Subscribe
Share
Share
Embed
Future-proofing the humans behind the tech. Follow Phil Gamache and Darrell Alfonso on their mission to help future-proof the humans behind the tech and have successful careers in the constantly expanding universe of martech.
[00:00:00] Phil: Garbage in, garbage out. AI is only as good as the data. You kind of feed it like everyone says, you need to fix your data. almost no one talks about what it actually means in practice to fix your data
[00:00:09] Jason: not the best label. it kind of makes this effort sound like some mythic authoritarian quest for the Holy Grail. Uh, you know, the issue is that teams jump from, Hey, here's a big pile of our data to, let's throw this agent on top that should now be able to like, act on any scenario without defining that layer in between.
[00:00:28] Darrell: what's the list of non-negotiables? You have to start with upfront.
[00:00:31] Jason: First I think. What's very important? Shared definitions, Do we mean the same things by what is pipeline or what's a qualified lead second, trusted access. can the system actually see the right records with the right joins, third, I think named ownership. you know when something goes wrong or needs to be fixed it needs to be a person or maybe a. Few named people, fourth, authority boundaries.
[00:00:51] what's the system allowed to do on its own and then fifth is you need a way to test whether the output's actually good enough.
[00:00:58]
[00:01:24] In This Episode
---
[00:01:24] Phil: What's up everyone? Today we have the pleasure of sitting down with Jason Dobbs, head of marketing and GTM Engineering at Kumo Before Kumo, Jason's held marketing ops, rev ops and lifecycle roles at big brands like Logitech, Seagate, apple, and Google. In this conversation, we cover the five and non-negotiables for AI readiness in marketing ops. When to ship AI before your data is ready Why marketing Ops is the context architect And we'll also get Jason's take on what the heck GTM Engineering even means. All that and a bunch more stuff after a quick word from two of our awesome partners.
[00:01:57] Sponsor: MoEngage
---
[00:01:57]
[00:02:54] Sponsor: Knak
---
[00:02:54]
[00:04:00] Phil: Jason, thank you so much for your time today, sir. Really excited to chat.
[00:04:04] Jason: I am so pumped to be here. Uh, really love, you know, the work that you know, you Phil, and you Darrell, are are doing in the community. I've been watching from the sidelines for a long time, um, you know, cheering you and, and everyone else on. Um, and yeah, like I said, so excited to to be on the show today.
[00:04:22] Phil: I really appreciate that Jason means a lot coming from you. Uh, so yeah, like we were chatting uh, a lot before, uh, you kinda came on here about what's the main topic gonna be here. And we kinda landed on this idea of AI data readiness. And
[00:04:35] 1 — How Undefined Data Definitions Make AI Confidently Wrong
---
[00:04:35] Phil: one thing a lot of marketing ops people talk about when they come on the show, as soon as we mention AI is.
[00:04:41] Garbage in, garbage out. AI is only as good as the data. You kind of feed it like everyone says, you need to fix your data. We've beaten that horse to death. Uh, almost no one talks about like what it actually means in practice to fix your data. Like who owns it, where do you start, how much is enough before you move to.
[00:04:57] Prevent AI and agents from, uh, you know, just [00:05:00] spitting out chaos in, in your system. Maybe we can start with that like chaos element, because you said before we recorded that you learned this the hard way at your current startup. So anyone that's using LMS on a regular basis, we've all had our first encounter with an LLM being confidently wrong.
[00:05:15] And at Kumo you said that you greenlit ent predictive workflows before the data was ready for it. Can you take us back to that moment, Jason? Maybe we can start there.
[00:05:24] Jason: Yeah, a hundred percent. And you know, I love the example of, of leading, um, with failure. Uh, you know, because I know, you know, for, for stubborn. People like myself anyways, uh, failure is, is unfortunately the best, uh, and quickest way to learn. Um, you know, going back to to, to some of the, you know, we talked about some of our personal interests, um, you know, kind of over the show, you know, I ride these high speed, you know.
