Colin Nederkoorn, CEO and founder of Customer.io, on AI's effect on marketing automation

In June 2023, we chatted with Customer.io CEO & founder Colin Nederkoorn about the platform's newly launched CDP and its positioning amid Segment, Klaviyo, and Braze.

We recently caught up with Colin to discuss how AI is reshaping customer engagement platforms and the future of marketing automation interfaces.

Key points from our conversation via Sacra AI:

Customer.io has begun positioning as an AI-powered customer engagement platform. Break that down for us. What does it mean to be AI-powered and where have you seen the strongest AI use cases?

The overall strategy we've taken, at least initially in the first six months after the launch of ChatGPT, was a fast follower strategy. What we were seeing was people were throwing a lot of spaghetti at the wall, and it wasn't clear initially what were going to be useful applications of this new technology. We were very excited by AI, but didn't release AI features immediately when we first started to see LLMs and some vendors put AI features on the market.

When we figured out our AI approach, after doing a little bit of experimentation, what we realized was you can do a much better job when you have context about the customer in your queries. Everything gets better when you have context. OpenAI sees this in the way they're evolving ChatGPT. They can give me a better response with all of that history of everything I've ever asked them. In the same way with Customer.io, our focus initially has been on building a context layer first to enable us to be more effective at everything else that we're doing with AI for our customers.

Now in the product, there's an area where we show you what we know about your business. We show you what we think you do, which is scraped from your website most likely, and what we think your tone is, the way you talk to your customers. There are all of these characteristics about you as a business. Are you funny? Are you serious? All of these things that we can then use to inform everything else we're doing for you.

We wanted to start with the context layer. We also felt the end state was highly tailored experiences. But if you do that first, if you put all of your energy into very narrow tailored experiences, you miss so much low hanging fruit that you get just by AI enabling things—giving an LLM access to things.

The other approach that we took initially was building a lot of general purpose stuff. We made MCP broadly available. Now our customers can do things inside of the product using LLMs from the outside. We also added an assistant inside of the product. Over time, we're adding more tools to the MCP server and to the assistant. Initially, it was just access to the documentation. Then we added the ability to create segments via the assistant and via MCP. Over time, we're continuing to increase the number of tools that we give to the LLMs, both outside and inside Customer.io, to take actions on behalf of our customers.

We have all this context, so we can do a better job when we enable our assistant. Over time, we'll build more personal, curated, highly tailored experiences for specific tasks that our customers are trying to do. But we felt it was important to do all of that on a strong foundation, and not just slap a subject line generator over here and an image generator over there. Those things are both AI, but they have no idea that each other exists. If we really want to do a good job for our customer, every AI thing that we're doing in the product needs to share context with every other AI thing. Tools need to feel unified.

A lot of the excitement that vendors had was just, "Oh my gosh, this thing is now possible. Let's go build this thing." But it doesn't make sense. It's not something that's going to be there for ten years. It's something that they'll probably throw away and replace because it was a fun experiment and the novelty wore off. The big risk with building AI features is they're novelty-based features. They don't actually provide durable value for your customers.

One of the strongest use cases we've seen in the work that we've done so far is AI-powered translations to content. We can translate a message in Customer.io into over 200 different languages. When we do the translation, we're also leveraging the context about your business. If you have words that are industry specific words, it's doing a better job of avoiding a mistranslation because it also has the context of your business when we're asking for the translation.

We've seen massive adoption with that. Companies that weren't translating before are now translating into other languages. This workflow is so much simpler than when you had external services, human translators. In order to get all of your content translated, or even just a single piece, it was many steps, lots of dollars, and now it's one click and free. That's a powerful unlock for companies that operate in different countries but weren't translating in the past.

There’s this idea of AI doing the content itself. Is that important to your vision here? Is there a benefit to being able to personalize the message because you have data on customers in the same place where you’re writing?

I think about it in probably three different layers. Within a given message, there's the structure of the message and the shared components. One of the things we're doing there is with our Design Studio. We scrape our customer's website, we extract their styles, and they can configure those styles if they want to override what we think their color scheme is and their font choice and all of that stuff. From that, we can apply that to standard blocks.

If you think about a CTA with a button, there are all of these pretty typical things that companies want to do, and they want to do them in a standard way. We can generate a set of standard blocks, and customers could also prompt new blocks into creation, or into existence. When they have a specific message to create, there's the structure of that message. For example, I have a message, I want to create a win-back message. There's the structure of that message, all of these blocks with a CTA at the end. Then there's the content of the message. That's the second thing.

