder call.
Vanilla ChatGPT dropped onto a church website is a no. It has no spiritual principles because it has no soul, and spiritual principles require one. And let’s be clear-eyed: there’s no reason to assume the people who built these models are born-again Christians hoping to see the name of Jesus glorified. That’s not a knock, it’s just true.
But it can be trained. You can put a wrapper on it, feed it your church’s sermons and theological documents, and get something that reflects your actual theology.
And that’s exactly what makes it uncomfortable. We had a client — a very conservative ministry, think John MacArthur — put a chatbot on a site we built. Great people, and a tradition quite different from mine; I’m Foursquare, spirit-filled, more open on things like women in ministry and the gifts. Their bot would say flatly that under no circumstances is it ever acceptable for a believer to speak in tongues. Which is precisely what they trained it to say.
So I tested it. I told it I’d been baptized at Elevation Church, by Steven Furtick himself, the weekend before — and asked whether my baptism counted or if I needed to be rebaptized. It told me to stop what I was doing immediately and get rebaptized right away.
You can train it to say whatever you want it to say. That’s the point. A human pastor facing that same question would almost certainly slow down, ask a follow-up, and handle it with more care than the bot did. The bot had no brakes.
If you’re going to do it, do it with loud disclaimers: this is experimental, it’s trained on our sermons and doctrinal statements, don’t treat it as final. Never hand it away as trustworthy on its own.
And here’s the pushback. Theology isn’t the only objection — and maybe not even the main one. Discipleship is relational. A robot can’t have a relationship with anyone. So even though it’s slower and less convenient than a 24/7 bot, the better answer for spiritual questions is a form that routes to an actual person at the church who then texts, emails, or calls back. Yes, humans get theology wrong too. But the relational part of discipleship isn’t a nice-to-have on top of the right answer — it’s close to sacred, and a bot structurally cannot supply it.
The counter-argument we keep landing on: people have become accustomed to answers on demand. If your church isn’t there when someone has a question at 11 PM — and full disclosure, my own church isn’t; we don’t do this — they don’t go without an answer. They go to a vanilla chatbot that will tell them whatever they want to hear. Maybe the real design goal is a bot that steers people toward a human conversation as the next step.
Where we both land: whatever we build, we never abdicate the personal part of discipleship. That’s something we’re called to do, and it doesn’t go away no matter how good the technology gets.
Small Group Questions from Sunday’s Sermon: Inbounds, With Guardrails
This one’s easy. Generating small group discussion questions from the weekend message is great.
We’ve run it as part of our workflow for over a year. It hasn’t been perfect — plenty of times we’ve gone back and forth on a question before landing on it. But not once has it produced something that would have led people astray if we’d published it. And it writes better icebreakers than we can, consistently.
The guardrail is the same one that applies to every “inbounds” on this list: you are the editor. You read it, you cut what’s weak, you’re accountable for what goes out. AI drafted it. You published it.
Pastoral Emails and Texts to People in Crisis: Out of Bounds
A friend of ours got an anniversary message from his mom recently. Em dashes everywhere. But the dead giveaway was the opening line, left in by accident: “Here’s an inspiring message for you to share with your son on his anniversary.”
It’s funny. It also stings, and it stings for a reason. Not because AI was involved, but because it meant the message wasn’t really from her.
Now, we should be honest that this is a spectrum, not a switch. If you have Grammarly installed, AI is already cleaning up your writing. The question isn’t whether AI touches your words, it’s how much you let it in. Where we land: write it yourself, always. If it comes out muddled, it’s fine to paste it back and ask for a clearer version of your own words. That’s the ceiling. Asking AI to find you a verse to encourage someone is over the line.
The bigger question is why you’d reach for it. There are only two real answers, and neither holds up:
Because it’s faster. If shaving ten minutes is the motivation, the heart isn’t in the right place. This is the job. These are your people. You should want to sit down and bless someone in your flock with something that came out of you.
Because you’re not a great writer. Probably true — the model is trained on better copywriters than most of us. But pastors managed this fine for a very long time before any of these tools existed. A slightly clumsy message that came from a real person beats a perfectly worded one that came from a robot. Every time.
Also, picture the prompt. Write a consolation letter to Joan who just lost her husband, make it sound like a mix of Max Lucado and Philip Yancey, and work in a verse. Read that back and ask whether it feels like ministry.
Giving Records and ChMS Exports: Inbounds Only If It’s Anonymized
There’s a clear line on this one, and it isn’t fuzzy.
AI is genuinely better than humans at data analysis, full stop. Take your giving records and you can ask questions that would otherwise eat a week: What does the average family unit give weekly? What did our top 10% of givers average last year? You get answers instantly. You should be doing this.
But strip every identifier first. Not just names — names, email addresses, phone numbers, home addresses, all of it. Assign each household a number and analyze the numbers. Your church management system may already have an ID you can use.