[00:05:48] Um, F1 skateboards ago, 50, 60 miles an hour. And yeah, you know, I've crashed or, or fell off, fell off a, a few times, but, you know, I can tell you I've never crashed or, or fallen off the same way twice. [00:06:00] Uh, so I try to, to take that same philosophy through the rest, the rest of kind of what we're talking about here.
[00:06:06] Uh, you know, and especially, you know, the age of AI that we're in, you know, everything's moving so fast, there's so much. Um, testing going on, new products coming out, new models coming out, uh, almost every week. Uh, you know, it's a constant, uh, process to go through that. Um, so. My own experience there, you know, talking about green lighting, kind of agentic, um, flows that weren't quite ready.
[00:06:29] You know, it can essentially be dangerous and, you know, what made it dangerous wasn't really like the obvious hallucination that we all see or used to with lms. Obvious hallucinations, easy. Um, you kind of reject it, you move on. I mean, I think what, what we were seeing looked polished enough to be operational.
[00:06:45] Um, at first glance, you know, the scores looked precise, the summary sounded coherent. Uh, the recommendations felt data backed. But the moment you ask, like the simple follow up questions, like, uh, and that's why explainability is so important. Like, okay, [00:07:00] well why did you choose this account? Or, um, why is that happening now?
[00:07:05] Or what data drove this decision? Um, you know, the logic started to thin out. So that's really the layer, the, that you need to, to be able to, to dig into. Um, so, you know, the lesson for me wasn't like the data was bad or the warehouse was bad, or the model was bad. Uh, what's, what, what was wrong is we were trying to automate ambiguity.
[00:07:25] Um, you were, we were asking a AI to solve for confusion that we hadn't yet ourselves solved for, uh, internally. Um, and, you know, and once you do that, you kind of enter the danger zone because, you know, the, the failure is essentially believable, nonsense. Uh, and that, and that's dangerous because people trust it.
[00:07:43] Um, if you, if you don't know what to look to look out for, so you know what the top of me is, you know, trusted data, shared definitions. Guardrails, uh, workflow ownership, that's not the boring back of office work. Um, especially, you know, that's a lot of the, the lifeblood, uh, the, [00:08:00] you know, m ops folks, ops folks bring the table.
[00:08:03] That is the product. Uh, you know, once the models are good and you know, the data's good enough, like those are the things that matter. Um, and analytics at the end of the day only matter, you know, if they drive a real life decision
[00:08:15] Darrell: Right.
[00:08:15] Jason: and impact.
[00:08:16] Darrell: Really good insight there. Um,
[00:08:18] 2 — Why Context Engineering Replaces Prompt Engineering as the AI Bottleneck
---
[00:08:18] Darrell: Jason, I wanted to ask you, you said in your pre-interview we were trying to automate ambiguity. We were asking AI to solve for confusion that we hadn't solved operationally. So, um, I wanted to ask you like, who actually does the data fixing on a team? Like, what's their roles and what does act, what, what does fix the data really mean?
[00:08:38] Jason: Yeah, great question. You know, I think. The fix the data may, may be, uh, not the best label. Uh, you know, it kind of, it kind of makes this effort sound like some mythic authoritarian quest for the Holy Grail. Uh, you know, it's the one quest we all seek out for, but clever never quite actually seem to [00:09:00] make it to the end.
[00:09:01] Uh, you know, regardless of, of kind of how good we are. Uh, you know, I think the reality is that. The warehouse is often where the richest relational signal already lives, or, you know, breakout kind of wherever. I mean, it's different for different organizations or wherever your core, uh, data sets live. Um, so the, you know, the issue isn't usually that the business has zero signal.
[00:09:22] Uh, I think the issue is that teams jump from, Hey, here's a big pile of our data to, uh, let's throw this agent on top that should now be able to like, act on any scenario without defining that layer in between. Right. So I think it's not really like how do we clean everything? Um, it's more like, what minimum context and control does a workflow actually need to become liable?
