AI Agents Explained: The Future Of Customer Experience — Dom Steil | Why Scaling Brands Need AI, Why AI Agents Make Support Proactive, How iCommerce Powers DTC Growth, What Sets CX AI Agents Apart From Chatbots, How AI Links Shopify, 3PL, And ERP (#437)

In this episode, we dive into how AI agents are changing customer support (CX) and operations for DTC brands.

Dom Steil, CEO and Founder of StateSet, shares how his platform automates complex tasks like warranty replacements and subscription changes while keeping a personal touch.

He explains the difference between AI agents and chatbots, how to maintain data safety and security, and why brands in hyper-scaling mode should adopt this technology to handle increased ticket volume.

Topics discussed in this episode:  

• How AI agents automate complex customer outcomes.
• What iCommerce (intelligent commerce) means for DTC.
• Why speed and 24/7 service builds customer trust.
• How AI agents save customers from subscription cancellation.
• What integration with DTC tech stacks looks like.
• How multi-language support works for international brands.
• What guardrails prevent AI from making errors or "hallucinating".
• Why hyper-scaling brands need AI to avoid scaling OPEX.
• What the future of AI-to-AI agent communication is.
• How to prepare operations for BFCM volume surge.

Links & Resources 

Website: https://www.stateset.com/
LinkedIn: https://linkedin.com/in/domsteil
X/Twitter: https://x.com/domsteil

Get access to more free resources by visiting the show notes at https://tinyurl.com/55m5sdds

______________________________________________________

LOVE THE SHOW? HERE ARE THE NEXT STEPS!

Follow the podcast to get every bonus episode. Tap follow now and don’t miss out!

Rate & Review: Help others discover the show by rating the show on Apple Podcasts at https://tinyurl.com/ecb-apple-podcasts

Join our Free Newsletter: https://newsletter.ecommercecoffeebreak.com/

Support The Show On Patreon: https://www.patreon.com/EcommerceCoffeeBreak

Partner with us: https://ecommercecoffeebreak.com/podcast-sponsorship/

Transcript

00;00;00;01 - 00;00;24;23
Speaker 1
AI agents now can actually perform the outcome on behalf of the customer autonomously in the background. Now, with our agents, we can automatically place that order in the downstream systems and provide that confirmation back to the customer almost instantly. We have kind of passed the Turing test. Customers don't even know that they're interacting with an AI. One of our best use cases, we actually save customers from canceling builds trust with the brand.

00;00;24;23 - 00;00;25;23
Speaker 1
And the ROI is great.

00;00;25;24 - 00;00;51;22
Speaker 2
Will be there a time where AI agents communicate with AI agents on the customer side, like so? Hello and welcome to another episode of the eCommerce Coffee Break podcast. Customer support has always been one of the toughest parts of running a DTC brand. It's costly, hard to scale, and often leave shoppers waiting too long. But imagine if you could automated order changes, returns, even subscriptions, while still keeping that personal touch.

00;00;51;24 - 00;01;09;21
Speaker 2
That's exactly where AI agents come in. And to learn more about this on how it works, I'm joined by Dom Steil. He's the founder and CEO of stateset.com, a platform that helps growing Shopify brands to use AI to run operations smoother. So we have a lot to cover. Let's dive into it. Hi, Dom, how are you today?

00;01;09;24 - 00;01;11;18
Speaker 1
Doing great. How about yourself?

00;01;11;20 - 00;01;21;06
Speaker 2
Very well done. A lot of brands see customer support just as a cost, and most brands struggle most with that. How can I fix it?

00;01;21;08 - 00;01;41;03
Speaker 1
So I look at it as kind of the pre transformer era and post transformer era. So in 2023 with like the advent of GPT four, this kind of thing got dropped into the world. And from a technology standpoint, I think it really changed the way we look at how software should be built and how it's delivered, and how it can create different customer experiences.

00;01;41;10 - 00;02;05;14
Speaker 1
I think prior to that, the best practice was you set up your systems. You set up some different middlewares or trying to integrate them. But really there's kind of that missing customer facing orchestration side of things where you're actually having the solver perform some outcome on behalf of the customer. And I think that's what changed, where agents now can actually perform the outcome on behalf of the customer, just kind of autonomously in the background.