In certain cases, or in many cases, you probably want, at the very least, human review of the content. But the LLMs are going to be able to do a pretty good job of templating out—no more Lorem Ipsum. It's going to be, "I'm sending a message to win back an abandoned customer," so generate the structure, then generate placeholder text based on the company style guide, based on other messages we've created before. Then as a skilled marketer, I'm going to go in and review and edit that and make sure it says it the way that I want it to say it.

In addition to that, the third thing is content within that message that's personalized just for me, based on my activity stream that is in Customer.io, based on the things that I've done inside of the product that I've been using and the product that this message is about. The specific message of why you should come back to the product, in all likelihood, that part of the message is going to be AI generated based on your activity stream or based on things you've told us in the past, ways we've interacted with you.

Those three layers is what's going to create a really powerful but predictable set of messages that a business is going to send. You don't have this complete uncertainty of, "Hey, we want to win back our customers. Please make that happen." We're saying, at this point, and maybe you're testing the timing, but when people have abandoned, when we've determined people have abandoned, we want to send a win-back message. We think it should be structured like this. Here's the standard content, here's the hyper personalized content. Businesses, at least for the foreseeable future, are going to be much more comfortable when there are some guardrails around how the AI is working and the way it's generating content.

Do you see Customer.io also doing some of the more meta, growth work around segments, campaigns, triggers, A/B testing, and so on?

We expect to be involved in both the inputs into what messages get created and why, and then the outputs of how to analyze your performance, how to iterate and improve on your messages. We have all of this data that we can really help our customers.

The way to think about it, the new opportunity, is that instead of someone coming to us and saying, "Hey, our data people have done all this analysis, my boss told me we want to drive repeat purchases, or we want to drive win-back, and I have a target of 10%, up from 5%." You may tell us those things, and then we can help you drive towards that goal. Where in the past, we wouldn't know any of that information. We wouldn't have the goal that our customers have. We could just infer it from what they're doing inside of the product. For us to do a good job for them in the future, we need that context of what's the business outcome you're trying to drive. If we have that, our LLMs can be much more effective in helping you get there, whether in an advisory role or in an execution role.

Is there a vision you also have of the whole AI agent as your employee type paradigm where it's your point of contact or it's coming up with the new ideas and stuff like that, executing autonomously on things. Do you think about that?

I do think about this a lot. I do think we are heading that way, but it's not as fast. It will be technically able to do it well before that's the way many businesses are working. Right now, our world is very deterministic. Businesses are comfortable with the determinism because they're setting these things up. Part of Customer.io's success over the years has been, when you set campaigns up with very specific rules, we will execute them predictably. The behavior you expect to happen in our system will happen without fail. If we do something weird, if a message doesn't send, that's a breach of our relationship with our customer.

If you, given today's technology, if you set an agent loose, they would do all sorts of things for which you would fire that human employee, and their manager would be getting in trouble. There are all of these structures within organizations to stop unpredictable stuff from happening. An agent today is very unpredictable.

You think about things like self-driving. Most people are still very uncomfortable with self-driving cars. Even though they are by all accounts, and they have been for a few years now, safer than human drivers. But they still, every once in a while, do something unpredictable. You take over, and you're thinking, that was a little odd.

The same thing, we're going to see that with marketing agents, especially when they're operating in a world that has different variability from self-driving. The rules are a little different. Companies really care about their reputations. They're not going to want to just delegate them to a marketing agent that may end up emailing 100% of their user base with “LOL, click here, bro”—because it was determined that would drive more conversions or something. You're just not going to want that really unpredictable thing to happen.

To get back to something I said earlier, you want to be selective with where you use these autonomous agents, and you want to have really strict guardrails. Maybe it's the agent deciding between these four paths based on what this user has done, which one should I send them down? Or maybe it's deciding, of these five CTAs that I'm testing, which one should this user get? At first, it's going to look like that. Over time, as we all gain confidence and the agents get better, and the supporting infrastructure gets better, we're going to untether them a little more and more.

What can (or will) a Customer.io customer who uses Journeys, Data Pipelines and Design Studio be able to do with AI that they won't be able to do if they only use Journeys individually? How does value compound via AI if you vertically integrate your stack into Customer.io?

One interesting thing compared to when we spoke before, we've actually pulled back from multi-product. We went multi-product and realized that, one, it was more complex to sell. And these products were all very close and adjacent to each other. Our customers were coming to us primarily for the value that Journeys provides. We decided to fold all of the products together. Now we sell the Customer.io Platform, but you can no longer buy data pipelines separately from Journeys. It's all meant to work together.

