Chris Strom:
Hey, everyone. We're back with another episode of our RevOps Hero Podcast. The topic we'll be talking through today is a super timely one. We'll be talking about context for AI agents.
That's basically a fancy way of talking about how to train AI agents to understand your company's messy systems and tribal knowledge. My guest today that we'll be talking about that with is Ronnie Duke. Ronnie's been starting to get into a lot of projects in this area and build out solutions in this area. So he has a ton of knowledge to talk through in this very relevant and very timely and much needed topic here. So, Ronnie, thanks for joining us on this show here.
Ronnie Duke:
Yeah. Thanks so much for having me, Chris. Great to chat with you, and looking forward to our conversation.
Chris Strom:
Yeah. Me too. So we can start off with talking about your background, how you got into marketing operations and revenue operations to start with.
Ronnie Duke:
Yeah. I've been in the industry for roughly 20 years now. I actually started off in graphic design and website design. I owned my own agency for about eight years, and I ended up selling it to a software company in California about 2014, and that was kind of a transition time for me. So I joined this company, and they had just onboarded Marketo, and they've been wanting to onboard different demand generation programs, lead generation, lead scoring.
So they really had nobody to use this platform. So they kicked it over to me and said, "Hey, Ronnie, go figure out how to use this thing." That was my first sort of dive into marketing operations and working with some of these systems. And I'd always done a bit of email marketing, general-purpose marketing with my agency before, but I really fell in love with it.
I loved the operational aspect of making data work, getting things from point A to point B. There's different sections along the way where you capture the lead, you score the lead, and you send it over to sales, and then you're watching to see, "Does that deal close or not?" And it was always really fascinating to me. And so, I kind of launched my marketing operations career from there.
I worked for a couple of different agencies, which was great, because I got to work for tons of different clients that would do things different ways. I got lots of exposure to different methodologies and ways of doing things, and really quickly versus doing things internally. And then about 2019, I got my first in-house role doing marketing operations, led and built from scratch marketing operations function there. And since then, I've been working at various companies, mostly high-growth technology, B2B, building different solutions for marketing ops and RevOps, and recently decided to go back out on my own and to take the knowledge that I've learned all over the years.
And also, in this new age of AI, one of the things that I learned from my agency days was that whatever new technologies or new processes are evolving, it's always better to be on the front lines of that and be able to see how different companies are working with this new technology and things that they're implementing. So I really wanted to reignite that exposure for myself to go in and work with different companies as they're thinking about AI and AI readiness, and that's what I've been focusing my career on recently.
Chris Strom:
So speaking of that, let's talk about how specifically you and the teams you've worked with are starting to use different AI tools nowadays.
Ronnie Duke:
Yeah. It's interesting, because the way that I think about it, it's like if electricity were just invented, there's so many different things you could do with it, and it's fascinating to see how different companies are approaching their use to this new technology, and I'm seeing it in a couple of different ways. I think the first wave is more like efficiency gains, the busy work that people have really not wanted to do anyway.
They're the email checking, the writing documents, documentation, and things like that, and they're using AI to help assist them in these various things. And for their productivity enhancements, I think that there's a lot to be found about what that's actually getting for the company or the outcomes that they're looking for. But nonetheless, it's allowing freedom to do other things and trying to figure out what those are.
And there's also implementing AI into existing workflows where you maybe already have some automation set up through marketing automation platforms or sales ops, and you're trying to do these reach-outs to AI to help answer certain questions along the way or work with certain data and implement those into the workflow, and then there's the big agentic movement that people are trying to think about how to automate all of these different things on a regular cadence and allow these agents to be more autonomous within those workflows.
And I think that's an area that's still being learned and developed and trying to figure out what's going to work well for that, because as we'll talk about, there's a lot of context that is needed for these agents to be successful, because they're nondeterministic. They really rely on that tribal knowledge within the org to produce the outputs that you would expect if you were to do that job manually. So I think that's an area where I'm really excited to dig into, the infrastructure underneath the agentic workflow, and get companies ready for that type of work to be done.