00;02;05;16 - 00;02;22;09
Speaker 1
Now, how that's accomplished. You know, you know, it takes it takes time. It takes effort to essentially do that sort of, the upfront work to get it right. But once you have it in place, I think the benefits and ROI is, really great even today. And it will continue to improve over the next couple of years.

00;02;22;12 - 00;02;28;22
Speaker 2
I found on your website the term ecommerce intelligent commerce. What does that mean? What does it involve?

00;02;28;25 - 00;02;47;17
Speaker 1
Yeah. So I think ecommerce is intelligent commerce. I think what we're looking at is a new category of software that exists in this new transformer world. And what it means is the software is integrated, it's intelligent, it's instantly connected to the different systems you're using, and it's iterative in that it's consistently getting better and better as the agents are learning more and more.

00;02;47;17 - 00;03;01;00
Speaker 1
So I think this is this new category that many companies are going to be looking at. How do we kind of build this intelligent system, this OS that's powering our business, something that maybe wouldn't have been able to be possible prior to that.

00;03;01;02 - 00;03;14;19
Speaker 2
Now, a lot of our listeners might have heard the term AI agents, but they might think about chat bots left in the past. But it's a completely different story. You talk me through a use case. What AI agents actually can do in a DTC business.

00;03;14;22 - 00;03;33;13
Speaker 1
So one of our customers, they're using our agents for doing, warranty replacements. So prior to this, if maybe a chatbot would, a customer writes in, hey, I got my order. It's, you know, it broke. Can I do a replacement? And there would be a chat back like, yeah, let me, you know, send you this link or. Hey, what for this someone on our team.

00;03;33;13 - 00;03;54;07
Speaker 1
And then they have to manually go in and go into three PL onto the order. Make sure the order gets into the ERP. Now with our agents we can actually take the intent of this is a warranty replacement intent. We can look up the skew either based on the data that they've provided or an image of the product, or even, their previous order history.

00;03;54;13 - 00;04;13;22
Speaker 1
And then we can automatically place that order in the downstream systems and provide that confirmation back to the customer. So this is just one example where it's gone from just kind of being conversational and like a chat link to something where it actually can perform, taking action in the various systems and provide that information back to the customer almost instantly.

00;04;13;27 - 00;04;30;10
Speaker 1
So it saves the team time internally, especially over the weekend. If they come in Monday morning, they don't have, you know, a hundred warranties that they have to manually process. It's completely streamlined. And, you know, maybe for the couple of outliers, if it's not sure what the correct skew is, it can actually escalate it with different logic.

00;04;30;13 - 00;04;50;08
Speaker 2
But it definitely helps a lot. But if you want to scale or really, this optimization takes a lot of work from a manual, employee to, from manual work to, to the I know a lot of brands or most brands have very, very different tech stacks. How does an AI agent connect to all these different technologies that are running in the background?

00;04;50;11 - 00;05;06;26
Speaker 1
So I think what we've done at the company is we've really focused on kind of direct consumer stock. So like the main subscription apps that are in the ecosystem, the main three plus or kind of WMR systems, and even on the storefront side, we've made sure that we've integrated with kind of the key platforms there.

00;05;07;03 - 00;05;40;12
Speaker 1
I think there's different methodologies today. We're seeing like MXGp protocol as a way to give agents, kind of context that they need, or even connecting to it just via, tool calling, an API calling, which I think was a key innovation over the last year and a half where being able to actually perform an action, make an API call, connect to Shopify, or recharge whatever the app is, and either check it to see the status of been order or to actually make a state change in that system, canceling the order, updating the status of subscription, or even moving a subscription out.

00;05;40;14 - 00;05;46;20
Speaker 1
You know, a few months of the customer doesn't want to cancel. Those are some of the ways that we're looking at, integrating the different systems o.

00;05;46;20 - 00;05;58;15
Speaker 2
Brands, brand owners, founders that have not used AI agents yet, that might have a fear that they're losing the human touch. How will I agents make sure that it still feels personal and on brand?

00;05;58;17 - 00;06;19;02
Speaker 1
I think it does today. I think it's almost an interesting even more macro potentially impact is like we have kind of passed the Turing test, quote unquote, and that a lot of the times the customers don't really even know that they're interacting with an AI. We have other customers that will say, hey, this is like automated assistant at the brand name.

