Make Agents vs. Classic Automation: The Truth About AI Workflows

Speaker A: Welcome, everybody. This is Ryan Staley and I am back. I have Tom Andrews, who is the VP of tech and ops at Pavilion, as well as the dean of Pavilion's AI courses. For those that you missed Episode one. Tom's a ninja with AI. Basically broke down how he started with Make Agents, didn't get the result he want, went to Gemini, and then basically leveraged Mana's to get the result he wanted for very complex tasks. Tom, welcome. Happy to have you back, man. Excited to do part two with you right now.

Speaker B: Thanks so much. So excited to be back.

Speaker A: Yes. So, um, for part two, what we're gonna do is Tom is gonna walk through make.com agents because he's done a ton of work on that platform. And I was super excited about, like, when we talked about this, Tom, because, I don't know, it's really hard sometimes to get the platform specific value if you're not on that exact platform all the time. Right. And so I know you spend an extensive amount of time on make, so would love to see, like, what you're up to, have you share your screen and walk through some of the things that you're seeing. Work well with Make Agents because that's a new feature that just got released recently as well. So we'd love to have you. What do you got, man? What are you working on?

Speaker B: Amazing. I'm just opening everything up that we can have a look at. So some background. For anyone who doesn't really know me very well, I'm, uh, a big Make.com fanboy and Salesforce fanboy. Actually, I got a Salesforce on my desk even. But we use make all the time, every day. So we were super excited when we started to see all of the Agentic features coming out. We currently leverage MAKE for pretty much all of our cross system, um, automations. Everything from member onboarding through to collecting MPS scores. MAKE will be involved somewhere in the ecosystem. Pavilion used to be this. This kind of smorgasbord of different tools that you would get logins to as a member. Used to be eight different platforms, eight different usernames. Now it's all synthesized and integrated through make.com as this essential kind of API fabric. Much easier to use than a real API design system, but still gives you all of the flexibility that you would get out of real engineering. The thing that's really cool for me is I run an operations team. Most of my team are, uh, very skilled in Salesforce. Their CRM administrators make as much easier for them to use than maybe going into APIs GitHub, trying to build something out in Visual Studio etc. So, so in the uh, the last one I talked a bit about mentoring and matchmaking. We'll dive a little bit more into the agents right now and I'll show you what I'm up to.

Speaker A: Love it ma'. Am. That's great. I, I think. And how many, how many automations do you have for make?

Speaker B: Why don't I just show you so you have a seizure. There's quite a lot, there's probably people with a lot more but I think for the size of our company, just for context for everyone, we're a 29 person company and uh, we've got that many big automations and then all of these down here, they go on and on and on, are either testing beds or like they don't need to be live, they're one offs. You can see there's a lot here of like adding people to Slack channels. Um, Slack is very difficult to work with. The API is horrible. Make.com makes it a dream. Typewrite is what our member hub is built on. Really awesome technology, very easy to work with, API, very nice to do in here and then all sorts of like mini things. There's an integration with Gemini, I play with it sometimes so all sorts of different things and you can see it touches everything from our financial systems, couponing, our exams, member hub automations and then these here, uh, anytime you see this like really strong purple, this is where it's um, make. One of the things that's really interesting about MAKE is you build out your um, your agents over in this section and with these you put in a prompt just as you would expect and then you give it context if you want to. This is background documents, grounding, all of that kind of thing. Then you give it system tools and these system tools can be any of your other scenarios you're giving it access to. So it works like some other builders of agents where you need to give IT skills to use. All of these have to be um, on demand or custom webhook and many of them should interact with databases. So you can see here I'm listing just two different things. These are actually separate scenarios you can then see here. So another agent that we've got and we've not built a ton out, I've just spent a lot of time with these few that we've built so far is this one. So the idea of this is it would take all of our members, it would look for people with similar interests and it would try to match them out. It again uses a short database pull, so enlist members. You can see it's basically pulling from a data store, aggregating the text into a very simple version of kind of baby JSON, let's call it, and then it's returning an output. So that would then be fed into the agent for it to do something. One of the quirks is you can't just run an agent from this interface, so you need a separate scenario to then, um, run the agent and then it will give you the output directly through a module. And you have to do some work to reprocess that if you want to put it into other systems. So it's been really interesting to play with and kind of see it come together. There's not a ton of learning material on it yet, which for anyone who's taken my course, I know you that were there for a lot of it. Ryan the number one way I assess the maturity of a platform is the training content and whether someone could go and learn about how to build an agent. Because even if you know a lot about agents, you don't necessarily know how to use an agent builder from someone like make. So uh, I think there's some maturation in that side of it to come, but it gets reasonably good outputs. So let me find the run scenario here.