The benefit here is that you have that full customer context. If you're using a different product that doesn't have the CDP capabilities, there's been a move that companies put everything in a data warehouse and only give certain amounts of data to the tools that are downstream from their data warehouse. The value of having data pipelines or our CDP functionality in the mix is that Journeys has full context and can act on that customer data. Our customers don't need to work with their data teams to make sure that we've got the right thing in Journeys.

We continue to expand our integrations with pipelines. Salesforce has been one that we've made very robust over the past year or two, both getting data in and getting data out. Hubspot as well. Giving customers a lot of the capability to get data where they need it via Customer.io means that Journeys has the ability to act on that data with very sophisticated logic. It puts us in a better position to serve our customers.

To maybe answer your question, the AI can be as good as the context is. By having the ability to help our customers get data from A to B via our system, it makes us better positioned to have the context to power really good AI outcomes for them.

Customer.io has historically won by appealing to the "technical marketer"—someone comfortable with data, logic, and maybe a little Liquid code. There’s a theory that AI lowers the floor for technical skills, making everyone a "developer." Does AI expand your TAM to non-technical marketers who can now prompt their way through complexity, or does it simply make your core technical user base 10x faster?

A couple of years ago, our goal was to redesign big parts of the product to make it more accessible to less technical users. Or maybe it was three years ago now. Pre the realization that LLMs are incredible at gobbling up complexity and giving you understanding and clarity. Our solution to the sophistication and the power in our product was to use design as a way to make that sophistication more approachable .

Our strategy shifted. We were not that successful at doing it through that approach. Partly because our most demanding customers were always asking for more capabilities. As we were trying to redesign and simplify user interfaces, at the same time we're adding more bells and whistles over here and expanding the capabilities of the platform. We're adding pipelines. We've got Parcel being integrated now. We're just getting more sophisticated. We're getting sophisticated at a faster rate than we're using traditional methods to make that sophistication more approachable.

The opportunity that we have with LLMs and the shift that we made is we're not eliminating the sophistication, but we're using LLMs to make it more accessible to a less technical audience. Part of that is our initial approach to that is the assistant, where on any screen in the application, you can ask the assistant questions. You can say, "What does this mean?" It has access to our documentation. In the same way when I'm doing a home repair, I'm using ChatGPT the entire time to be asking, "Wait, do I turn my plumbing wrench to the right or the left? What tightens it?" The assistant is a really powerful, experienced marketer that you have right next to you as you're trying to set stuff up in Customer.io.

The other more tailored experiences that we're building leverage LLMs to make some of these more complex configurations more possible out of the box. There's all of the generation stuff—generating HTML that works well with Outlook or whatever it might be. All of this stuff is now possible where it required expertise or required you to phone a friend when you were attempting to do that in the product.

Our approach now is our sophistication is our strength, and LLMs will make that more accessible to everyone.

A common criticism of generative AI in marketing is that it reduces the marginal cost of creating content to zero, which inevitably leads to more noise and spam. If Customer.io makes it effortless to generate 100 personalized variations of an email, how do you prevent your platform from becoming a weapon of mass distraction that degrades the very channels (email, SMS) you rely on?

I do worry about it. It's always been in our terms of service that the messages that our customers are sending are opt-in. We're fortunately insulated a little bit in that the risk is that we, as the broader business community, over-create a spam cannon on every messaging channel possible, and people just do not want to hear from any businesses ever. That's the worst case outcome, a tragedy of the commons, where everyone abuses it just a little bit, and as a result, nobody wants emails. Nobody wants texts.

For our customers, the opportunity is that we can, through the use of AI, help them understand the value of the work that they're doing in Customer.io. By having better understanding of the data, they stop doing the stuff that's ineffective. We reduce spam cannon behavior on our platform. By leveraging generated content in the messages, they can improve the desirability of what they're sending to people. People want the messages they're getting from those companies. They get more messages they want, fewer that they don't. By tying it to real business value, we're able to deliver a better experience both for the business and for the customer.

I've always had an optimistic view that when someone signs up for a website or an app, they're coming to that app to solve a problem. It's your responsibility to help them solve that problem. If you send them an email that is furthering that goal, you're fulfilling your obligation to this person who came to you to solve a problem. That's okay. That's not spam. It's when you deviate from that, or you lose their trust, or whatever happens in that relationship along the way that they no longer are interested to talk to you and you keep going.

We have the opportunity to be better about that as businesses now. My optimistic view is that it actually gets better, not worse.