00;06;19;04 - 00;06;40;24
Speaker 1
We're here to help. And if it does accomplish the outcome that the customer wants, I don't see that they completely. I think it once it goes beyond a certain number of messages, usually we do have rules in place where this is probably something more nuanced. And we do want to escalate it to a team member. So they still feel like they're not just in this loop where, they're not getting, what they, what they need done.

00;06;40;26 - 00;07;09;28
Speaker 1
But I think it's going to be a combination of basically, you'll have your AI agents for things like repetitive tasks like subscription modifications, warranty replacements, order cancellations, and even some of the more infrastructure side of things like orchestrating orders between different systems. I think that's where a lot of the benefit we're seeing today. And then for more nuanced things, but just require time, or the AI doesn't even have the API call, you know, in its available list of tools that will always still need to be escalated to, to human.

00;07;10;00 - 00;07;30;08
Speaker 2
Brings me to the next ideal topic here is data security. Data safety. Obviously was connecting a lot of systems and then having AI in there. Some of the stats might say I'm not safe about that and I'm not feeling good about it. Specifically, as people might have heard about, I had to say anything. How do you make sure that all the data transfer that's happening there is safe?

00;07;30;14 - 00;07;47;05
Speaker 1
So there's a couple ways we look at this. The first thing is creating the right guardrails and rules around what the AI can and can answer to. For a lot of customers like, you know, not answering medical is, very important to kind of, you know, table stakes thing, especially if they're like a health and wellness brand or beauty brand.

00;07;47;07 - 00;08;08;22
Speaker 1
Having the right guardrails in place, even using more traditional software methods like regex and being able to check, like, you know, for certain keywords like reaction or, you know, you know, frothing or even if for escalation policies like frustrated. Having that in place I think is table stakes. Without it, you're not able to go live, especially for customers that are more at health and wellness side of things.

00;08;08;25 - 00;08;29;09
Speaker 1
I think the other thing is, creating the right evals. So creating evals around, like, can we grade the response and see does it pass all of our checks around like politeness or, does it pass all of our rules around the correct tone so you can run those evals when the response is being created. And if it doesn't pass the correct amount of evals, it doesn't get sent to the customer.

00;08;29;14 - 00;08;47;03
Speaker 1
So I think that's another really important thing. In addition to kind of building your prompt in the rules and everything, having the right values in place sets you up for success. And then the last part of it is being able to use those evals to fine tune the models over time. Six months ago, I probably wouldn't have said fine tuning was that big of a thing.

00;08;47;03 - 00;09;15;19
Speaker 1
I thought the models will get better and you can just use the base models and, just build really good prompts and rules around it. I think what we've found is that if you can have the team create the correct evals because they know better than anyone what is a correct response is just the correct tone. If you create those evals, you can actually use them over time to fine tune the models, and you can actually get even better results in, you know, feel better or more confident about is it going to be on point and is it not going to like hallucinate something?

00;09;15;19 - 00;09;22;18
Speaker 1
So that's something that we've leaned into a lot over the past year or so. Is it fine tuning in evolves?

00;09;22;20 - 00;09;41;24
Speaker 2
It's also interesting because obviously in the first place it should you save your time and make your customer service better. But and the second step actually makes your service better overall by just optimize your business in so many aspects and even in a communication. Now when we talk about this optimization, how does the onboarding process work? How long does it take?

00;09;41;26 - 00;09;45;25
Speaker 2
How long does it take to get the data into do define tunings. How's that.

00;09;45;25 - 00;10;01;28
Speaker 1
Working? Yeah. So we have certain tools that we've used like different website scrapers where you can just put in a website URL and it will loop through the site and pull in like product information or map macros and FAQs. We also have APIs to help that, so we can pull in the knowledge base and help this, information from there.

00;10;01;28 - 00;10;20;02
Speaker 1
So it's not just copy and paste everything manually. So that's some of the upfront work that we've done. I think in the initial 2 to 3 week build out time, that's usually around what it takes for us to build the knowledge base, build some rules. Test the AI kind of with like drafting different responses and creating evals off of those responses.