Speaker A: While you're doing that, let me ask it like, so what is your take on that, man? Like, I mean, with the state of where we're at right now with, I mean we got make, which is automation and then make agents. Same thing with Zapier. Right? Zapier is going down the same route. There's other tools like gum loop and so what do you recommend for someone who is just getting started now or has familiarity with LLMs to get like go cross over and be like fully automated and know how to build these agents?

Speaker B: Yeah, great question. I probably have a controversial answer. Uh, considering we talk a lot about agents. I would learn make because automation, in my opinion and other opinions are available, is way more mature than agents. And I can solve any business issue within a few hours with, okay, maybe not any business issue in a few hours. I can solve many major business problems within a few hours just with old school, very well built automations. I don't generally need to get an AI involved unless I'm doing text passing and I would still then prefer to wrap the AI in an automation. So I've got at the moment kind of bots running in our slack where people can send a message, AI will digest it it will categorize it for me. And based on the categorization of that text, I run an automation. So I would start with a platform that does both automation and agents because an automation skill set is immediately and clearly valuable. Agents are still finding their feet and it's very easy to explain why an automation did what it did. Very hard to explain that with an agent. And sometimes it's the explainability that's really core to the use case. If an agent makes a sales forecast, you can't necessarily defend that to the board unless the board is fully bought in on the value of agents. It's very easy to defend if you built a machine learning model and they can actually like get into it and see how it's built and play with it themselves. I found a lot more confidence in automation and ah, machine learning than I have in agents in very executive level discussions. Which is why I think I'm still in my comfort zone of automation and I play with agents. I just don't find as many uses for them. So a platform like MAKE lets you go both ways. Love it.

Speaker A: Great example. And then so for, let's say take it one step farther while you're pulling this up. So say I wanted to go learn MAKE today, would you just recommend following like through the learning path that they have or the academy? Is that what you would recommend?

Speaker B: Yeah, totally. Go do Make Academy. I did the whole thing waiting for a delayed flight for four hours on my phone. Had a great time. Admittedly I'm very familiar with make, so I didn't do any of the practicals but um, it's very easy to go through all their learning, really get into it, ah, understand all the concepts and, and get it all out of the Academy. You can also choose to go into all the projects and do the practical examples which can also be really nice. So this is now finished loading. So you can see we've got a response in JSON and it's pulled the data out. And for the person it started to pull all of the data together. It's interesting. It's pulled the word mentor out. This is a different thing but mentor, member and mentee are so similar. I'm convinced it pulls the wrong actions into itself quite frequently. But you can see this is quite hard to work with. It's like JSON. It's not super well structured. It would be very hard to take this output and do a lot with it without passing the JSON and passing the text. Obviously that is something you can then automate and make, but it is one of the things that is a bit prohibitive. Now what you can see is that this person who is obviously a test Account is the CEO at Ah, TestCo. And you can see a bit further down that it's matched them with this person test. So it's found something in the word test and it's decided you should meet everyone else for test in the email. Which is really strange but also kind of cool. Down here you can see leadership is coming up a lot. It's like a topic everyone's interested in so it can get there. It's still got some ways to go compared to some of the other agent builders, but I think I've also got some way to go learning about all the things it can do. Uh, and I probably need to spend a bit more time in it. Yeah, it's cool. The other bit I don't love is like you can see the message that went in, you can see the tools it has, the additional system instructions. But what I really want is an output like Manus, where you can see it invoking all the things and all the actions it takes. Like I want more of a real breakdown of exactly all the steps that it's taking. One of my favorite things generally speaking in make is when I'm in my bigger uh, complex. Here's my big one. This is like the biggest automation we've got. You can see it's multi screen. It's got a huge amount of nodes and modules. Managing this is my pride and joy. I love this thing. You can tell when someone in my team has worked on them because they're less anal about making everything line up nicely. But it also makes it easy to find what they're working on. And then this one, when someone runs through, you can see each step it goes through and where it errors. It's like classic automation. I would love something like that. When I'm looking at this, like I want to know it's using the tool. It's done this, it's done this. It's very hard for me to know how to improve this with, without the thinking and cast your mind back to being at school. I used to constantly get told in maths exams, show your working. And what I really want from agents, especially in make, where I'm so used to seeing it's working, is to see a bit more of its working because it would be easier then to refine the output by improving the input. It's very hard to write a uh, progressively better prompt when you're not really sure what it is about the prompt that's not leading to the output. Just seeing input and output. Uh, to me a little bit, little bit difficult to work.

Speaker A: Yeah, like it's. Why do you think that is, man? Why do you think they like, like it's almost like the agent's way more complicated than the workflow, I mean, or the automation.