Chris Strom:
Let's get into that topic of the problem and challenge of context for AI agents here. What a lot of people are starting out with is they sign up for Claude or something, and then they say, "Cool. I'll just connect Claude to Salesforce, and then I'll just start asking it questions," or "I'll connect it to Google Drive, and then I just start asking it questions." So what's the problem with just plugging in a tool into your system and going to town on it?
Ronnie Duke:
Yeah. It's funny, because I see a lot of this out there, and the capabilities are really cool with being able to plug these different things in with AI and do different things. But the way I like to think about it is, imagine if you were just starting at a new company, at a new role, and if someone gave you all logins to all the systems in their go-to-market stack, would you as an individual be able to understand the history behind what you've just gotten access to and how things got the way that they are or certain nuances behind how we work within the data in these systems?
That's a lot how these AI agents, the LLMs, operate, is they have access to all the data, but they don't really have the context behind the why. Why do these fields exist? What are they actually saying? Because as we know, working with these systems manually over the years, sometimes something simple like field naming conventions are kind of ambiguous.
I was at a company where there was a checkbox field on an opportunity called COVID-19, and nobody knew what that meant. It was just something somebody put there to track something. The humans that were involved in that, they maybe have a mental model of what that means, but the agent wouldn't know what to do with that. Every system that I've worked in, there's messiness. That's just part of the evolving nature of business, because business is very fluid.
I don't know that I've ever worked at a company where we've done the same exact thing, the same exact way for any period of time, and that accumulates in the tools. You're adding new fields. You're adding new workflows and processes and triggers and things like that, and they just continue to stack up. And as much as we want to spend the time to document everything and clean house and keep those things fresh, other priorities get in the way. We're always moving from one thing to the next and putting out fires in our roles, and that has always taken a back seat in the industry.
And so, what happens is, when you give an agent access to all of that, then they have access to all the messiness as well, and they don't have the full context of what to do with that. If you want to run a report that says, "What are our deals' close rate by industry?" But if, in your CRM, you have four different industry fields all with conflicting data between them, because they're coming in from different enrichment tools or however you have it set up, the LLM is either just going to pick one or it's not going to know how to actually interpret that data and use it.
And what it's going to do is, it's going to deliver you an answer sometimes without even telling you the methodology that it used. And so, you might get this really confident-looking answer to whatever you asked and have no idea the methodology behind it. So it can get you in trouble as a practitioner, because you're having to deliver these reports to other stakeholders, and sometimes you could go months on end with reporting off of wrong foundational data. And so, having that context is really important in building that knowledge base, that middleware in order for the LLMs to be able to interpret all of this correctly.
Chris Strom:
Yeah. So that's the problem there. I like how you described it as, just plugging the tools in is kind of like hiring a new employee at the company, just giving them all the logins, and trusting whatever they say after that.
Ronnie Duke:
Yup. Exactly.
Chris Strom:
That's the challenge that companies will need to address first, that they need to note that exists to begin with, and then the next step is how to address it. So let's talk about how to give the AI tools the context and the tribal knowledge that they need to work correctly.
Ronnie Duke:
Yeah. I think a lot of it boils down to some form of documentation one way or the other in a way that is clear and noncontradictory to itself. A lot of teams, they store documentation in cloud-based solutions like Notion or Confluence, but there's no real governance behind what gets posted in there, and you may have two different people posting two different documentations about something that don't agree with one another or they have contradictions between the two of them.
And so, it really becomes an exercise that the entire team needs to rally around. How do we create a source of truth for these different things, so that way, when we do enable our LLMs to read this data and process things on our behalf, that it's doing it in the way that we want it done consistently every time? So I think there's an element of structural documentation around the tools.
So that could be background knowledge around the flow of information that comes into every person's stack. What tool does this data go into? What data does it transform? What data does it create on its own versus gets inserted from another platform, and what happens there? So if you were to take and draw out a tool mapping diagram and a flow of information, I think getting that into written form in a way that the LLM can understand, I think, is step one.
I think step two is, if you were to pick a thing that you wanted to automate or get that sort of 10x multiplier with AI, what is the process that you would do it at 1x? Can you describe it in a way that would be reproducible every single time the same exact way? And a lot of times, that's a very hard thing to answer, because there are new... You can take something simple like building an ABM webinar.