00;10;20;05 - 00;10;38;15
Speaker 1
I think that process is it's kind of an iterative process to get the tone right. And also make sure that like the email signatures, right. You know, it's able to make the function calls that it's doing. And then lastly is kind of just before you go live, making sure that the correct channels is on for email, chat and social.

00;10;38;15 - 00;10;53;29
Speaker 1
Or is it just email? Is it responding to these types of intents around orders and subscriptions, or do we have it answering everything? So I think getting very granular about not just how it's responding, but what types of requests it's responding to is also super important. From a live standpoint.

00;10;54;00 - 00;11;01;17
Speaker 2
Most brands out there are selling internationally, omnichannel and internationally. How can an AI agent support this?

00;11;01;19 - 00;11;17;11
Speaker 1
Yeah, I mean, works in every single language. We've seen it work in some of we've done demos. I think, we did one it was a can finish and it was like perfect finish for, for the customer that we do. We didn't know signing them as a customer, but they were pretty amazed by the demo of at least getting it pretty accurate.

00;11;17;11 - 00;11;36;28
Speaker 1
And that was around a year ago. So I'm sure as the models continue to get better and better, it's, going to improve. So yeah, multi-language works really well. I think it's also a really great it just kind of understanding context and nuance because we'll give it all of Shopify customer, all the orders from Shopify, all the subscription data from Shopify.

00;11;37;00 - 00;11;50;11
Speaker 1
It has all of that context. And that's a really important thing. If you can build that and have that in place, it doesn't matter what country they're in. It's kind of the same structured data that will need to create a response. So, I think a really key part of that as well.

00;11;50;13 - 00;12;13;00
Speaker 2
It's kind of of oppressive, of course. I mean, finding a Finnish support agent might be a little bit difficult, depending where you are. And and you might only have like a handful of requests per year. So it's not for us to get somebody. And it basically just covers all of that. Are there any areas that where AI is improving where, new jobs or new tasks for AI agents coming up that it does not cover yet?

00;12;13;02 - 00;12;34;00
Speaker 1
Yeah. So I think one use case that we've done this year was around on more on the B2B side of like parsing, shuck unstructured kind of emails and orders for, purchase orders. And that use case was great, but sometimes you would get lined up like a customer that would write in, let's say like 50 line. And I was like, here's what we need for this next week for our order.

00;12;34;02 - 00;12;55;13
Speaker 1
Some of the models that we tried early on just wouldn't they weren't able to actually pass through the email and find the correct SKUs. We found with GPT five and some of the latest reasoning models, it actually was able to do those things. So I think that pattern of finding use cases where it's like almost working and it's almost there, but as the models get continuously better, you know, eventually it will work.

00;12;55;13 - 00;13;11;11
Speaker 1
That was one that we found early on. Did it work? But as the models and the reasoning got better, it was actually able to determine like, okay, I see these are the five products the customer wants to order. I can do a lookup into the database to find the correct SKUs, and then I can create the order in the ERP system.

00;13;11;11 - 00;13;27;24
Speaker 1
So, I think some of those sort of use cases where it's unstructured data, or even vision model use cases as well, where it's unstructured, those are the ones where it's going to take a little bit more time, but I don't think there years out, maybe, you know, months out or as the next model, gets released.

00;13;27;27 - 00;13;35;20
Speaker 2
As AI and agents are relatively new technology, can you give some case studies or success stories of customers that you have worked with and what kind of results there?

00;13;35;20 - 00;14;02;12
Speaker 1
So yeah, absolutely. So I think for a lot of our customers on the subscription side, all of it's around retention. So can we actually save customers from canceling? I think that's probably one of our best use cases in the ROI there. Obviously it's not losing a customer. And also they it builds that trust with the brand. If say on a weekend you write in, you get an email notification like, hey, your next subscription orders on the way and you write back like, hey, can you cancel this?

00;14;02;12 - 00;14;18;00
Speaker 1
Or, you know, I thought I cancel the subscription prior to having like an AI agent in place. Maybe the team doesn't get back to that till Monday. The order shifts, the customer's frustrated. Hey, you waited. You you know, you waited and you didn't get back and you didn't cancel the order. Now, this is getting shipped to me. Cancel.

