Beyond Chatbots: How AI Agents Can Transform Service Businesses

Jeffro (00:01.44)
AI is everywhere right now, but most of what we see is pretty basic. It answers questions, spits out copy, or automates simple tasks. But what happens when AI doesn’t just respond, but actually thinks, decides, and acts like a teammate? That’s what we’re talking about today. My guest is Derek Osgood, founder and CEO of OO.AI, a platform that lets non-technical teams build automated workflows using true agentic AI, just by texting instructions like you would to a coworker.

Now, my guest Derek has led launches for over 100 products at places like PlayStation and Rippling, generating over a billion in revenue. And now he’s on a mission to make sophisticated AI automation accessible to everyone. If you’ve ever wondered how to get AI to actually do the work from sales research to marketing content, data entry, without needing a developer on SpeedDial, you’ll want to listen to this one. So Derek, welcome to Digital Dominance.

Jeffro (00:56.843)
Yeah, I’m excited too. mean, agentic AI is kind of a buzzword right now. know, chat GBT just released their agent tools. It just showed up in my chat GBT account. But to make sure our listeners are up to speed, can you quickly describe what actually makes agentic AI different from the tools people have been using over the past year or so?

Derek Osgood (00:58.547)
Yeah, thanks for having me, Trevor. Pumped a champ.

Derek Osgood (01:18.667)
Yeah, the funny thing, mean, the word agent is getting thrown around in like 10,000 different ways. And even, you know, when you talk about like the way that like OpenAI uses the word agent, it’s actually different from the way that like we use the word agent, which is different from the way that, you know, probably like 10 of the other tools that everybody’s familiar with are. So like the simplest definition that I use when I talk about like agentic AI is it is fundamentally about like giving, you know, an LLM, which, you know, is the large language model is like doing all the thinking when you’re doing kind of Q and A in, know, like Chad GPT, for example, it’s giving it like arms and legs and letting it actually complete tasks. And so you’re basically turning the LLM into the brain and you’re allowing it to like intake some information could be structured unstructured, whatever you want. And then, you know, it can decide what to do. And the real difference between like a true agent and just like an LLM doing a simple kind of function call is that it can actually choose from an array of possible things that it could do. And it kind of decides what is the right action that I need to take. And then it goes and takes that action using, you know, some tools that are accessible to it. ultimately, let’s say about like, you know, allowing like LLMs to actually like kind of take action on your behalf instead of just like respond to a question from you.

Jeffro (02:27.316)
Right. So it’s moving past just a chat bot or content generator into something that can take the response based on your prompt and, know, it’s integrated with Excel or Google docs or a website or whatever, so that it can actually do the things that you want it to do. So can you give us an example of how a true AI agent system would work for a service business? Kind of just walk us through that.

Derek Osgood (02:46.944)
Yeah.

Derek Osgood (02:55.305)
Yeah. So I’ll give you a couple of examples actually, because I’ll kind of talk about like the different types of agents, right? So let’s like start with kind of like the OpenAI version, which is the, basically what they have done is they created what we call like a browser agent, where what it does, it kind of opens up a browser in the background and it goes and kind of navigates through the browser and like, you know, takes action on the websites that you have given, you’ve asked it to go do stuff on. So the example that they throw around all the time is like having it like book a trip for you, you know, it can go and like, you know, search for Uh, search for flights on the web and then, you know, basically like, you know, determine which flights are the ones that fit the criteria that you gave it and then go and book it. Um, those kinds of agents are, you know, right now, like they’re really cool, but they’re pretty tough to like make work reliably. And, um, there’s a bunch of technical reasons why, but like, ultimately it’s just, you’re kind of relying on the LLM to make too many decisions. Um, LLMs are great at making some decisions, but they’re not great at doing it kind of like on a really long tail basis. So the other kinds of agents that are out there, the other really common one that people are seeing is functionally like kind of chat or voice based agents where, you know, it’s really like a chat bot similar to chat GBT, but you’re kind of adding a layer of like having tool access to it where you might be integrating a couple of different tools like Google docs, et cetera. And maybe it can create Google docs. could like, you know, kind of read Google docs, et cetera. I think, you know, those kinds of tools are really, really great for deploying customer support agents where you can actually have it be fed with knowledge about your company. It can answer questions about it, but then maybe it can also book meetings, for example. So that customer can come in and say, hey, I’d like to talk to a person, and it’ll actually book time on your calendar. The third kind of agent is much more workflow-oriented, where it’s an agent that is usually behaving through APIs. So it’s a little bit different from the browser agent, where it’s actually working

Jeffro (04:32.693)
Yeah.

