Summary:
A chatbot can tell you what to do, but an agent can actually do it. In Part 3 of the refreshed Fundamentals of AI series, Amith Nagarajan and Mallory Mejias move from the brain in a jar to the AI worker, building a practical picture of agents for association leaders. They break down the four components of every agent—model, instructions, knowledge, and tools—explain how agents differ from traditional automation, and unpack agent loops, sub-agents, and why not every step needs a frontier model. They also trace why unified data and a semantic layer matter so much, how MCP and computer use connect agents to even the most dated systems, and how to set human checkpoints as an agent earns your trust. Along the way, Amith makes the case for member service as a low-risk, high-impact place to start.
Timestamps:
00:00 - Back for Part Three
02:53 - The Four Parts of an Agent
07:22 - Agents vs. Automation
13:43 - Agent Loops in Action
18:51 - Can Anyone Build an Agent?
26:16 - Why Agents Need Unified Data
33:27 - Tools, MCP, and Human Checkpoints
38:07 - Computer Use for Legacy Systems
41:27 - Member Service as a First Agent
👥Provide comprehensive AI education for your team
https://learn.sidecar.ai/teams
📅 Register for digitalNow 2026:
https://digitalnow.sidecar.ai/digitalnow
🤖 Join the AI Mastermind:
https://sidecar.ai/association-ai-mas...
🎀 Use code AIPOD50 for $50 off your Association AI Professional (AAiP) certification
📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE
🛠 AI Tools and Resources Mentioned in This Episode:
Sidecar Sync Ep. 142 - The Data Episode ➔ https://shorturl.at/WZZ7v
ChatGPT ➔ https://chatgpt.com
Claude ➔ https://claude.ai
MemberJunction ➔ https://memberjunction.org
GPT-5.6 ➔ https://openai.com/index/gpt-5-6
Kimi K3 ➔ https://www.kimi.com
Gemma 4 ➔ https://ai.google.dev/gemma
Claude Code ➔ https://claude.com/product/claude-code
LangGraph ➔ https://www.langchain.com/langgraph
Model Context Protocol ➔ https://modelcontextprotocol.io
Codex ➔ https://openai.com/codex
Claude in Chrome ➔ https://claude.ai/chrome
ChatGPT Work ➔ https://openai.com/chatgpt-work
Sierra ➔ https://sierra.ai
Decagon ➔ https://decagon.ai
https://www.linkedin.com/company/sidecar-global
https://twitter.com/sidecarglobal
https://www.youtube.com/@SidecarSync
⚙️ Other Resources from Sidecar:
More about Your Hosts:
Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.
📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan
Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.
📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias
🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖
[00:00:00:14 - 00:00:09:17]
Mallory
Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.
[00:00:09:17 - 00:00:24:09]
Amith
My name is Amith Nagarajan.
[00:00:24:09 - 00:00:26:08]
Mallory
And my name is Mallory Mejias.
[00:00:26:08 - 00:00:56:06]
Amith
And we are your hosts. And if you've made it this far, you are on part three of three of the foundations of artificial intelligence for associations. Hopefully you've been enjoying the series so far. We have been and we're excited to land the plane bring it home and hopefully share some final nuggets of wisdom that complete that foundational layer for all of you in the association market, working to transform your organizations with this amazing technology.
[00:00:56:06 - 00:01:10:00]
Mallory
Mm-hmm. And actually Amith, I will say, I was wanting to continue the series. So this may be a part three of four, a part three of five. I think when you talk about fundamentals, we could really go on forever with this stuff. Don't you agree?
[00:01:10:00 - 00:01:16:18]
Amith
I'm on board with that. So it'll be part four of four until it's part five of five. And I like that. We'll keep going a bit.
[00:01:16:18 - 00:01:31:18]
Mallory
Honestly, this could probably be its own break off podcasts, like thinking about the fundamentals of AI as they pertain to associations, because sometimes on the pod, we get into some technical topics or kind of go off on tangents. But I do think with fundamentals, there's a lot to cover.
[00:01:31:18 - 00:01:49:02]
Amith
Yeah, in these episodes, when we do fundamentals or we talk about data or topics like that that have a lot of durability, we tend to find that there are a lot more evergreen than we have listeners going back. And that's why we're redoing the series, of course, from I think very early on in the pod journey. We did a fundamentals
[00:01:50:10 - 00:02:49:09]
Amith
pod episode or two, and we still get tons of hits on that every single week. So it's time for a refresh, although I think that the older episodes might still get listens even with the new ones out there. Who knows? We'll have to see if we can go update those and throw a little ad in there to listen to the new ones instead. But, you know, Mallory, I think that the idea of fundamentals and to your point, it's such a critically important thing because it doesn't really have a lot of the sizzle when you hear about it. You know, it's like taking, oh, I gotta go and take like English one in college or math one or whatever the topics are, where it's like these are foundational elements of whatever you're gonna go focus on that's much more skill specific or profession specific. But that fundamental layer will serve you well. So you're making an investment right now, digging into it and understanding the basics really well. And those basics, to Mallory's point, they keep changing. So keep tuning in, we'll keep updating this series, we'll keep adding to it. And at some point, we'll probably redo the whole thing again because the field is changing that fast.