00;14;18;04 - 00;14;33;23
Speaker 1
Cancel my subscription. Versus what we've found is we can actually say, hey, do you want to just pause for six months? We found, you know, if you have enough product already, instead of canceling, do you want to pause or just skip the shorter? And the fact that we can get back to them right away, and then also perform that option for them and confirm it.

00;14;33;25 - 00;14;50;22
Speaker 1
It builds the trust, it saves the customer. And the ROI is great because then the team doesn't have to spend time, you know, manually going back and saying, hey, we're so sorry that, you know, we didn't catch this. It was over. We can or we'll refund you the order. You know, sure thing. So I think that's a really good use case around subscription and order management.

00;14;50;23 - 00;15;14;25
Speaker 2
I love this example. Because obviously speed is really key here. And that's the system that's working 24 seven. And you don't have to call it into a call center in, I don't know, somewhere abroad. And there's just all these kind of headaches that you had to the past. An agent is just there and does the work. Now for founders that are listening, when is the best time to start using an AI agent in your I think, brand journey?

00;15;15;02 - 00;15;31;04
Speaker 1
Yeah. So I think in the brand journey where we found it works great as brands that are basically kind of in that like hyper scaling mode, and they're looking at how do we do either go and get a couple more apps to try and help with this? Do we go hire more headcount, or do we work with the BPO to kind of get temporary outsource headcount?

00;15;31;11 - 00;16;01;29
Speaker 1
I think that's potentially where an AI agent, on the operations financial side can come in and kind of put the infrastructure in place. So as the brand is scaling, they don't need to continue to increase all action. So it's not linearly, scaling with them. I think the other the other side of it would be around like Bfsi, and holidays having something in place ahead of that kind of for the same reason, if you're if you are looking at it from a standpoint of like, we're going to have a lot more tickets, we need to be all hands on deck, having an AI agent in place to basically handle the surge in

00;16;01;29 - 00;16;10;04
Speaker 1
volume and make sure that everything's being orchestrated and customers are getting back to, quickly. I think that's another, you know, key. Key part.

00;16;10;06 - 00;16;21;02
Speaker 2
Basically, now's the right time. If you're looking at Black Friday, Cyber Monday, we should get started right now. Who's your perfect customer in that? So is there a specific industries or verticals that you work more than others?

00;16;21;05 - 00;16;50;14
Speaker 1
Yeah, I would say health and wellness and beauty brands north of 10 million GMV with like thousands of tickets per month is like this sweet spot. If there's lower volume, it might not make a ton of sense. Just because I think teams that are able to handle the volume sometimes end up fighting the AI if they're like, hey, you know, it's we don't need this in place now or hey, you know, took a batch of tickets that I was working on or hey, this process was something that, you know, I did or something along those lines.

00;16;50;17 - 00;17;07;05
Speaker 1
I think it fits the team kind of inundated in there. Like we have a triage set up. We are like firefighting kind of all these tickets coming in like we're, you know, every day it's like that. Then it's it's a lot more, of a just total value add in a win win on our side and on their side.

00;17;07;07 - 00;17;18;22
Speaker 1
I think the, the other side of it is, yeah, primarily work with direct consumer. North of 10 million GMV health and wellness. Beauty is like our perfect sweet spot I think is kind of where we found it. Works great.

00;17;18;24 - 00;17;27;18
Speaker 2
Makes perfect sense. And hyper scaling mode. Obviously that's the best place to to use. I, how does your pricing structure work? How do you charge for this?

00;17;27;21 - 00;17;49;12
Speaker 1
Yeah. So we originally we're doing kind of just a flat, monthly fee around, basically 1500 a month. And that covers unlimited ticket volume. That includes infrastructure, set up everything. It's just a basically a flat monthly fee. We've been recently kind of moving more towards this outcomes based model, where you have a pre-approved, amount of outcomes that could be performed by the agent.

00;17;49;12 - 00;18;08;00
Speaker 1
And so it would still be that 1500 a month. And the only difference is if you go, over that, there is, some, you know, charge per outcome. But the idea is that if we want to basically like, align the ROI and the incentives of us in the brand, we would only pay for if an outcome is performed.