I think most of the ingredients are repeatable for most teams, but there's always some sort of gotcha or different thing that the team wants to try and do something differently. That doesn't translate to an agentic workflow, because the LLM can't make those decisions on your behalf. It can make decisions, but it's not always going to make the right one based on the variance that you're introducing in this workflow.
So it's really a matter of trying to document the process that you would want done the same way every single time, and that's going to be the hardest part for a lot of teams, because they don't always know what that is, or they don't always do it the same way. And so, I think the joining of that together of, how would we want this done if we were to do it manually? What are the steps involved? What tools need to be involved? Where does the data come from?
And then the third thing is, are our platforms even ready to support this? Because you might have a process documented of, "Okay. We go to this tool. We create a webinar. We sync it to this other tool." Do these tools that we're using have the right APIs or MCPs or ability for an agent to even go and perform the task that you even would want it to do? And sometimes the answer is no. And so, you have to understand, where does our martech and go-to-market tech environment sit within these capabilities, and can it actually be involved in this workflow if we were to leverage AI to do it?
Chris Strom:
Yes. That's a lot of the process knowledge, and then can you talk a little bit about even just the systems operations knowledge, like glossaries or data dictionaries, defining what CRM fields are even supposed to be used or not used?
Ronnie Duke:
Yeah. That's a great point. And again, it comes back to that context of, we create fields for things or areas in certain applications for things that don't always... You can't always tell what it is just by looking at the field. And so, creating some sort of a mapping dictionary that says, "This field is used for this purpose," in any kind of context behind that, like, "It was introduced in late 2025. So any data before that is unreliable." These are all nuances that the LLM is... They're very good at leveraging that in context at the time, but you have to have it written in some way for them to be able to know, "Okay. I shouldn't report on anything before this date if I'm using this field."
And so, yeah, it's having that away, and also just keeping it up-to-date, keeping it fresh with any changes that you do make to it. A field was populated by ZoomInfo, and now it's populated by Clay, or there's different things that we recycle fields sometimes, and there's cutoff points and there's annotations that can be useful for the LLM when it's reasoning and making those decisions, that it can incorporate that historical context with the data that it's working with.
Chris Strom:
Oh, yeah. That's a good point. I was just thinking of definitions and glossaries, but you mentioning the point of also including the backstory on these things too. I hadn't even thought about that part.
Ronnie Duke:
Yeah. And it's so common, because, again, these teams, they move really fast. I've been on teams that pivot their strategies almost every quarter. And even manually before AI, the questions were hard to answer, because people want to look at historical trends, "What are our MQLs this year versus last year at this time?" But that question is hard to answer when the motion behind the MQL or the criteria behind that changed three different times throughout the year. And so, documenting what has happened and what has changed is very useful, because it adds more clarity to the question being asked.
Chris Strom:
So that's a lot of the different types of knowledge and context you need to start documenting. And then where should companies store and organize and reference all of this?
Ronnie Duke:
That's a great question, and I think it's something that the industry is still trying to iron out. I'm seeing a lot of bespoke solutions being implemented today that depends on the size of the team and how the team likes to work. But the big idea is, when AI was being rolled out, it was very individualized. You had your own ChatGPT. You had your own instance of Claude Code on your laptop, and these things didn't connect in centralized repositories.
Your conversation that you're having with your LLM is private between you. The outcome of that isn't automatically in a place where somebody else can piggyback off of that and reference that context. So it really becomes a game of, how do we take both the inputs, all of the structure, the data structure, the glossaries, the context, but then also the outcomes? When I run a report or when I come to a conclusion or develop some sort of insight, how do I put that as an artifact that other teams and their working environments can access?
And there's low-tech ways of doing this. There's GitHub repositories, but that tends to lean a little bit more to technical audiences, and there's things that people would have to remember to, "Oh, I have to do a Git pull," or "I'll push every time I make a change so I keep that current." I'm seeing teams think about maybe a shared Dropbox or a Box account where the files in those folders continuously sync so the teams then have access to that.
I'm seeing cloud solutions that are starting to break out, and maybe that'll be a way to centralize things, because one of the other aspects of that is access. So something we often overlook is, when we're on a team and we're part of a company and we have our SSO logins to the data warehouse or to the marketing operations platforms or whatever, those are all unique to every individual.