[00:02:49:09 - 00:02:51:18]
Mallory
Yeah, maybe we'll do that next week, so stay tuned.
[00:02:53:02 - 00:02:53:03]
(Laughing)
[00:02:53:03 - 00:05:19:24]
Mallory
Well, if you haven't listened to the first two parts of this series, I would highly recommend that you go and do that. In today's episode, we are gonna be talking more about AI agents as a focus topic. However, it's building on what we covered in those first two. So I highly recommend that you check those out first. If you haven't, last time we rebuilt the mental model of kind of what AI is. And we left off kind of with something that's like a brain in a jar. So we talked about language prediction. We talked about how models can think longer when the problem is hard. And that's kind of where we left it off. So if that resonates with you, maybe you are someone who opens Chat GPT or Claude and you ask a question and you get an answer that's very impressive and you think this is amazing. But notice what happens next. Maybe you copy that output somewhere else, then you go and you do the actual work. What you were talking to essentially is a brilliant and kind of completely helpless worker. It's a world-class consultant on their first day who's never seen your organization and can't open a single one of your systems. So what do you have to add to turn a brain into a worker? That's what we wanna talk about in this part of the series. So first, what an agent actually is or the components of an agent. The word agent is everywhere right now. So I felt like it was helpful for us to use a framework to break this word down, which is that any agent comes down to four components. You've got the model, instructions, knowledge, and tools. The easiest way to think about this is comparing an AI agent to a new hire in your association. So the model is kind of like the person or the brain, who you hired, the raw intelligence and judgment. The instructions are the job description and the onboarding, what the job actually is, how you want it done, what to do when they're not sure. The third is knowledge or context, which is everything about your world specifically, your members, your bylaws, your history, the reason you do things the way that you do them. A brilliant hire with no institutional context gives you confident, generic advice. And the fourth is tools, which is system access. The logins and permissions that let them actually touch something. Without tools, your agent can advise you all day, but it can't actually do anything. So Amith, we covered kind of very basically what makes up an agent, but what can an agent actually do that a large language model cannot?
[00:05:21:08 - 00:07:22:00]
Amith
Yeah, I think you framed it really well with the components that you described. And so the absence of tools is similar to hiring a brilliant human, putting them at a desk in your office and not giving them a computer. They wouldn't get too far. I don't care how smart they are or how great their background is. If you don't give them tools, they won't be able to do much for you. They might be able to come back to you and say, "Hey, I thought about this problem for a while, and here's what I think you should do about it." And you could in turn talk to that person and go back and forth, and they might be a wonderful thought partner, but they couldn't actually execute the work. That's the big difference. So agents are able to run tools, and I'll give you some examples of tools. Let's start off with some simple things. So sending an email on your behalf. That may sound a little bit scary because you might not want an AI system to send an email for you right now. Or some of you might say, "Yeah, I'm totally comfortable with the kind of responses I'm getting. I'd love for the AI to be able to just fire off that email." So when I'm talking to Claude and Claude says, "Oh yeah, this is the way I'd write that response," you can actually have it send the email for you. So that's a very simple example of a tool, send an email. Another example of a tool might be something like, register somebody for an event. So you're talking to a member, and the member says, "I'd love to come to the annual conference, and how can we make this really easy? Let's just go ahead and register them for that." And that might require accessing data from your AMS. It might require going to a separate system like an event management software of some sort, and invoking an API to actually create a registration in that system. And then it might result in having to email a confirmation, or perhaps the other system already does that. So it's essentially connecting the AI to the tools you, as an association staff member, actually use. So it's really opening up the door to connecting the brain in the jar with tools that you as a human are accustomed to using, and giving some select sampling of those tools to the agent.
[00:07:22:00 - 00:07:35:11]
Mallory
Mm-hmm. Associations have been using automation for years, email triggers and workflows and that kind of thing. And they may be thinking of automation as you're talking. Can you explain how an agent is actually different from automation?
[00:07:35:11 - 00:08:02:24]
Amith
The biggest difference is that automations really are recipe-based, where you can essentially say, "I'm going to follow this set of steps to build this outcome." So like a recipe would typically have, it says, "Here's your ingredients. You need these five things. This is how you prepare your ingredients. This is how you mix them. This is how you cook it. And this is how long to cook it for." And the outcome is, "Here's a beautiful souffle," or whatever.
