The “Christian AI” market is getting crowded.
At Gloo, users are “getting more personal” with their chatbot “than with most pastors.” Pray.com’s AI-generated Bible videos are styled like video games and replete with dramatic monsters and sexy ladies.
New to the scene is Creed, which bills itself as a “digital companion for churches and believers.” CT spoke over Zoom with Creed’s CEO, Adi Agrawal, about the app’s tech, business model, user base, and goals.
This interview has been edited and condensed.
First, to make sure our readers have a good sense of what Creed does and how you do it, tell me about your tech stack and training. What large language model are you using as the base for the app, and what are the specific resources that you use to train it?
Think of the product as a Tamagotchi meets Duolingo for a Christian.
Think of it as a best friend and youth pastor in your pocket that you can talk to about any questions, be it your personal life, be it faith questions, be it your Bible studies. It’ll give you answers rooted in Scripture, but it’ll also build on that relationship over time. It remembers things you tell it. The more you talk to it, the more personalized it gets. It develops personalized faith paths for you, and it’ll help you answer questions and find community around you so you can discover Christian events. If you’re not affiliated with a church, it’ll recommend churches for you to go to. So it’s not just talking to a screen.
On the back end, we’re not building our own models. We’re using off-the-shelf models from OpenAI, Google, etc. But we noticed that everyone’s using ChatGPT, and people are asking it super personal questions like “Should I get a divorce? Was I the wrong one in this situation?”
ChatGPT will give you an answer, but whose values determine that answer? Who is determining what those values are? That’s a little bit of a black box, but the values are determined by a few companies in Silicon Valley.
So how do you set guardrails around that? How do you make sure the answers you get back from the AI are something your pastor or your church leader or your parents would agree with?
We take off-the-shelf models and fine-tune them on Scriptures—30 different versions of the Bible. And then on top, a few nonbiblical texts as well, like the texts of C. S. Lewis or a few other, broader Christian authors.
Then we also use denomination-specific teachings. Different denominations have pastors and priests whose teachings are widely accepted, so we use a few of those. Our offshore teams annotate those texts and label them to ensure some level of denomination-specific accuracy. For the same question, what you might expect as a Southern Baptist is going to be different from what you expect as a Catholic.
Then the third layer is church-specific nuances. We work with churches in our partner network directly so they can go set their values on “topic X” on top of that denomination-specific nuance. Right now we only have 12 or so churches in our network, and we’re building that out.
And then we set very, very strict guardrails. One of the big issues with AI is hallucination. How do you ensure it’s giving you actual answers? How do you ensure it’s not just making things up? So every answer passes through three filters of checks.
And any kind of sensitive topic—say, transgender rights or gay marriage—we are not going to give you like, “Oh, you should do this” or “You should not do this.” We’ll cite Scripture, and then we’ll tell you to go talk to your pastor or go talk to your parents. A very middle-ground approach where we as a company are not being prescriptive. We want to be very sort of value-neutral, and we want you to go talk to actual trusted authorities.
If there’s any mention of self-harm or harm to others, we have a human who will intervene and will be like, “Oh, do you want me to call up your parents?” or “Do you want me to call up this hotline for mental health issues?” We want that human intervention layer in case there’s something that’s super alarming.
You’ve mentioned parents, and that plus your description of Creed as an AI youth pastor makes me think your target audience is teenagers. Is that accurate? Can you give me some sense of your typical or ideal user demographics?
Yeah, initially we went to market with a focus on Gen Z. It’s a bit surprising though: About 60 percent of our users are 15- to 30-year-olds. Then I’d say the fastest growing demographic is 50- to 70-year-old women. That’s another 30 percent, and then 10 percent is 30- to 50-year-olds, a mix of males and females.
We found unexpected success with demos other than Gen Z, so now we’re trying to make it more age- and gender-neutral in terms of the design and the tone of the AI’s answers. If you use the app today, the way it talks is very Gen Z. But we’re implementing a new voice where it’ll essentially reflect back your own personality and your own voice. It’ll adapt to you. If you use emojis, it’s going to use emojis. If you use full sentences with capitalization, it’ll like that.
You said in your fact sheet that you had 200,000 users in the first four months. Do you have a number for daily active usership yet?
Daily active users are around 25,000.