So if you were to think about an agent, whose login are they going to use? So you're starting to think about having agent login accounts that have access to this, and where are those stored? And so, I think that's an industry that is ripe for development, is to, how do we manage that shared context, but then how do we do it in a controllable and scalable way?
Because anybody that's worked with localized AI, the Claude Code or Codex or even Claude Cowork, AI can generate a lot of stuff really fast, and that comes a lot of documents, a lot of files, and people thought a Confluence space gets messy. Imagine somebody's individual Claude folder constantly being updated and shared and everything. So I think that's an area that needs to be carved out in a way that teams can do that efficiently.
Chris Strom:
Yeah. I agree that's going to become a bigger and bigger issue going forward.
Ronnie Duke:
Yup. Exactly.
Chris Strom:
So we've been talking here about how to plan and start building out that knowledge layer and the context layer for agents to use, but that's just the beginning of the journey. So I'd like to hear from you then the next step on, how should you plan out how to keep it up-to-date and continually refreshed and current as situations change going forward?
Ronnie Duke:
Yeah. It's a great question, and I think it kind of dovetails from the last one, is getting that shared context in a place that multiple people can access it, but then how do you maintain governance over who gets to update that knowledge base? If there's, let's say, a data dictionary or data glossary, how do you lock down who has the right to change those definitions or add that additional context?
You have to get your team on board with how this is going to be governed. I think in the future, as more solutions get developed, then we can start thinking about technical governance behind this, like actual role access to this information wherever it's stored and who has the granular abilities to create artifacts and reports from the outcomes versus who has access to change those core definitions, those core files. And so, I think it's something that teams need to start thinking about now, because it's an evolving problem that people are going to need to solve.
Chris Strom:
What do you think people should think about in terms of a cadence or a schedule of updating the context and documentation?
Ronnie Duke:
I think that's sort of the process path, is that if a team can agree, "Hey, once a week or once a month, once a quarter, whatever that cadence is, we all have our responsibilities to make sure this is up-to-date," or build in processes to where it self-updates as you go. That's one of the beautiful things about some of this AI technology, is that you can build skills and repeatable patterns for the actual usage of it.
I use a lot of projects where I've built in my own governance into the memory and the logic of the folder itself. So with Claude, for example, I can tell it, "Hey, whenever I say something that changes one of the to-do list items or the documentation of this, go in and actually update the documentation to reflect..." Or at least ask me, "Hey, should I go in and make a change to this reference?" So it's sort of self-healing as it goes.
And I think there's not a blanket recipe for everybody to use, but I think just the knowledge of what these LLMs can do in those contexts, I think, is helpful, is that, "Hey, build in your rules for when I ask for something, or when I change something, ask me back." "Hey, should I update this in GitHub, or should I update this data dictionary definition for this and build in processes that way?"
Chris Strom:
Oh, yeah. That's a great idea. Basically instruct your agents and tools themselves to be proactive in updating the documentation as they go rather than just relying on your own memory to do it or relying on some sort of predetermined schedule.
Ronnie Duke:
Yeah. I think the reality is, if you think back even before AI, what was your cadence on updating documentation in whatever tool you were using at that time? And a lot of it wasn't frequent. I think everyone had the intent to update these things, but these are just things that always fell through on the back burner.
So those problems didn't go away with AI. And that's one of the biggest things that I try to stress with my clients, is that you can't just point AI at problems that preexisted before the technology was there, and these are process governance things that teams are going to have to adopt, like, "Hey, we need to be rigorous about what tools we're updating, what tools we're giving access to, what core knowledge base we rely on to make this function." And that job doesn't go away with AI.
Chris Strom:
Yeah. In fact, it'll probably expose a lot of things that had been sitting in your documentation and systems for a long time that you had just forgotten about.
Ronnie Duke:
100%. And one of the clients that I was working with recently, on day one, I started connecting tools to their go-to-market, to their HubSpot. And when I approached the knowledge-based original design, I say, "Take an inventory of all these fields. Find me any kind of conflicts or things like that." And there was fields that were contradicting each other. There was all these things.