[00:08:04:01 - 00:09:05:09]
Amith
So your recipe doesn't really change. It's the same deterministic steps that you follow over and over. And many things in business, this is perfectly fine. This is a great approach, where you can say it's step one through step 20, and that produces a great widget at the end of the conveyor belt. Wonderful. For those types of things, it isn't so much that an agent can do them better, because it's still 20 steps, but it's that the agent potentially can invoke those steps for you, so the agent can kind of like fire off that process. But when you think about deterministic or repeatable recipe cookbook-y things, agents are a great invocation point, meaning they're where the process can be started or managed, but they don't really necessarily provide you a whole lot of leverage. Where it gets much more interesting, in my opinion, and the difference between automation and agents really starts to become apparent, is when you're dealing with non-deterministic or variable
[00:09:06:14 - 00:10:19:04]
Amith
types of steps. So unlike a recipe that says, "To bake this cake, I need to follow this exact set of instructions," if I were to say to you, "Hey Mallory, can you research the current state of my industry, find out what the biggest concerns are that people have, and then put together a report," you could try to break that down into a set of steps. You could say, "Okay, well, the first thing I have to do is I have to conduct some interviews, and I have to do some web research. Then I have to compile the data, and I have to do maybe some statistical analysis to really try to analyze the data that I collected. And then I'm going to formulate some hypotheses and see if I can test those with some more data. And then finally, I'm going to put together a report." You might come up with a stepwise progression, kind of like that. But the thing is, is there's a lot of judgment that goes into each of those steps. And in each of those steps, how you approach it could vary. Like you have 10 interviews and you might say, "You know, those 10 interviews don't really provide me what I was looking for. I need to do another 10 or another 20." Or maybe actually, instead of that, you also should access a database that we have that has prior surveys of members and look at that data. That's another thing I have available. So
[00:10:20:05 - 00:12:24:17]
Amith
in the example I'm giving, as a human that has both judgment and creativity, you might choose to use other tools that are at your disposal that you're aware of, or you might look for additional tools, thinking intuition-wise that such tools may exist that you just personally haven't used. Like you might, like we often will go ask a caller. You can say, "Hey, I'm trying to solve this problem. Have you ever dealt with something like this?" And that caller, you may say, "Oh, you know, like three years ago, I did a project kind of like that. And I use this one database that we have over here. You may not have heard about it. And you might have to like, you know, dust the cobwebs off, but you can still use it. And you go look at that, right? So it's not a linear progression with a predetermined set of steps. So this is where agents really, really shine. You give them an objective and they can autonomously or semi-autonomously get to that outcome, taking the steps that they decide makes sense. So it's basically the agent has the autonomy to basically keep going, keep working on your problem, using the tools that you have available or you make available and making judgment calls along the way, using its creativity to then come to the outcome you're asking for. We actually have an agent built into member junction that does exactly what it is described. It's called research agent. And this research agent is endowed with several different tools, tools that can query databases, tools that can search Google, tools that can search your private repositories of content like SharePoint, for example, other tools that can write code to do data analytics at scale based on the data that's collected, other tools that allow it to generate images or to search images. And then finally, tools that help it craft a really stunning professional report. And so this agent can run sometimes for, you know, 10, 20, 30 minutes, even hours to produce pretty extraordinary results. But that's the power of agents, Mallory, is that agents can do what we do, where you say, I have a goal, go figure it out, and here's a set of tools, and here's some knowledge you can access. That is the big difference.
[00:12:24:17 - 00:12:38:14]
Mallory
AI agents can take action and they can also utilize, whether you wanna call it judgment or not, they can utilize at some degree of judgment in an open capacity to kind of make the decisions and tools they need to achieve the objective.
[00:12:38:14 - 00:12:53:14]
Amith
That's 100% judgment in my mind. And I think there's a high degree of creativity in a lot of these models. And it's also how you set the agent up. You can make AI actually be much more creative or much less creative. There's ways of tuning the ways these models, they're not
[00:12:54:15 - 00:13:42:23]
Amith
always work all these the same way. There's ways to optimize for different goals. So if you're doing a research task at the front end of it, you might want your models to optimize for very creative thinking. Whereas later on, when you're doing statistical analytics type work, you might want the models to not do that. You might want them to be much more deterministic in how they work. But the big difference between classical automation is you can draw it out on a flow chart. You can draw on a whiteboard, step one, step two, step three, what the decision criteria is. Whereas with what I described earlier, you need a really big whiteboard because there's kind of an infinite number of possible paths that can be taken in that research example. And there's a lot of things in business that are like that where there is a significant requirement for judgment and a pretty significant requirement for creativity many times.