Derek Osgood (04:54.635)
through the tools directly instead of working in a browser that it opened in the background. There’s a bunch of reasons that ends up being more reliable, but ultimately what it does is it allows for higher degrees of accuracy in a lot of the stuff that it’s doing. This is the area that we mostly play with with DoubleO. And essentially, examples there are you can use agents to, for example, every time you get off a sales call, it could go update your CRM with all the data about that sale, about that leak. You could have it… intake a booking from your website and then qualify that person by going out and researching their company and who they are and say, this person like seems like they’re actually a qualified lead. And then it could actually go email them, you know, collect some information from them, book some time on your calendar. So you can do a lot more like complex multi-step workflows where you’re not just like going and reading information from places, but you’re actually like taking action in those tools, sending emails.

updating SaaS tools, all sorts of different stuff off.

Jeffro (05:54.524)
Yeah, and I mean, that’s a great overview. I think it’s easy to just assume, like, if you’re not technical, you’re like, wow, I can just do everything. That’s great. But obviously, there’s more to it than that. I think I want to kind of illustrate the difference between the browser-based approach with OpenAI versus what you’re talking about with APIs. So the browser-based, it’s kind of like the computer is pretending to be a human and looking at what’s on the browser and figuring out what to click on and what to read and stuff.

Humans are great at that. Computers, like obviously it’s getting better, but it’s not the same. it doesn’t know when things don’t go as they’re supposed to. Maybe an image got blocked by your firewall for some reason and that could get it hung up. You as a human would not care necessarily. Maybe you can keep going and find what you need anyways, but little things like that make a big difference and so with APIs, like you’ve got specific endpoints to plug into that do specific things with a very well-defined set of rules in terms of what’s allowed, what’s not, how to ask, and what your possible options are. so computers are great at, you know, discrete, finite options like that. And so you’re able to do a lot more. It’s kind of like having R2D2 plugging into the port. Like he could do that super fast and now he’s in, right? Cause that’s his access point. That’s the API, as opposed to him trying to like talk to the person behind the desk and tell them what he wants. Like he’s going to be beeping away. They won’t know. It’s just not effective or efficient. Go for it.

Derek Osgood (07:31.147)
I’m gonna have to steal this R2D2 analogy. That’s the best analogy I’ve heard for this yet. Because it is totally that. It’s like APIs are how computers were designed to talk to computers and browsers are how computers were designed to talk to humans and they don’t behave the same way and they speak different languages, they think differently. So yeah, totally.

Jeffro (07:46.738)
Awesome. So, I mean, one of things you talk about is, you know, having non-technical teams being able to set up these workflows. tell us what that actually looks like. So how do you make this feel easy? Because on the surface, it sounds complicated.