Usage has been a powerful driver behind Customer.io's strong revenue expansion dynamics which indexes on customer profile and email send volume. You don't yet have a usage metric for AI usage. How do you intend to incorporate AI usage into your usage-based pricing model?

Our pricing has already been reasonably well aligned with our customer success. We only want to win when our customers win. The way we've priced historically has been profiles. If our customers' business is growing, then their Customer.io bill goes up.

We're always experimenting with things. We're running an experiment on charging for monthly active users instead of total profiles. For some businesses, that makes a lot of sense. When we looked at should we charge for these AI features that we're implementing now, the distinction that we make here is, if it's AI that helps someone use our product more effectively, then we don't charge for it. Or if it's AI that basically increases usage of our product.

Think about AI translation, segment generation, the assistant, MCP access. We're not charging even though, MCP doesn't really cost us on the LLM side, but those other things do. They have a real cost. We're paying for something where we weren't before in order to deliver service to our customers. But we're absorbing those costs because that drives value for customers, which increases LTV. It increases their usage of the platform.

When we look at AI to deliver customization and personalization, we ran some analysis and if every single message sent through Customer.io—this was a couple of generations of models ago, so the pricing has come down a fair amount—but it was something like $100 million a year in token costs to personalize every single message sent from Customer.io. That's a real cost to us if customers decide to have a campaign that goes to 10 million people, and they want to personalize the content using an LLM query for every message. In that case, we're looking at usage-based pricing for that type of personalization. That will come down over time.

In the places where we can generate additional value via LLMs or increase performance via LLMs, we're exploring new additional pricing. Then in the places where it's just about using the product more, we're absorbing that. But in our business, I don't see in the short to medium term pressure to fundamentally change our unit of value for customers. The per profile or per MAU, it's pretty reasonably aligned regardless of their team size, regardless of all of these other changes.

I'm a one person company, and I'm $100 million ARR or whatever. If that's your situation, then offloading this part of your business to Customer.io, which is typically not the core—or to maybe take this a different direction—you could get something working with Claude Code, but you then are putting it into production.

You've got to run this infrastructure in production. You're iterating on it over time. You've got to report on the performance. Maybe you hire one other person. Maybe you want agents to do it. We're just going to keep improving the experience whether it's for humans or for agents. We're going to take this thing that's not core to your value for customers, but is critical for your business, and we're going to keep making that better over time.

At a very basic level, it's possible to generate a triggered email with Claude Code and deploy it in your app. But in the long run, this alignment with our success and your success, charging for profiles or MAUs, is very well aligned with our customers, and I'm glad I'm not in a per seat model.

On the deliverability front, it'll be interesting if there's a day where Verizon's postmaster is just an agent, and when you're having deliverability problems with Verizon's postmaster, you reach out to their agent, not a person. But today, if your job is deliverability for someone like Customer.io, you're maintaining relationships with humans at these big ISPs in order to solve problems that come up when their automated systems incorrectly flag incoming messages and are causing deliverability problems. You sort that out with human to human relationships, and until that changes, it's important. As a business, you want to get your messages to your customers.

How do you describe your long-term vision for how AI, its role and how it will benefit your customers? Is it largely a shared vision across companies like Braze, Iterable and Klaviyo and the differentiation is in the execution, or is there a difference of opinion on the vision and the role of AI?

I think about what's going to stay the same and what's going to be different. What stays the same is that we're still going to be focused on serving product-led companies who care about engaging with their customers—sending good messages to their customers. We're still going to prioritize those customer outcomes, those positive customer outcomes over vanity metrics. So it's conversions, not opens or clicks or any of these other vanity metrics. We still believe that behavioral data, what your customers are doing inside of the product, is great signal to send targeted messages. We still believe in the value of having an extremely flexible platform that can be set up to do whatever you need it to do.

But then what's going to be different? The interface, the way that people interact with software like ours—today, our workflows are these very logical big trees of, you know, if this happens, then do this thing. Those interfaces are going to be much more prompt-based, or conversational, where they're going to be generated by people expressing their ideas, and then validating that the output seems right. They're going to be much less deterministic, and look much less like decision trees.

The rate at which companies are testing different content is just going to increase dramatically. The level of personalization and testing is going to be huge. What today is this cycle of, "Oh, well, we ran this campaign. Here's the feedback we got. How are we going to tune it? What do we want to test next?" That takes weeks in companies to go through a full cycle of trying to figure out what to change. That's going to be a continual process where it's more similar to the way the ad platforms work now where there are some boundary conditions, but they're continually testing and iterating to find what's the optimal path. That's what it's going to look more like in platforms like ours as well, or that's what the behavior is going to be like. Continuous optimization, not a discrete project that goes end to end.