[00:13:42:23 - 00:14:49:12]
Mallory
We kind of started to touch on this already, but I want to spend some time on how the work actually gets done. So we're going to talk about agent loops and then also teams of agents. So let's say you've hired this new person and they're brilliant and you've told them what the job is, you've given them the context and system access, how are they actually working? It's a lot actually like how you work. You get handed a task, you make some kind of plan, you take the first step, look at what you produced and often you decide it isn't quite right. So maybe you revise it, you keep circling until you've got something that actually meets the goal and then you turn it in. That is essentially what an agent does and what you kind of touched on Amith and it's called an agent loop. So plan, take action, look at the results, decide whether that worked, adjust, go again until the job is done. A single prompt with an LLM might get you a one shot at an answer, a loop gets you something that can attempt a thing, notice it fell short and try a different approach without anyone telling it to necessarily. So Amith, you talked about this a bit with the research agent example. Did I leave anything out or are there any other examples of agent loops you could share with our audience?
[00:14:49:12 - 00:17:47:15]
Amith
I'll throw another example at our audience, something that we're in the process of building and deploying for Blue Cypress right now, which is a marketing agent that kind of acts like a chief marketing officer. And what you do with it is you build a strategy. So you have a strategy and it might be company wide or it might be for a particular product line. And then once the strategy is built, the job of the marketing agent essentially is to continue to monitor on an ongoing basis over time signals, these signals come from different marketing programs that you have. So you might have a blog on your website, you might have paid media, you might have email campaigns, you have a variety of different programs. Some of these programs may be well established, long running programs, others might be more experimental. And what you want the marketing agent to do, and in our case is to continually have a creative element where it can hypothesize new experimental ideas for programs that might supplement or possibly replace existing programs. And then the facilities to actually run these programs. So for example, if the marketing agent is looking at our data and says, you know, we tend to have really good results in converting people on the Sidecar website when they're looking at the Digital Now content, if they've been to the site at least twice. So based on that knowledge, let's focus on a retargeting paid media campaign and retargeting, if you're not familiar with it, is a way of bringing people back to your website through very targeted ads online on websites other than your own. And it encourages people to come back to you if they've already been to your site at least once. And so there's a lot of different techniques for that. But so the agent, this is a very typical paid marketing play that a lot of CMOs would say, hey, that makes sense to try that out. So the marketing agent might hypothesize that a retargeting campaign could be helpful. And so it might propose a program that has perhaps a small budget. It might say, well, let's try a thousand dollars or some small budget, and let's run this for four weeks and see what the results are. And it proposes the targeting strategy, the creatives, et cetera. It just basically builds the whole program. Just that one piece to build the program is a loop. It's very much like, hey, build the program, which is what we would call in our terminology, a sub agent. Think of it as an org chart, where it's a specialist on the team who knows really well how to do this particular task, like build a marketing program outline, right? And then the overall marketing agent, in our case, will then launch the program. Now we can allow the agent to just launch the program on its own, or we could say, well, there has to be a human approval gate. In our case, that's what we're doing, because it's a trivial amount of time for a human to review it, just to make sure the agent isn't doing something totally nuts.
[00:17:49:05 - 00:18:51:03]
Amith
And then once it's approved, the agent turns on the program, it starts spending the money, because it has, to your earlier point, it has the tools, the access to actually set up the ads on the retargeting platforms, and then start monitoring them, and seeing, hey, how's it going yet? The hypothesis was, this $1,000 or $2,000 would result in this number of conversions over this period of time. And it starts monitoring that, and then ultimately it's gonna choose, oh, well, this is going outstandingly well. Let's increase the budget. It's going so well, we should double or triple or 10x the budget and see if the value holds, if the conversion rates hold. Or this isn't going very well. Should we pause it? Should we let it stay the course? Et cetera. These are the things that marketing leaders would do. And so over a long running loop, that's an example of not just achieving a task like a research report, but continuing to operate essentially as part of your team on an ongoing basis. So that's where the process of deploying this right now for Blue Cypress across almost a dozen different individual brands.
[00:18:51:03 - 00:19:10:21]
Mallory
Okay, so it sounds like in terms of how the work gets done, you've got agent loops, which is the agent's ability to iterate on the work that it's doing. But then you've also got teams of agents. So you could have kind of a lead AI agent, like the marketing agent in your example, but also sub-agents that break out and do tasks based on what the lead agent plans, is that correct?
[00:19:10:21 - 00:19:41:07]
Amith
Yeah, exactly. So the lead agent might come up with the idea for what it wants and then say, I want to build a program that does these things. And so the sub-agent isn't thinking about, is a retargeting campaign or program a good idea? The sub-agent's just saying, my job is to build a retargeting campaign. And then the overall agent, what we call the CMO level agent, is actually coming up with the ideas and monitoring the results. But the sub-agents might be specialists in particular specific skills.
[00:19:42:14 - 00:20:07:03]
Mallory
I can imagine someone listening to this one might think, okay, this sounds quite complicated and I hear you. And they also might be thinking, wow, using the frontier models out there for all of this work sounds like quite a bit of work in general. Can you talk about the importance of considering model size when you're thinking about teams of agents and how we don't necessarily always need the frontier model for every part of that process?