Derek Osgood (08:04.415)
Yeah, so the way that we, and this is why we kind of took the workflow approach, right? Chatbots are a lot harder to build really, really powerful chatbots because you have to think about all the possible scenarios and questions that that chatbot could get asked and what it should do and how it should behave in each of those different situations. With workflows, you’re describing a process. And most people are pretty good at thinking in terms of what are the steps that I do when I’m gonna go and complete a task. And so it’s like, you know, let’s say I’m using that lead qualification example. It’s like, you know, I’m going to get a lead and it’s probably going come in through, say email and I’m to get that lead and then I need to go check who that person is. So I’m to go look at their LinkedIn. Then I’m going to go check, you know, like a little bit about their company. So I’m going to go look at their company’s LinkedIn, maybe look at their company’s website. And then I need to decide if that person is a fit for my business. And so I need to like, you know, interpret the information that I found and then I need to go and like update my CRM with it. And then I need to, you know, if they were a fit, go email that person. like, that’s a series of like steps that a human would take. And so the, what we do is we basically like, you know, built a bunch of stuff under the hood that takes those simple instructions and kind of translate them into the agent instructions under underneath, you know, kind of the orchestration system. So you can just describe the steps that the agent should take where it’s like step one, go research this person’s LinkedIn. Step two, research this person’s company. Step three, decide based on this information and who my customer is, whether they are a good customer. Step four, go update my CRM with that information. That’s all you need to give us is that level of instruction. And we take that natural language instruction and then we push that downstream into all the agentic workflow. you just write, you literally write what you want the agents to do and you just instruct them like you’re instructing a teammate. And then we handle all the complexity of like, do you make that work? And it just kind of happens magically under the hood. But it’s still not magic. Like you still have to tell the agent what you need to do, what you want it to do. I think that’s where like a lot of people are still in this, there’s like you were saying, there’s this kind of weird gap right now with AI expectations where, you know, people are still believing like AI is magic and AI can like think and actually come up with.

Derek Osgood (10:24.625)
answers with very limited context, but you still need to give it context. You still need to give it direction. You still need to get instruction in the same way that you would like an intern. You know, if you were hiring like a new hire straight out of college and never done the job, like what would you tell that person to do? They don’t have any clue how to do that job. So you need to give them like a little bit of direction and what resources they need in hand to be able to go do it.

Jeffro (10:43.656)
And they still might be limited by what services actually have API endpoints, right? And you’ve got to share your credentials essentially by authenticating on behalf of the tool so it can go do stuff.

Derek Osgood (11:00.939)
It is, it is one of the funnier, one of the funnier user behaviors that I have noticed since, you know, like starting double O and working with AI is like, see suddenly with AI, everybody assumes that they don’t need to log into the tools that they want the AI tool to use, which is you have to log in. You still have to give it access in order for it to be able to use it on your behalf. Yeah.

Jeffro (11:18.15)
Otherwise, your stuff wouldn’t be very secure. All right, so here’s another question. Do you see businesses wasting time or money using AI wrappers instead of true automation or even this agentic approach?

Derek Osgood (11:37.609)
Yeah, I I think everybody, like I encourage companies to like experiment with a bunch of different approaches to using AI. And like, do think like agentic AI, like we were talking about, like it takes a lot of different forms and the form factor like is going to like, there’s different form factors that are better for different use cases. And I think like where people are falling down and like are wasting time is there’s a lot of tools out there that, you they are really just software as service tools, like traditional SaaS tools that have added AI into themselves. And they’re not actually agentic. They’re not like really actually speeding up the process at all, because they’re not taking advantage of, you know, agentic AI’s ability to actually think through problems. So you’re really just kind of trapping yourself in like a much more rigid structured workflow in that sense, which is going to be really hard for you to like migrate down the road. I think it’s more so like, are you creating headache for yourself? a year from now, not necessarily are you wasting time right now? Because you can still solve the problem. Like obviously software has been around for a while and you know, think like people building their processes in existing software are still solving the problem. They’re just maybe doing it in a slightly less efficient way. But I do think like it’s really important for people to not take the approach of like, I’m just going to build this whole process in ChatGBT because it’s like ChatGBT is not agentic. It’s not actually able to like…proceduralize a repeatable process. And when you’re just using the LLM itself, you’re gonna end up like spending so much time just getting your prompts right and getting your process right that you’re not actually gonna be automating it and you’re just gonna be like spending time experimenting, which is not necessarily bad for just getting your head around what AI is capable of. But it’s definitely not the most efficient way to roll this out.

Jeffro (13:23.249)
Yeah, well, and one other thing you touched on there is talking about what this is going to do long term. Like, because I want to hear more about how you maintain the automations and the workflows that you create, because it’s one thing to say, OK, you give it a natural language, it builds it. All right, that’s great. I come back next week, next month. I want to change that. Do I have to scrap that, start from scratch? Can I go in and just edit this one part? What does that actually look like?

