What We Won't Let the AI Do

I want to publish the opposite document.

Holdfast has eight AI features. I'm proud of them. But the decisions I actually agonized over weren't about what the AI could do — they were about what it must never be allowed to do. Those decisions are the product. So here they are, written down, in a place you can hold me to.

1. It cannot invent a fact about your church

This is the one that matters most and it's the one nobody talks about.

A general-purpose chatbot, asked a question it doesn't know the answer to, will produce something plausible. That's not a bug in the usual sense — it's what the technology does. It's also completely unacceptable in a church context.

If a visitor asks your website what time the nursery opens, and the AI guesses, you have a family standing in a locked hallway on a Sunday morning holding a diaper bag. If someone asks when the last baptism was, and it invents a date, you have a pastor citing a fiction from the pulpit.

So Holdfast's AI answers only from your own records — your sermons, your events, your groups, your pages. Nothing from the open internet. And when it doesn't know, it says so plainly. Ask it about a mission trip that isn't in your records and it will tell you it can't find one, and tell you what would need to be recorded for it to answer next time.

Saying "I don't know" is a feature. It cost real engineering effort to build a system that would rather be unhelpful than wrong.

2. It cannot publish anything on its own

Nothing the AI produces goes live without a human clicking publish. Not a calendar event, not a sermon summary, not a page, not a social post.

Upload Sunday's bulletin and the importer will turn it into events for you — but they sit there waiting for your approval. Upload the sermon audio and you get drafts of notes, small-group questions, a blog post. Drafts. Somebody who was in the room decides whether they're right.

There's a version of this product that autopublishes and is technically more impressive. I don't want to build it. The church's public words should be chosen by the church.

3. It cannot make up faces for your congregation

We include image generation, because a lot of small churches genuinely have no photographer and no budget for stock. But it will produce buildings, light, and landscapes only.

It will not generate people and let you present them as your congregation. A photo of smiling strangers who don't exist, on a page inviting real people to come meet your church, is a small lie told at the front door. It's also the exact thing that makes church websites feel hollow — the stock-photo problem, now automated and infinite.

If you want pictures of your people, take pictures of your people. One real, slightly imperfect photo of your actual sanctuary beats ten perfect ones of nobody.

4. It cannot answer from the internet

The chat assistant on your site answers from your pages and your records. Not from a general knowledge base, not from what other churches say, not from whatever the model absorbed in training.

This is a narrower assistant than it could be. That narrowness is the point. When a visitor asks your church a question at eleven o'clock on a Saturday night, the answer should come from your church.

5. It cannot train anyone's model on your people

Your congregation's data — the directory, the giving records, the prayer requests, the people — is not training data. Not for us, not for anyone we work with.

And giving is never broken out by person in anything the AI can see. Volunteers get a smaller view than administrators. The people who trusted your church with their information did not consent to being anyone's dataset, and they're not going to become one because we needed a feature.

Why write this down

Because I've spent twenty years in ministry and about twenty in IT, and I know exactly how this conversation goes in an elder meeting.

Somebody proposes the new website. Somebody else says "doesn't that use AI?" And then the room divides between the person who read an article about deepfakes and the person who thinks the church is twenty years behind, and neither of them has the information to settle it.

So this page exists to be forwarded. Print it. Take it to the meeting. If somebody in your church is worried about what AI would be doing on your website, this is the honest answer, and I'd rather they read it before signing up than discover it afterward.

The thing underneath all five

Every one of these constraints comes from the same conviction, which is that a church's technology should reduce the number of things that can quietly go wrong, not increase them.

That's actually the whole reason Holdfast exists. Small churches don't get undone by dramatic failures. They get undone slowly — a service time that's been wrong since 2020, a login nobody has, a sermon archive that stopped in March. Adding a system that confidently generates plausible-sounding wrong answers into that environment would make the underlying problem worse while looking like progress.

Nobody went into ministry to retype a bulletin into a calendar. That part, the AI should absolutely do. But the moment the machine starts deciding what your church says about itself, the tool has stopped serving the church and started replacing its voice.

AI drafts. You publish. That's the deal, and it isn't going to change.

---

Every AI feature described here is part of the Enhanced plan. If you'd like to see them working before you decide anything, there's a 14-day trial with no card required — or email me and I'll walk you through it on a call.