[00:20:07:03 - 00:21:12:00]
Amith
Yeah, that's like asking Albert Einstein to make you a cup of coffee. I mean, you don't need the most powerful model to do the simplest tasks. Sometimes that's really inefficient, right? Because a lot of people want to talk to Einstein, so there's a backlog and it's slower and it's much more expensive, obviously. So the idea ultimately is, is that use the right model for the task. And when you build an agent, depending on the agent tool or agent framework that you use, you will have choices to be able to say, well, for this part of the work, we'll use GPT 5.6. For this part of the work, we'll use Kimmy K3. For this part of the work, we'll use something really small like Gemma 4, which is an open source model from Google. That's very small, very fast, and also quite smart. Many of the smaller decisions that agents have to make, like which tool should I use? Let's invoke the tool. Let's see what the results are. Are very simple and don't necessarily require the high powered model. So to your point, Mallory, we generally recommend people will set up their agents with a kind of a medley of different models that are optimized for the particular task at hand.
[00:21:12:00 - 00:21:28:20]
Mallory
That makes sense. And for the person listening to this thinking, okay, this sounds pretty complicated. Amith, would you say an average person working at an association could build an AI agent, or do you think this is a job for developers that are very skilled with AI?
[00:21:28:20 - 00:24:26:16]
Amith
I think the average person can definitely do it. The question is, would you want them to, and how do you protect the organization when you're building these agents, which are all topics that we can touch on in this episode, but with tools like Clodcode or ClodCo work, you can in fact build agents just like what I'm describing. Now, if you do it in a proprietary platform like Clod, that means you're using Clod for everything, which is basically frontier level intelligence. They do have different size models with different effort levels, so it has some optimization opportunities, but you're using Clod at that point. Same thing is true if you use chat GPT, you can do similar things there. And so the average person could use something like those tools to build the basics of what I'm describing, to deploy it in a secure and governed way. And what I mean by governed is that we know exactly what the agent is doing, the agent has extremely tight controls on what it's allowed to do, all of that. You do want to consider an agent framework of some sort within our open source member junction platform, which is the AI data platform we talk about a lot. There is a full agent framework that's part of that totally free open source environment. The reason we built it was we wanted to have the agent framework layer live in the same place as the data layer, because that's just kind of a natural combination. For example, with data governance rules, you want to be able to control exactly what data the agent's allowed to see. So it makes sense to have them together. And the agent framework in MJ is not only free, but it's extremely powerful. And it's actually pretty easy for people to use that are not developers. There's actually an agent called the agent manager, and you simply have a conversation with this agent and tell it what kind of agent you want. And the agent manager can build new agents for you, can modify existing agents, and off you go with your agent. You can actually have cloud code build agents that then get deployed inside MJ as well. So there's a lot of ways to use the member junction framework for this. You can also use other tools like Langgraph is a very popular framework that is much more developer oriented. Member junction is more of an enterprise framework that's designed for enterprise governance and safety, and very powerful as well. Whereas Langgraph is much more of like a power tool for developers that is at the code level is where you use it. So there's many different ways to do it. Bottom line in your question is building agents that you can use for your own personal needs is totally within the capabilities of anyone who can speak to a computer in natural language. To deploy it at the association wide level, to know that it's done securely, safely, there's a bit of work that has to go into that. And so it's probably a good idea to get a little bit of help from somebody who has expertise in building this stuff, at least at first, and then learn how to crawl, then learn how to walk, and then learn how to run with it.
[00:24:26:16 - 00:25:34:01]
Mallory
I want to spend some time talking about what your agent can see and do. So two of those four components that we talked about earlier, decide almost everything about how useful an agent will be for your organization. The first is knowledge or context, meaning what it can see. Think about where your member information actually lives, some in the AMS, some in the email platform, some in the event system, some in the online community, and a meaningful amount, probably in a spreadsheet on somebody's desktop that maybe only one person fully understands. An agent that can only see one of those systems essentially an employee locked in one room. It will do the task you gave it confidently with a fraction of the picture, which is the whole reason you're hearing so much about unified data and AI data platforms, Amith, that you just mentioned. Now, I want to let you all know we have a whole episode on data, it's episode 142. If you're feeling like, "Oh, I feel like I need a deeper dive on this, I want to understand more," highly recommend that you go check out that episode. But at a high level, Amith, can you just speak about why, when we're thinking about AI agents, it's an essential step to be thinking about unifying your data as well.