Okay. And then turning toward the business side, you mentioned in your fact sheet that you got funding from Andreessen Horowitz Speedrun. Is your cohort with them ongoing? And I’m also curious about how much funding you were awarded.
We started Speedrun in July of this year. When you start, you get $500,000 from them. We finished Speedrun in October, and then after we finished, we raised $4.2 million.
That is speedy. Accurately named! I would guess that as you’re working with these venture capital folks, you’ve presented a business model for what to do after that initial funding. What is the model going forward?
It’s very similar to Duolingo. It’s monthly and annual subscriptions on the consumer end. And then we also offer in-app purchases in addition to the subscription. So think of e-commerce, or very similar to gaming [with] in-app purchases.
I haven’t used Duolingo in a long time, so I’m not sure: Is there any advertising? Is there sale of user data, or is it just those direct user purchases?
Just those [direct purchases]. We were going back and forth on whether we should show ads. There are pros and cons. If you show ads, you could potentially make more money, but it really ruins the user experience if you’re chatting about deeply personal issues and suddenly seeing ads. People start to question where their data is going. So we’re going to hold off on ads and purely go with subscription revenue for now.
Turning from tech and business, I’m interested in your goals for Creed. I think the single most surprising thing in the fact sheet you sent was this sentence: “If you tell your companion that you are feeling sad, it will pray for you.” For me, that raises the question of what your team understands prayer to be. What do you think it is and how you think it works?
When it says the companion will pray for you, it’s more like it’ll pray with you. It’s not going to pray for you. It’s more like, “Oh, do you want to pray with me? And how are you feeling?”It’s almost like generating personalized prayers for you that fit your mood and fit the way you like to be prayed with.
I think the beauty of AI is it’s not pushing something at you. Rather, it’s working with you to build something in collaboration. So this prayer would be very customized to how you’re feeling and the way you like to pray. You can tell the companion, “This is how my church or my pastor taught me how to pray,” or you can give it a YouTube-style sample prayer.
There’s no simple, one-size-fits-all approach to this. How you need to pray can look very different from person to person.
Just to make sure I understand, you’re saying it would write a prayer that the user would say as opposed to the chatbot praying in its own voice?
Yeah, this would write a prayer, and it’ll basically speak the prayer out loud with the user. So it’s almost like a friend would do it. Like you’d ask a friend, “Hey, can you pray for me?”And then you both would pray that together—
Speaking in unison.
Yeah, speaking in unison. Then there’s this other feature we just launched. It’s almost like a prayer wall where you can have other users pray for you. Say you’re in a remote town in Tennessee, and you don’t have a church community or a lot of friends around there, and your mother’s sick, and you want prayer. Other users will join you in that prayer. We launched that two days ago, and we’ve already seen a good adoption of that feature.
I read an article in which a pastor was thinking about this kind of product and specifically about prayer. He made a point that I’ve not seen elsewhere, and I’m interested in your response to it. He said the difference, perhaps, is that one of these prayers is by a soul saved by Christ and the other is written by a program; and, whatever the implications of that, which might God prefer? Does God want to receive a prayer written by a machine?
That’s a very valid point, and you could go philosophically down that whole rabbit hole, but I think our end goal is helping people out in times of duress.
And sure, a machine-made prayer is definitely not up to the standards of a prayer written by an actual human. But in that situation, if that machine didn’t write a prayer, that prayer would not have happened. So we’d much rather have a prayer, even if it’s of a slightly lower quality, than no prayer at all.
I think that’s how I think about it, but I think that’s a very valid point in terms of what God prefers. God would probably prefer a human-quality prayer, but it’s hard to scale that service to 3 billion Christians worldwide.
Of course. I think maybe the alternative would not be that they would turn to a custom human-written prayer, but that they would just pray.
Half of our users are folks who very recently turned to God and Christianity. We’ve noticed they use this almost as an introduction to exploring more about the body of Christ and reading the Bible. A lot of them are not familiar with that whole process of praying or what’s the right way to pray—or even getting the words into their minds, in terms of how you compose that prayer. So we don’t want to supplant them praying for themselves; rather we want to build that habit into them.
That’s good transition to another question I had. I was struck by the fact sheet’s mention of “Duolingo-style daily quests such as daily prayers/devotionals—the more of these quests that you complete, the more points you accrue, opening up richer devotionals, faith milestones, and tailored guidance.”