00;18;08;00 - 00;18;24;22
Speaker 1
So that's kind of the option we're giving our customers now. All of our existing customers are kind of on that grandfathered plan, where it's just a flat monthly rate. But we are looking at the idea of, does it make sense to basically say, you know, not per conversation or not per ticket, but can we actually create the replacement or can we actually change the subscription?

00;18;24;22 - 00;18;31;04
Speaker 1
Can we create a return? If we're able to perform that outcome, then we would charge like a dollar per question.

00;18;31;04 - 00;18;41;20
Speaker 2
And you are on the forefront of AI. And I'm just curious to hear from you. We'll be there at time and probably very soon. Where AI agents communicate with AI agents on the customer side.

00;18;41;23 - 00;19;06;28
Speaker 1
I think. So I like maybe negotiating, pricing, negotiating, maybe like you have like a sales concierge in communicating with like, asking for like a discount or applying a discount and, you know, post an order with, a agent and maybe they, you know, neither of them know that they're agents, but maybe they the concierge agent and, you know, hey, you know, my my, I place this order.

00;19;07;00 - 00;19;35;07
Speaker 1
Can you please apply the discount? I didn't do it at checkout. And maybe the AI would be able to perform that outcome and, you know, sends a notification. Hey, I got your discount for you. Just, you know, thanks. You know, thanks. Maybe that could be a future, I think. More on the orchestration side as well. Maybe AI agents from one system of record to another, like a three PL and, like a warehouse agent with like, six agents saying, hey, we just got this request in from this customer, to cancel an order.

00;19;35;10 - 00;19;50;06
Speaker 1
Has it shipped yet? And then the agent can respond back saying, yeah, it's it's still being picked. We're we're able to sell it for number or cancel it for them. So that's maybe a little bit like horizon one. Horizon two. But that could be a cool feature as well as it's just kind of working autonomously in the background like that.

00;19;50;08 - 00;20;04;05
Speaker 2
Yeah I think that's because most of the traffic on the interwebs is bots by now. So I think it's going to happen very soon, and it will be interesting to see how that works out before a coffee break comes to an end today. Is there anything that you want to share with our listeners that we haven't covered yet?

00;20;04;07 - 00;20;31;24
Speaker 1
Yeah, absolutely. I think, kind of we we talked about like categorically, I think this is a new era of software. Like prior to this, you have your SAS kind of dashboard for that. You have your on premise systems. I think now it's kind of looking as like you'll have these kind of environments where you can like monitor agents, you can find to them, and you can kind of create the evals and the right kind of, you put the rules and infrastructure in place, and then it's more of a sort of like, how do we continuously improve it.

00;20;31;26 - 00;20;51;15
Speaker 1
So and then how do you focus on the strategy of like this use case around orders and subscriptions working really well? What else could we add. You know long term. So that kind of shift the mindset of I have to go in and I have to manually kind of do all these different things to how can we optimize this so that I can get time back to look at more of the strategic and growth side of things.

00;20;51;18 - 00;21;02;25
Speaker 1
And so yeah, I think that's really what we're, we're working on is how we that, that system that is up for a leading scaling DTC brand. So, yeah, that's that's what I'm working on this year.

00;21;02;27 - 00;21;20;04
Speaker 2
Yeah. It's exciting times. And I hear I agents left, right and center. And it's definitely the advantages of having them are there. And I think that's will be different where e-commerce ecommerce is going. And definitely something to look into. Where can people go and find out more about you?

00;21;20;07 - 00;21;38;17
Speaker 1
Yeah. So, so, stay sitcom. As far as all that, you can reach me at Dom at state's at dot com, and then also. Right, on LinkedIn, we just launched a newsletter called eCommerce signal, which talks a lot about these different use cases, that we're, we're working on. And then, yeah, on, on Twitter as well.

00;21;38;20 - 00;21;55;03
Speaker 2
Okay. I will put the links in the show notes and consider me as a new subscriber for your newsletter, because I want to stay up on the latest what's happening there for all this. You want to scale. If you want to look into AI agents, reach out to Dom. I think that's a good start. To get this integrated into your e-commerce business and it will save you a lot of time.

00;21;55;06 - 00;21;56;28
Speaker 2
Thanks so much for your time today.

00;21;57;01 - 00;21;57;29
Speaker 1
Thanks for the shirt. It.