[00:25:35:02 - 00:29:15:06]
Amith
Yeah, and I think you said it well, Mallory, that having an employee, no matter how smart, locked in a room without access to the tools they need, they're not gonna get too far. And they're gonna be pretty frustrated too. So ultimately, yes, you can utilize AI agents without unified data, it's just way harder for them to get their job done. It's also extremely difficult for you to apply governance around it, so you don't necessarily know how they're accessing data because they're accessing a bunch of individual APIs to get to one system or another system. You don't typically have logging or audit trails tracking what they're doing when you're doing it this way. So it's really, really helpful to unify your data so that your AI agents don't have to worry about that. You'll also get better performance, better consistency. The other thing that's really key about data unification isn't just the idea of co-lodicating data physically. Like imagine if you had a bunch of boxes and you had three different storage units all around town and you had 100 boxes in each of those storage units, just by taking all 100 boxes from each unit and putting them in a bigger storage units, now you have them in your garage, you have 300 boxes, you've unified your data, but you haven't necessarily made any sense of it. And so the idea of an AI data platform is you have the AI look at those 300 boxes and help you organize it, to help you make sense of it, to help you first of all, take an inventory of what's in each box and what does it actually mean? Because after a while, a lot of times, what might've been once a carefully maintained spreadsheet or database, no longer is understood because there's been staff turnover, it's been just, you know, time has passed. And so how can AI look at this data and determine what it is, catalog it, build what we call a semantic layer, which essentially is a layer of metadata that understands what's in the data. And that's a big part of data unification. It's not about changing the data, by the way, you're not saying, hey, we're gonna go through those 300 boxes and we're gonna painstakingly reclassify or recode all the data in those boxes I'm talking about, but rather you're leaving those boxes alone, you're just looking inside and saying, hey, what's in these boxes? And how do these boxes relate to one another? And so when I need to look for something, I know that I have a box that has this information and I know what's in it, and I know how that box's contents relates to the other 299 boxes. That's what this so-called semantic layer does. Any good data platform will have something like that. In MJ, we just call it the semantic layer, and it's just something that happens automatically when you ingest your data into the platform. But with that, just like a human, right, if you have a really smart human and you say, hey, do work for me, here's 300 boxes full of really great stuff, you will find all sorts of treasures in there, but you tell them nothing else. That's like starting in a lot of companies, right? And they just have to like go box by box and figure it out. That's if they're super motivated and you give them a lot of time. Most of the time they're just frustrated and they can't find anything. But instead, if you give them this really great Rosetta Stone or a roadmap that says, this is exactly what all these boxes are, how they interrelate, when you need this, go here, when you need this other thing, go here, here's how these different things connect. If you have that type of semantic mapping and layer, then the human, as well as the AI agent, can really do some cool stuff. That's the missing link in a lot of AI agents people are trying to deploy right now, is they're trying to stick them directly on top of APIs to their AMS or even like a data warehouse that someone built for them that doesn't have what I'm describing as an auto-maintained, auto-generated semantic layer. If you don't have that, your AI agent won't get too far, unfortunately.
[00:29:16:08 - 00:29:34:17]
Mallory
I feel like we may have lost some people here talking about AI data platforms and they're thinking, oh no, no, no, we just replaced our AMS, that took years. We cannot put another system in place in this association. But can you just briefly talk about how an AI data platform is not replacing any of your systems and how it doesn't take that long to implement?
[00:29:34:17 - 00:30:59:15]
Amith
Yeah, it's just building on that example I gave of the physical boxes. I've imagined the 100 boxes from your AMS, the 100 boxes from your LMS and another 100 boxes from your CMS, your website in what I'm describing. And so what we're saying isn't that you're literally moving them, but that the AI data platform is able to essentially create a copy of all those things and then understand them, but it doesn't touch the origin system at all. It doesn't modify your AMS, it doesn't change where you do membership renewals, it doesn't mess with your LMS, it doesn't change where you deliver courses online. It is essentially taking a snapshot, like taking a photograph, but continuously of all of these systems and then being able to do the stuff I described. So it's very lightweight. It is not like an AMS implementation that is, quite frankly, is a little side note. If you're considering replacing your AMS because you think you need to do that in order to pursue some of the stuff we talk about, I would strongly urge you to reconsider because AMS migrations tend not to be transformative events in terms of capability growth or certainly AI capabilities. Even if the new AMS has some AI capabilities in it, it's gonna take you a long, long time to do that, a lot of money, and it's oftentimes the only thing you can do while you're in that type of migration. If you're midway through that swing right now and you gotta complete it, I get it. But
[00:31:00:20 - 00:31:38:06]
Amith
those types of major systems changes to your point. Nobody wants to do them, they're really painful, they're really expensive, and they take all of your energy and put it into that one bucket for a period of time. So AI data platforms are nothing like that. Quite frankly, Mallory, personally, I would want nothing to do with any of those projects because I spent 23 years implementing AMS's with associations and it's important work. I admire the people who do that, but it is at the same time something that is radically different than when we talk about implementing an AI data platform literally within a handful of weeks. You can get something up and running that's delivering value to your association.