What I don't know is whether people are going to use us more via our API, and we'll just be an API MCP layer that's part of their conversation agent that they're running all of their marketing activities or all of their business activities from. Or if people will spend more time in our UI. That's something I'm curious to see how that evolves. We're well positioned to do either.

People are going to be even more focused than they are today on understanding what is the business outcome that's generated by the tool that I'm using. Because of all the data flowing through, it's going to be easier to tie specific campaigns or specific messages to particular business outcomes. Attribution, which has been, for my entire career, so squishy—I think people are going to let the AI figure out attribution and trust what it comes out with at this point. We're going to be making decisions more around, how many dollars does this path drive versus this path? It's going to feel much more—I don't know if that'll make it more cold and more sciency and less of an art. But it's overall a good thing to be able to get closer to dollars and away from, "Oh, but this one got more opens, and this one got more clicks."

As far as channels, people have been calling email dead for a while. Since we started the company, people have been saying, "Oh, email's done." But we haven't seen email be totally replaced. We've seen additional channels become important. Email is still important in 2030. There's probably more one-to-one text channels or text-like, or video—you can imagine video interactions with a business where you record your question and an AI video agent responds. We'll see more of whatever rich looks like in five years from now. Could be VR, augmented reality. Those interactions are going to feel more conversational, and people will get very comfortable talking to AI agents that represent companies.

The through line is: the things that we believed when we started the business are still very much true. It's behavioral data. You're using it to drive or to create better customer experiences and drive shared desirable outcomes for both you and your customers. AI just makes that way more achievable than it ever has been to date.

As far as are we thinking about this differently from our competitors? The companies are all seeing these same opportunities. I don't know that the way I talk about the future will be fundamentally different than the way Braze or Iterable or Klaviyo might see the future. Obviously, Klaviyo is much more focused on e-commerce than we are. Some of the things that they'll build will be very focused on e-comm.

I do think that this is about execution. It's about identifying—we're all talking about, "Oh, we're going to build a more personal experience for your customers. You're going to be able to use data to hyper personalize, to target." Everyone's going to have really similar messaging. But if we can deliver on it by leveraging this context layer, by making it so that when you go in, we can get beyond the tech demo. We can actually deliver the experience that we show you in a video, and you experience that when you go in and use the product. That's going to come through execution and your pace at which you can take feedback from your experiments and feedback from customers and adopt new developments in LLMs, the way you can roll that all into the product.

If we're doing it right, we're building on top of this foundation where we don't need to reinvent the foundation, and the features that we build all relate to each other. In a feature checklist world, they may look similar, but it's the unification of all these features that we can provide that will be the differentiator. It may require touching the product to see that. But that's the way the world is going anyways, that PLG is very much the past and the present and the future.

I assume an email developer might be in Cursor. An email marketer might be in the web UI of Customer.io. But also, what is that role? What's the 2030—is there still two different roles in 2030, or is it really one thing?

With very few exceptions, the email developer role disappears by 2030. There's code blocks, but there are the standard blocks that a company has. Here's my block for a CTA. Here's my image on the right, text on the left. What does that look like? There are all of those standard things. But that's both generate—there's 10 of these that are pretty normal for any company to have, and then there's your custom ones. The custom ones should be able to be prompt generated and tuned.

What's the tuning you're going to do? Either you're going to prompt it to say, "Actually, I need more padding, but I don't know how to do that," or "I need a yellow border around this, but I don't know how to do that. Can you do it for me?" And the LLM can do it for you. Or you're configuring in the UI, you're making some small tweaks by yourself. But very rarely are you going to need to be in the code to accomplish it. I do think the email developer goes away.

When you have all of these blocks, it's really either you're dragging and dropping the blocks around in the message, or your LLM is generating for you guys, where you have different articles or different news snippets. What I imagine is there's the content you want to insert, and there's a notion of these are discrete news snippets, and the LLM will know how to insert the news snippets and what blocks to use. Even though there's code underlying, you should never be limited by not understanding the code. You should never really need to see it.

This transcript is for information purposes only and does not constitute advice of any type or trade recommendation and should not form the basis of any investment decision. Sacra accepts no liability for the transcript or for any errors, omissions or inaccuracies in respect of it. The views of the experts expressed in the transcript are those of the experts and they are not endorsed by, nor do they represent the opinion of Sacra. Sacra reserves all copyright, intellectual property rights in the transcript. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any transcript is strictly prohibited.