Derek Osgood (13:52.297)
Yeah, it’s a super good question. this is, this is like, think where fundamentally the real value to agentic workflow comes in relative to traditional workflows. So like, if you think about like the Zapier’s of the world, which like Zapier has been around for forever. It’s a great automation tool. love, I’ve been a power user for Zapier for a long time. Let’s say using the example of like updating a CRM, right? Like in Zapier, you have to define, okay, I want to go update this object, like the deal in my CRM and I want to update this field in that deal. And that’s how you set up the workflow. So like there’s a step which is like update the value field on the deal object. And then there’s a step which is update the notes field on the deal object. And if you ever change the way that your CRM is set up, like for example, let’s say you change the name of the value field to like deal size, you have to go back in and you have to go update that workflow.

And you have to say, okay, like I changed this. Now there’s a new mapping to the value that I need to maintain. It gets super painful when you start talking about like complex systems that you’ve built out of these tools. And they’re just very brittle. So with agentic tools like 00, like because we have abstracted a bunch of that away and like we’ve basically said, Hey, like you’re at the level of instruction you need to give us is update my CRM with this information. And the agents handle figuring out like, how is your CRM set up? What fields are relevant to go update based on the information that we’re updating with. What CRM do you have integrated into the tool? Then you don’t need to go update the workflow. All you need to do is like, maybe you just disconnect your CRM, your previous CRM, if you change CRMs, connect the new one, and then the workflow will still behave the exact same way. It’ll still go read your CRM every time it goes and conducts that task and figure out like, okay, like there’s a new configuration to the CRM.

I need to go update these fields instead of those fields. So that’s where like agents become really valuable is because they’re a lot more like scalable over the long term, like because you don’t need to change them every time something changes in your business. You just give them kind of high level instructions and they’re able to take that action under the hood.

Jeffro (15:58.778)
Yeah, that makes a lot of sense. And it kind of goes right back to that browser example, right? If someone changed the text on a CTA button and now your steps don’t work, you know, the workflow is broken, but with the agent, can see, they just reworded this. This is the right button. I’m going to keep going.

Okay, cool. So for someone who’s listening to this conversation and they run a service business, what would be an easy starting point for them to kind of dip their toes into getting an agentic workflow set up without getting overwhelmed?

Derek Osgood (16:18.302)
Exactly, exactly.

Derek Osgood (16:32.713)
Yeah. So I, like, I, I always suggest, you know, everybody, the starting point should, as, me just talking about how chat GPT is not worthy, you should be like scaling stuff out. Like I actually do think it’s a good place for you to just get your feet wet on like understanding kind of how, how LLMs think. I’m like how they’re going to interpret a request from you. So I tend to suggest people like play around with trying to do this stuff, like manually in chat GPT once or twice. That just gives you a sense for like how it’s going to interpret a response.

Jeffro (16:37.061)
you

Derek Osgood (17:02.291)
And then I think from there, like, you know, like obviously I’m biased. think double O’s like the easiest way to get set up on this stuff. You know, you can come in, we even co-build most workflows alongside you. So you will help, actually help you kind of like get up to speed as you’re setting up your first one or two workflows. but you know, like I think finding the right agentic tool that like is mapped to the way that you think about process. like define your process first. Once you have a clear picture of like, what is your process? then go find a tool that is ideally a flexible agentic tool, kind of like Double-O, where it’s able to map to that process. I think where people end up having trouble is like when they’re buying a bunch of like kind of point solution tools, which are again, mostly kind of SaaS tools that have added AI into themselves. Then you have to start thinking about like, how are all these tools going to interact? Like what’s the overall tech stack that I’m going to build instead of just thinking about like, what’s my process and how are agents going to kind of like fit into.

So, you know, I think generally like with our tool, you just come in and you describe the process. If you have that process documented, you can just upload it. And then, you know, we take that and we convert it into the whole workflow.

Jeffro (18:04.293)
So here’s another question that kind of along those same lines, but it’s more of once you’ve created this, right? It’s working great. Awesome. I’ve spending more of my time focusing over here. Something might break at some point, right? Do you have to be explicit about failure cases and what to do? Like, hey, if you’re unable to complete this CRM update, send an email to Derek to let him know or is the agent just going to kind of know that, this is what I should do in case something goes wrong?