Speaker B: Yeah, exactly. And like, you know, as I've said before, this is an opinion and other opinions are available, but I half wonder if they did a little bit of what Salesforce did. Like they said, uh, we're going to build an agent platform, we're going to rush it out to market and we'll be one of the first to get an agent platform out there and people will start using it and we'll improve it over time and eventually it will be good. Like Agent Force is still not the best. It can have some incredible outcomes and you can build some incredible agents. If you deeply understand the Salesforce ecosystem and Salesforce automation, you can't just spin up an agent. You've got to build a well prompted agent, arm it with a library of grounded data, build it a series of actions that it can take using Salesforce Flow and Apex and synthesize all of that together into one agentic experience. But it really is agent orchestrated automation rather than something that's truly, truly agentic. And so I wonder if MAKE is kind of in the same boat where they're using make.comagents as agentic steps within, within larger, uh, classical automations so that they can hold market share and have a place for it as the technology improves. It's still a very approachable, very easy to use and relatively effective offering in the market. As good as many of the others out there. I, uh, just think that the entire kind of domain needs to level up a little bit and offer more to people who are very experienced with technology and like workflows and automations and machine learning so they can really get into it and understand like what's the lever I need to pull? What, what do I need to twist and refine? Like how do I get better and better and better at this without iteration? You're kind of taking away the core skillset of most people with a brain like ours. Because the world we have always worked in with machine learning and AI is very, is very iterative. There's like steps, it progresses over time.

Speaker A: Yeah, you got to get the feedback and then adjust and adapt and wait. So what do you think? I forgot. Are you guys a workspace user, like a Google workspace user? Are you a Microsoft? Okay. Have you heard of Agent Space at all from Google?

Speaker B: I've seen the like release notification. I've not touched it and I really want to make time and space to like give it a go. Um, because obviously I use Gemini the most as I've said in, in the previous one. So I really want to try it but I've just not had the time yet. Have you played with it?

Speaker A: So here's the thing. I'm um. It's. It's not general release. It's like private special release. Even though they quote unquote released it.

Speaker B: Yeah, I saw, I saw marketing that said they'd released it.

Speaker A: Yeah. So here's the thing though. On a positive note, I have a live walkthrough with Google coming up this week. So I'll let you know how it goes man.

Speaker B: Please do.

Speaker A: I'll let you know because like it's almost like you got to do it through Google Cloud. And so anyways I got a live walkthrough with Google uh, coming up and then the other I was going to ask you about is like because this combines like the whole ecosystem is Microsoft right With copilot agents and I. Do you have any experience working with those at all as well or. No. No.

Speaker B: Okay. No. We're a pure Google house. And then everything else is Salesforce because obviously Slack is Salesforce. We don't even offer like nothing pavilion wise would be available through teams. So I've just never really.

Speaker A: So here's.

Speaker B: I also would be terrified to.

Speaker A: Well here's the quick and dirty of it, man. Like I was so I was at the Gartner chief sales officer bad last last week and a lot of CROs, CSOs, chief commercial officers, whatever sales leaders, like they're really frustrated with Microsoft copilot just because like the team's not using it. The results aren't that good. And so like I. They've done a lot of updates with. They had Microsoft build last week and so I'm like all right, I'm going to see based on the updates as of last week with agents what, what is available. And so it's kind of interesting, man. They have like a deep research agent like OpenAI does but it's like hidden and you have to find it. So that's one thing that's weird, I don't quite understand that. And then number two is there are quote unquote agents really seem like custom GPTs with simple automations built in there. Now I haven't dug super deep on that. But that's my like initial experience in terms of working with them. Uh, it's just what I've noticed is Copilot struggles with just even basic prompting. So it's kind of a weird situation. Like I think they did a good job of making it accessible with like automating but that's kind of where I don't know if you've heard anything different but that's just kind of my, my 2 cents on it.

Speaker B: As of right now, not really the sphere I'm in. Most of the people I speak to are uh, salesforce nerds like me. But it's interesting. Like I think with a lot of these platforms they, they rushed. I think it will be a while until we really see the true capability of them. I half wonder if there's going to be like a halfway house where platforms like make instead of a p pure agent builder they start to build agentic modules for their existing automations and makes own AI interface that helps you build workflows is phenomenal. Really, really good. Makes very good suggestions. So I think we're still to see the ecosystem kind of mature to the point people really can build what's been sold to us. This could be one of those classic cases where the marketing team is super excited about what's going to be possible and the reality of the product that's in general availability just isn't quite there.

Speaker A: Love that man. It totally aligns with what I'm saying. So it's, it's good to know that um, I'm not hallucinating myself when it comes to this. So we're up on time though, my friend. Where can people find you where they find more about you and Pavilion?

Speaker B: They can find me on LinkedIn, uh, Tom Andrews or slash tommandrews. And if they're Pavilion members they can come speak to me in Slack or go to our website Joint pavilion dot com.

Speaker A: Well, thanks for being on the show again, Tom. And it was excellent um, to have you on and it was a blast man.

Speaker B: So good to be here. Thanks so much, Ryan.

Speaker A: All right, and I will see you on the next episode.