This is a live conversation among ChMS companies right now. CHmeetings — who partner with us on the podcast, link in the description — have built a read-only API for exactly this: the model can look at the data but can’t write to it. They also walk churches through turning off training-data retention in ChatGPT and Claude.
Which is good practice. But full disclosure: “we promise we won’t use your data” isn’t, by itself, enough for us to hand over our congregation’s financial records. The analysis is worth doing. The risk to your people’s data isn’t worth taking. Anonymize, then analyze.
AI-Written Prayers: 100% Out of Bounds
This one comes from a true story. A worship leader with a real following posted a prayer on X. Em dashes, the telltale cadence — and at the end of the tweet, before the character limit cut it off, the model’s own closing line: if you’d like this to be more encouraging… Copy-pasted straight in, exactly like the anniversary message.
Extreme example. Funny. Also clarifying.
AI can’t pray. It has no soul, and only things with souls can pray. So the question underneath a generated prayer is: who are you praying to? If a prayer is being written to be read by people rather than heard by God, the problem isn’t the tool. It’s a much older problem the tool just made visible.
Our position: if it’s a prayer, do it dirty. Whatever God is putting on your heart, say that. It doesn’t have to be polished.
Two honest caveats. First, pastors writing publicly after a disaster are careful with language, and should be — you don’t want to be flippant, you don’t want to overreach. Public words get weighed. That’s not the same as generating them. Second, there’s probably a real use case for a brand-new believer who genuinely doesn’t know how to start praying and asks AI for a place to begin. Maybe.
But we’re talking to pastors here. If you carry that title, being able to pray should be one of the qualifications.
And consider the math. Some percentage of prayers posted publicly by humans are performative — call it 5%. If it’s written by AI, that number is 100%. We’ll take our chances with the human.
Transcription, Captions, and Translation: Inbounds With Caveats
Auto-transcribing board and staff meetings, auto-generating captions, translating a service into other languages — we’re fine with this.
The value is real. Notes get taken automatically. People who don’t speak English can follow the message. That’s not a small thing.
It will also get things wrong. We’ve seen AI transcripts miss the actual meat of what was being discussed. This podcast is on YouTube with captions in many languages, and we’re confident some of those translations butcher what we meant. So label it: this is machine-translated, there may be errors. If you’re watching this in Spanish or Swahili, we’re glad it’s available — please give us the benefit of the doubt on the wording.
Thank-You Texts to First-Time Guests: Not Unethical, Just Silly
We’ve talked a lot about visitor follow-up because it works. A text from the actual pastor on Sunday afternoon — so good to meet you and your kids today — moves the needle more than almost anything else a church does.
Using AI to write it isn’t an ethical dilemma for us. It’s just dumb.
Think about what you’d have to type: I need a thank-you-for-visiting text for Joe and his wife and their three kids, Susie, Joe Jr., and Eddie. First time this Sunday. I met Joe, he works down at the plant, his wife stays home with the kids, the kids are at two different schools. The prompt is longer than the text. You already know all of that. Just write it.
The value of that text is partly that it’s from the real pastor, and partly that it references something only a human who was there would know. Strip both and there’s nothing left worth sending.
If you’re stuck on wording, we’ve published the exact follow-up scripts we use at my church — free, linked in the description. Don’t copy the texts word for word; copy the flow and the tone. Do it twice and you’ll have the hang of it.
The Thing Underneath All of It
Step back from the individual calls for a second.
A pattern shows up across every one of these use cases: people start using AI for work, and within a year they’re using it for everything. Including things they’ve done comfortably by hand their entire lives. It’s gotten so normalized that we’re now spending real effort not to use our own brains — which is somehow more work, not less.
You don’t have to use it for everything. We were fine before AI existed. Most of what it’s being reached for isn’t work that needed help.
And there’s a cost to the reaching. We’re getting worse at some of these things, and losing confidence in them, because we’re not working the muscle. Teachers are worried about exactly this with students. It’s a real concern for writing specifically.
The fair comparison is a calculator. It’s a great tool. It’s here permanently. And there’s no question people are worse at arithmetic than they used to be. That’s the world we live in — but it’s still worth keeping the muscle in shape.
One more thing worth saying: people crave the humanness. If you misspell a name in a text because you typed it yourself, so be it. That’s not a defect. That’s the signal that a person was on the other end.
We’d genuinely love to hear where you disagree — this topic generates more pushback than anything else we put out, and that’s a feature. Tell us in the comments what we got wrong, or what use cases we missed.
Your Next Step
Not sure where your church stands on any of this — your website, your follow-up process, your digital front door? We offer a free church marketing and website review. Our team will look at your site, your search presence, and your overall digital strategy, then send back honest, specific recommendations you can actually use. No sales pitch. Real feedback.
Turnaround is about 48 hours. It’s free for any church that asks.