[00:31:39:07 - 00:31:59:08]
Mallory
You can say that to me. 23 years is a long time to have worked in that space, so that gives you the credentials. But I was just thinking of anybody listening ever sees a Meath at an in-person event and you're like, "Oh, I want to strike up a conversation with him. I'm not sure what to say." Just go up to him and say, "Hey, I think we're gonna replace our AMS. We think it's like the most important thing we should do at our organization." And a Meath will run with it.
[00:31:59:08 - 00:32:02:03]
Amith
I might just turn around and walk in the other direction.
[00:32:02:03 - 00:32:57:02]
Mallory
No, you couldn't resist, I don't think. I talked about system access. I also want to talk about another component of agents, which is tools. So what the agent can actually do. This is the line between something that advises you and something that works for you. With tools, we talked about this earlier, an agent can send the email, update the member record, register someone for the event, pull the report, process the refund. Without them, it's back to being that consultant who has great ideas but no keys to the building. For a minute, hooking up an agent to your systems meant custom work for every single connection. That has changed with something called MCP, or Model Context Protocol, an open standard originally created by Anthropic, which is the company behind Claude, and now adopted pretty broadly across the industry. The short version is that it's a common plug, so a system gets connected once and any AI that speaks to the protocol can use it.
[00:32:58:06 - 00:33:11:24]
Mallory
I mean, tool access sounds great in theory, but once an agent can take action, it can also take wrong actions at machine speed. How do you recommend thinking about the human checkpoints within the whole world of AI agents?
[00:33:11:24 - 00:35:09:21]
Amith
A couple different things. First of all, you have to have a system that has governance built in. Governance consists of both controls, where you control which agent can do what, but also something called observability, or just tracking what happened. So you need both sides of that. Both critically important, so you need to make sure that you have really robust controls, where you have fine-grained control over what an agent can and cannot do. It needs to be very clear and very easy to understand. And then you need to have logging or observability, where you're tracking everything that an agent does. Now, what you want to start off with is, just like a brand new employee, you typically wouldn't give them everything. With a brand new agent, you want to start it off with a couple of things maybe, and let it get to work, see what it does. Optimize as you go. There's the four components you talked about, the instructions, right? Which is kind of like the prompt that goes into the agent's brain. That's something you continually are refining, and you'll refine it for a long time. An element of that context, by the way, is the memory of the agent. All of the good agent frameworks, MJ included, have really robust memory systems now. And the way to think about that is, it's like a Groundhog Day thing, the inverse to that, right? So you have an employee that shows up every day and doesn't remember what they learned the day before, it's kind of a problem, right? Even if they have tools, even if they're brilliant, not gonna be very helpful. Agents need to learn over time from their interactions with you and with other agents, and that's what memory systems allow them to do. Even though the models themselves, as a side note, aren't changing, the models don't accrue memory, the system does, the agent system learns over time. So for example, Member Junction has a memory system that has multiple different facets, including knowing what the organization's preferences and requirements are also at the individual level. So if it's agents talking to Mallory, it might have a different set of preferences loaded compared to talking to Amith.
[00:35:10:23 - 00:36:18:02]
Amith
But memory is an important part of it. But to come back to your question, I would start off slowly. I would give the agent limited capabilities and I would watch what it does. I would initially have it come back to me for anything of any significance. To me, that would be touching a member in any way, like communicating with a member, sending a renewal notice, sending an email. And this is not forever, this is just initially. Kind of like with a brand new employee, you might check their work, everything that they do for the first few weeks or something like that. And once you feel pretty comfortable that, yeah, they're producing pretty consistent output, you might loosen that control. You might say, you know what, I'm gonna give you the ability to send that email yourself. And you might tell the agent, hey, I want you to double check it, but then go ahead and send it. So you've now added a step where they're able to maybe even send the email to an editor agent that is independent from the agent itself that gives it feedback. And if the editor approves it, then the agent has the right to send it, right? So there's lots of ways of structuring this, but start off slow and then plan to grow over time, just like you would with a human employee as you see the agent perform well.
[00:36:18:02 - 00:36:39:15]
Mallory
For the person listening that's saying, there is no way there's a model context protocol connection for some of the ancient systems that we use at our association. For a minute, that might've been an issue, but with computer use, maybe it's not. So can you talk a little bit about computer use and how that works into this agent idea that we're talking about?