And so, what I was able to do though is create a formalized questions document, bring it back with the client, and he was able to go in and say, "Oh, yeah. This is something I started but never finished." And so, in real time, we start providing that context. I'm like, "Great. I take that feedback, and I bring it back to the LLM." And it is a back-and-forth that has to be done. Everyone likes to think about trying to one-shop things with AI. This is a build process, and we can get it to a state where it understands all of the different pieces that are involved.
Chris Strom:
Yeah. Even more so than just traditional code-based tools. Sometimes, or almost all the time, the training never ends.
Ronnie Duke:
Yeah. Exactly.
Chris Strom:
Well, as we're talking about people documenting their knowledge and their processes and everything and writing it down for the purposes of training AI tools on this, that leads into the fear that is on a lot of people's minds, which is, "Am I replacing myself with the AI tool by just giving it all the knowledge I have?" So I'd love to hear your thoughts on this question.
Ronnie Duke:
Yeah. It's certainly a hot topic. Right? Everyone's worried about AI replacing their jobs, and I think the reality is a lot less scary than the headlines make it out to be. And when I think about go-to-market specifically, I can't think of a single time in any client or business that they have done the same thing, the same way for any solid length of time.
And when the business pivots their motion or strategy or adds new features or changes anything, that all trickles down into all the things that produce the tasks that are involved, and AI is really good at executing tasks. It's not going to be able to keep up with your strategy or your own nuance from your teams on how you're changing the business.
And so, there's always going to be a human element to doing that. What is changing is that the humans are doing less of the task work that the AI is now starting to do on our behalf. And look, if your role is isolated to just doing a single task every day, then yeah, there might be some changes or some displacement that might happen.
But there's going to be other roles, I think, that open up to help counter that. I would not trust AI to just set it and forget it and just run my entire go-to-market, all my ads, my demand gen campaigns to just automatically do everything. It's going to need some level of oversight, course correction, quality assurance. I think some of these jobs are just going to change and evolve to work with that. I don't see a world where AI is completely replacing every human aspect of the things that these teams are producing.
Chris Strom:
Yeah. I'm in agreement with you on that. What the specifics of our jobs themselves look like, that's really hard to predict, but I think there'll be something for people as long as they're willing to adapt, of course.
Ronnie Duke:
I agree.
Chris Strom:
But then, like you said, if your job is just doing one specific task all the time and nothing else, you might be kind of happy for the chance to do something different than just that one same thing all the time as well.
Ronnie Duke:
Yeah. I'm not going to say there's not going to be disruption in some capacity or another, but if your job is checking emails every day for somebody, then AI is certainly going to be able to handle that that doesn't need a human. But what does that do? Well, that forces you to pivot your skill set into doing something different or maybe using AI to learn a new skill set, or there's so many different vocations that I think that are going to start to come out of all of this.
And it's like, when social media came out, nobody knew what a social media manager was 20 years ago. That wasn't a thing, and these are different roles that we just don't even know exist yet, and I'm excited to see what that turns out to be.
Chris Strom:
Yeah. I'm excited too. Well, I think this has been a great conversation on a topic that basically everybody needs, but a lot of people haven't even started to become aware that they need it yet. So I'm really glad we could take this chance to talk through this to show people what they need to know and how they need to start planning out how to document the context to give their AI agents so that they can actually utilize their messy systems and incorporate all of the tribal knowledge and institutional knowledge that companies didn't even know that they had in many cases.
Ronnie Duke:
I'm seeing two different worlds of people that are just getting started. They maybe just signed up for Claude and don't even know where to start, and I think that's a whole world of people, and I think there's another side of it of, they've been tokenmaxxing for the past six months, and now their CFO has started to ask questions around, "What is this money that we're pouring into this actually getting us?" And I think both of those camps could benefit from some consulting on where to go with their AI journeys.
Chris Strom:
Yeah. I agree. Thanks for taking the time to talk through all this with me, Ronnie. I feel like I've learned a lot here, and I think everyone who watches and listens to this will learn a lot themselves as well.
Ronnie Duke:
Yeah. This was really fun. Thanks so much for having me, Chris. Yeah. I really enjoyed the conversation.