Derek Osgood (18:42.475)
Yeah, it’s a good question. So this is where like, you know, a of the tool selection is really important when you’re thinking about agentic tools and like, you know, we’ve done a lot of work in double O to basically build like self-healing workflows where basically every step that happens, there’s like a manager agent that like gut checks the work that that step did. And it says, Hey, you know, like this was good. This was bad. This used the right tools. It didn’t use the right tools. And it will already like give course correction to that agent that’s doing the work and tell it to retry it or tell it to do something different. It’s not currently like sending a notification to the person who’s, you know, managing the workflow and saying, Hey, like, you you need to go redo this, but we’re talking about adding that. but yeah, I think like you generally do, like, this is where we get back to like the instruction, right? So like, when you think about building agents, again, think of them, like they are a brand new, straight out of school kid that’s trying to do the task that you’re doing. And they’re pretty smart but like they have no real world experience. So you do need to tell them like, what do you do in the scenario where it fails? So like, know, let’s say for example, like you said, like, you know, let’s say if you’re not able to get an email address off of a website, like, do you tell it, okay, go look at this other place for that email address? Or do you tell it, try and guess the email address? Or do you tell it like, skip the step and don’t give me a fake email address? So like, you know, you can basically give it, you can give agents instructions about like what to do in failure scenarios.

Um, but you do need to, this is why, like, you know, when you’re rolling out, you know, your first couple of agents, it’s really important to like test it, like small scale, just do it like, you know, very onesie twosie style and like, you know, try and get it working in a couple of scenarios. And, know, like where you have good data, where you have bad data and just get it working that way. And then start thinking about like scaling this into like full blown automations where you’re doing it at scale, because you do want to give them, you know, instruction around how to think about like, error.

Et cetera. And some tools will have some of that built in. Some tools won’t. It’s very much like Wild West right now. Some tools have it, some tools don’t. And I think generally you should just be thinking about your own process and what do you do if that thing fails?

Jeffro (20:49.655)
Yeah. Well, and as we’re talking about this, kind of made me think if you have a well-defined SOP, I just dropped that into an agent and now you’ve got a thing that will always follow it. Unlike people who forget about it that they read it once when they started, right? Like that’s kind of a dream in a lot of ways for managers with some of these tasks that are, you more easily defined.

Derek Osgood (21:16.469)
I mean, that’s exactly how you’ll see it. Like if you log into a product, like it is exactly set up that way where you can upload an SOP, the agents will just go do the SOP.

Jeffro (21:22.871)
Yeah, that’s amazing. Because I mean, you can even have chat TPT write you an SOP if you don’t have one today, right? And then you can edit it, customize it to your business. But then, yeah, to just be able to put that in there without spending hours trying to figure everything out yourself, like, that’s awesome. I think that can make a huge difference in a lot of businesses. So that’s great that you guys are doing that. What what should owners and business folks be watching for in the next year or two as agent AKA continues to evolve.

Derek Osgood (21:58.955)
I think the biggest things that I’m kind of excited about and like also just watching the evolution of the space is like all the agentic like tool calling stuff and being able to actually use the integrations that are connected to agents is like all very new. And so some models are better than other models at it. Some, you know, the agents themselves are like still kind of like not always perfect around it. So you have to build a lot of scaffolding like we have, you know, around kind of around the agents to make sure that they do the stuff reliably.

I think, you know, keeping your eyes peeled for like, how are the model companies like building new tooling to help the agents like can think through these like failure states and how do they think through like, you know, what to do when a tool call fails? And I think that there’s a lot of that coming. And, you know, the models are constantly getting better at this stuff. Like you just watch, you just watch like, you know, Gemini 1.5 was like terrible at tool calling and Gemini 2.5 was incredible at tool calling. And so all of this stuff is, you know, kind of like evolving very, very quickly.

I do think like, you know, I, this is one of the few times where, know, I, like, I generally nudge people away, especially, you know, established businesses away from like, you know, kind of trying to stay on the edge of, know, like what’s new. But I think this is a scenario where like the technology shift that’s happening right now is so fundamental that you actually do really want to just like stay on top of like, what are all the new tools that are launching constantly? And, know, like try and actually get out ahead of that because the existing SaaS tools that are adding AI into themselves, like you’re seeing, you know, for example, like Salesforce is launching like agent force and you know, like these big companies are, they just don’t have the speed and agility and they’re not close enough to the, to the actual like iron on building AI to be able to do it really well. And so you’re actually like just taping, just like kind of chaining yourself to a workload.