[00:36:39:15 - 00:39:37:19]
Amith
Yeah, you know, I was just at ASC and I asked a number of people I know in the AMS community if they have MCP support for their AMS, and most do not yet. So, but that's actually okay because most of them do offer APIs. And so an API, which stands for Application Programming Interface, is just a way for computers to talk to each other. And an MCP is simply kind of like a more modern version of that, if you will. And the reason MCPs are important is because they're just standardized in a way that allows them to be more pluggable, as you described. If a vendor you use has an API available, and that API is sufficiently robust, meaning it has enough functionality on it to do the things you wanna do, you can go to a tool like Cloud Code, or Codex, or a number of other coding agents, and you can have it create an MCP for the API, literally in minutes. It's extremely simple to create an MCP that wraps the API. So that's a little bit of a technical subject, but that is totally straightforward. This is not something that should cost tens of thousands of dollars with a consultant. This is something that should take maybe a few hours for someone to set up for you. So MCPs are trivial once you have an API, and most AMS's and LMS's, even the really archaic ones at this point, typically do have some sort of API. If they don't, coming back to your question about computer use, let's say you have a really archaic system, there is no API, and it's like a green screen on a terminal from the 1990s or 1980s even, and there's no API whatsoever. Mallory's point is that computers are actually getting pretty good at using computers. So AI systems are capable now at an extremely high level of actually navigating the desktop, navigating the browser, even being able to see what's on an old school green screen terminal to be able to type, to be able to click on your behalf. They don't physically have a mouse and a keyboard, but they're able to interact with these computer tools using simulations of keyboard presses and mouse clicks. And so you can try this out for yourself. If you're a cloud user, download the Chrome plugin, log in from the Chrome plugin, and ask cloud to navigate around the web and to do something for you. If you're already logged into Amazon, for example, say, "Hey, I need to buy some whatever," and watch it actually do it, you can actually authorize it to go all the way through. Clod is very risk averse by design. And so Clod will ask you for a bunch of approvals. There's some other tools out there that are a lot less concerned about that. And you can actually do this through Clod Co-Work. You can do it through Chat GPT Work, which is their agent products. These tools all have computer use kind of baked into them now, and it's getting better and better and faster and faster. So to Mallory's point, even those systems that seem like they are destined to not be connected to AI systems can be connected to AI systems through computer use.
[00:39:39:08 - 00:40:03:01]
Mallory
And we've spent this whole episode talking about AI agents. I'm curious for someone listening who would like to do some more work on AI agents within their association, but wants something relatively low risk that also packs a punch. When you're thinking about an AI agent in that context for associations, where does your mind go? What is a low risk agent that could actually drive impact?
[00:40:04:07 - 00:42:52:07]
Amith
I would start by thinking about the business outcome you're looking to affect. So for example, many associations are burdened by a tremendous amount of highly repetitive inquiries. Members or prospective members, emailing them or calling them, asking 80, 90% of the questions might be FAQ type things. They're not unimportant. They're extremely important to the person asking, but they're not really like high difficulty things to answer. And yet it's often quite common to hear associations say, our goal is to respond within one business day or two business days. And oftentimes they don't hit that goal because they're so inundated with so many emails, particularly during peak parts of their calendar, right around annual conference season, for example, associations, small teams get crushed. And so how could you potentially relieve that both simultaneously improving the quality of service and the speed and also relieving your team of having to do so much of the redundant work. So this member service challenge is what I would call it is a category I think is worth spending some time thinking about and saying, could we perhaps do something in this area to reduce the manual workload to semi automate this category and even taking a subset of it, perhaps taking a particular slice of these problems. And again, simultaneously affecting a pain point for members, which is to improve responsiveness and quality of response, and also to reduce the time required. Obviously this particular use case is one I've thought about a lot. I've talked about a lot. We actually have an AI agent called Izzy that literally does all of that, but you can also build this yourself. You can utilize a cloud cowork or a chat GPT work to prototype something like this. You can then build something that's more enterprise level on your own if you want. You can look for products like Izzy. There are other products in the world out there like Sierra and Decagon that are commercial products outside of the association space that are similar. There's a lot of options. Member service to me is one domain that is incredibly obvious to focus on because it's such an incredibly painful experience for the member a lot of times, and also for the staff. I really, really empathize with these team members at associations who are trying their best to keep up, but they have these asymmetric periods of time in the calendar where they just get destroyed by thousands of emails coming in. So I think this is an area AI can help a ton. That's just one example, but I think that's one that could provide an immediate benefit for everyone.
[00:42:52:07 - 00:43:21:18]
Mallory
Yep. I feel like that 100% resonates with our audience. Well, today we've learned that an agent is a model plus instructions, plus knowledge, plus tools. And the fastest way to understand it is to think just about onboarding a new hire. What makes it an agent rather than a chat bot is the loop, the planning, the acting, checking its own work and adjusting until the job is done. Agents increasingly work in specialized teams, each running a loop on its own. And what determines whether any of it works for you comes down to what your agents can see across your
[00:43:21:18 - 00:43:27:05]
(Music Playing)
[00:43:37:23 - 00:43:54:22]
Mallory
Thanks for tuning into the Sidecar Sync podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in-depth AI education for you, your entire team, or your members, head to sidecar.ai.
[00:43:54:22 - 00:43:58:03]
(Music Playing)