And I was baffled by this idea of gamifying the practice of faith and then apparently—correct me if I’m wrong here—withholding the best discipleship materials until people level up or maintain a streak, like in Duolingo. If we’re dealing with new Christians, people who really need help, wouldn’t we give them the richest devotionals off the bat?
We’re not withholding devotionals. The way it works is the more points you get, the more sort of rewards you win. These rewards aren’t discipleship frameworks. It’s more like you can customize your character. You can open up new characters. You can buy a hat for your characters. It’s more those sorts of rewards versus actual texts or devotionals. Those are open to everyone, starting day one.
The characters—you sent images of users holding the screen with what looks like a sheep. Is the sheep—and maybe later other characters—what you all had in mind when the fact sheet described Creed as “embodied” with a voice, personality, and memories?
Yeah. A lot of chatbots—like ChatGPT, Claude, all of them—are abstract chat forms. But our whole thing was: Can you make this more embodied? Can you give it a personality with a voice, proactively following up with users?
That is what we mean by “embodied” versus the more abstract chatbot forms. This idea of embodied companions has been one of the biggest trends in consumer AI over the past couple of years. Most of these embodied AI companies are AI boyfriend or AI girlfriend apps.
Just to clarify, you don’t mean a physical piece of hardware, like Friend. You mean the image on the screen?
Yeah, this is on the screen. This is not a physical thing.
So these other embodied AIs are sexualized, weird, romanticized companions. They’re all millions in revenue, but they’re so toxic, and it’s so unhealthy for the future—not just the future of Christianity, but also the future of our younger generations, that people are spending hours per day having weird, sexualized interactions with these AI boyfriends and girlfriends.
Our initial motivation was to appropriate that same technology but use it to get people closer to God, learn more about the Bible, get them out and discover community instead of having weird interactions.
Yeah, one thing I appreciate in the fact sheet was your policy of cutting people off after an hour of use. That that is a real distinctive you don’t see with a lot of the big names in AI.
The more time you spend, the more money they make. It is a very unhealthy cycle.
Right. That’s a great caution to have. I do wonder about blurring the lines by using “embodied” this way, because having an avatar isn’t usually what we mean when we say “embodied.” It’s a picture instead of text, but it is on a screen still, right? It’s not embodied.
Yeah, I agree with that interpretation. It’s not truly embodied, I guess.
Okay. As I was preparing for this, I was thinking about when Paul writes to the Galatians, and he’s very worried about them. He says he’s in the pains of childbirth until Christ is formed in them, because he’s their pastor and he loves them.
When I think about a young, lonely Christian who maybe has moved to a new city, is struggling to find friends, struggling to get plugged into a local church community—when I think about that person being told that this program is going to be their source for pastoral care, it honestly makes me sad.
And it strikes me as a pretty dark vision for the future of Christianity and the future of the local church. I do appreciate your sense of the importance of pushing people to go find a local congregation, pushing people to get off the screen. I want to affirm that. But on the other hand, it seems to me that merely by offering pastoral care with no pastor, you’re kind of giving up on the local church at a level so fundamental that those tweaks may not matter.
At the very outset, this is not a replacement for pastoral care. And when you download the app, one of the first things you’ll see is “Find church events near me,” “Find a church near me.”
We very much positioned this as a “Christian best friend” app. We are not that pastoral authority. Best case, we might be a youth pastor, but we’re definitely not any kind of authority. We want to push you to actual resources.
And we’re acting very much as lead generator for local churches. As a company, and for me personally, local churches are a backbone of our future, and they need to be revived. Obviously they’re going through a very challenging time now, and we want this app to pull people off TikTok and introduce them to the faith, introduce them to local churches, help provide local pastoral care. I think that’s the funneling role we see our app playing.
I assume everything’s very data-driven. You have the data to make sure things are working the way as intended. If at a certain point—a year, five years down the line—you are not reliably getting data that people are getting off the internet and becoming meaningfully involved—joining, volunteering—in a local church community, what do you do?
If at any point we notice we’re trying to funnel them to actual, real-life engagement, and they’re not doing that, I think that means we failed in our mission as founders and building this company.
And I think at that point, we’d probably want to pivot the business and do something else entirely. That mission is just so central to why we started doing this. And if we fail in that, I think we failed as founders of this company—and at that point it would be better to even shut this down than to get people addicted to some phone game and not go out into the actual world.