Jeffro (23:47.945)
Yeah, they’re just bolting stuff on. It’s not AI native.

Derek Osgood (23:55.263)
Yeah. Yeah. You have, you have to be using like AI native tools when you’re doing this, because like the product teams that are building this stuff, like they have to be living in it all day, every day in order for them to do a good job with it. And it’s just like big companies can’t do that in here.

Jeffro (24:07.529)
One other thing I forgot to ask you earlier. With OO and the agents you create, are they able to send tasks to other agents within OO?

Derek Osgood (24:20.907)
Yeah, so this is like a very, I think like architectural philosophical question about like how it even works, right? Like you have a lot of, you have a lot of, there’s many paradigms for how agents, you know, kind of interact with each other. So our agents are doing that. Like when you look at a process in our workflow, like, you know, step one is giving, is handing off to step two. And it’s like, Hey, task one is now going to task two. But what we’re not doing is allowing the agents to like say, you know, from task four, like.

Jeffro (24:22.187)
Yes.

Jeffro (24:32.384)
Mm-hmm.

Derek Osgood (24:47.273)
Hey, like task one agent that’s, you know, kind of focused on research. Like, can you go do additional research for me? Which is this concept of like, you know, the chat bots talking to chat bots idea where, we don’t allow that. And there’s a very specific reason for it. It’s not that we can’t, we actually have the technical infrastructure to do that. But going back to the browser agent thing, it’s one of those things where you don’t really want to give AI that much latitude.

Jeffro (25:15.051)
You

Derek Osgood (25:15.167)
Like if you give it that much latitude, it’s just giving it enough rope to hang itself. And it ends up just like, you know, spiraling off into like weird tangents. You end up burning through tons of credits and tokens in the actual usage that you’re in. your costs spiral out of control and you just end up like creating compounding failures where it just gets worse and worse as these things give bad answers back and forth to each other. So we purposefully don’t do that, but I do think like that is an area where like I I’m hopeful that the technology improved significantly over the next, you know, two to three years. And I think, you know, like that is the long-term future where theoretically agents are going to be able to behave a lot more like kind of employees. And they’ll be able to like actually be given abstract tasks and then go figure out like, Hey, I have this team. I’m to go figure out like who on the team I need to go get help from in order to go complete this thing. But it’s still not at that point yet.

Jeffro (26:04.298)
Sure. Okay. Well, that makes sense. Well, thank you, Derek, for chatting with me today. This is all fascinating. And I know for folks listening, if you’re new to this, it’s easy to get lost in some of the buzzwords, but hopefully this has given you a good idea of how Identic AI can become a real practical teammate that actually gets work done. And I’m personally very excited about this and all the ways I can use it. And it sounds like Double-O is making this really easy for you guys to use if you’re non-techie especially. So if you want to learn more about that, we’ll have all the links in the show notes. Last question for you Derek, before we wrap up. If there’s one thing that you want the listeners to take away from our conversation today, what would that be?

Derek Osgood (26:47.595)
I really think it’s just, you know, like you’re gonna mess up with AI for a little bit and like that’s okay. I think everybody’s still figuring it out. And like the best way that you get good at implementing AI in your business is just go get your feet dirty or get your hands dirty and like, you know, go experiment. And I do think it is like a trial and error experience to do this stuff well. And you can’t really wrap your head around like how this stuff’s really gonna work through best practices by researching it and reading about it because there’s so many different think pieces out there that have very, very different perspectives on how agentic AI should work. So the best teacher is your own experience and seeing how it’s gonna behave similar to a new hire. You gotta work with them and see how they’re gonna behave.

Jeffro (27:31.551)
Yeah, agreed. Well, thanks again, Derek, and thanks to all of you for listening. If you found this episode helpful, please share it with someone who is curious about using AI in their business. And that’s it for now. Take care. We’ll see you next time.