[00:09:45] So that includes, you know, things like what are the definitions, you know, what is the source of truth? Uh, what's the freshness of the data? What's, what is the permission that we're actually giving for this? Um, you know, defining those thresholds and then like figuring out like where [00:10:00] does human judgment still belong?
[00:10:01] Like, uh, in all this. So, you know, we talked a little bit about, you know, how prompt engineering can make something smart. I think contact engineering, context engineering is really what makes it dependable.
[00:10:13] Phil: Yeah, the, the context engineering part is, uh, like the, the hottest buzzword. Um, like, uh, Scott Brinker wrote a whole report on. On context engineering. And when we were trying to figure out like cool questions to ask you about that, uh, one thing that popped up in your job history is that you spent seven years as an air force intelligence officer before you went into tech, and some of your work included the president's daily intelligence briefing.
[00:10:40] And I had this crazy idea about like, Hey, why don't we ask Jason about. This job, like how it's probably one of the highest stakes content context assembly jobs in the world. Um, take us back to that time, like how much of what you're arguing about AI readiness and data quality is actually intelligence trade crafted with [00:11:00] like a different label.
[00:11:01] We chat about that for a bit.
[00:11:02] Jason: Yeah, you know, I actually love this example because I think it has, at least for me personally, formed, um, a lot of how I like tackle even, you know, business challenges. Uh, and, you know, we could probably nerd out and make an entire episode about. Uh, kind of the correlations between that also. But, you know, a lot of, a lot of what we are talking about is intelligence trade craft with a different system.
[00:11:25] I mean, an intelligence just like in, you know, operations, like you're almost never working with perfect information. Uh, you're building a decision picture from incomplete often, you know, contradictory, time sensitive inputs. So the real skill, not really collecting more information, it's knowing what to trust, what to discount, uh, what stale and you know, where uncertainties high enough that we need to invest more human cycles.
[00:11:53] Um, so in the intelligence world, just like business world, we have a lot of different data sources that we pull in. We call them, [00:12:00] uh, you know, intelligence collection sources, things like. Um, human intelligence, signals intelligence, geo imagery, intelligence, uh, measurement signature intelligence, open source intelligence.
[00:12:12] Uh, so, you know, that's, that's how I think about AI readiness too. Uh, on the business side, we have, you know, our CRM data, our product data, our, our engagement history, our all the support history, like tons and tons of these different data sources, you know, that by themselves don't really mean much, but you start to build that in, uh, you know, you, you bring all of the different, sort of connect the dots and bring all the breadcrumbs together.
[00:12:33] Um, and, you know, essentially that's, those are just enterprise versions of intelligent source streams also. So, you know, the job's the same. Assemble enough trustworthy context to make a good decision without fooling yourself. Yeah.
[00:12:47] Phil: We, we won't ask you a political question, uh, with the, uh, the, the climate today. We can keep that for another
[00:12:53] Jason: Yeah. No, a hundred percent. Yeah. Hmm.
[00:12:56] Darrell: So, um, we've been, uh,
[00:12:59] 3 — The Five Non-Negotiables for AI Readiness in Marketing Ops
---
[00:12:59] Darrell: so there's this concept of minimal viable readiness and, you know, I think that that's what a lot of folks are trying to achieve at minimum. Before layering on AI and agents, what do you think are the bare minimum requirements for you to feel like, okay, our d our data's in a good place, might not be perfect.
[00:13:18] Um, it'll probably never be perfect, but what's the list of non-negotiables? You have to start with upfront.
[00:13:24] Jason: Are you sure you don't wanna undertake the quest for the holy Grail, Darryl? Like it could be perfect.
[00:13:29] No, but actually I love this, uh, this term minimal viable readiness. I feel like we just need to put a stamp on this right now.
[00:13:36] Phil: Yeah, I actually didn't Google to see if this was like a
[00:13:38] thing or
[00:13:39] Jason: Phil. We'll let you take the credit for this one, man. Yeah, we'll just, we'll go ahead and put a trademark on it and move forward.
[00:13:44] Uh, but no, it's a great question. So, uh, you know, first I think. What's very important? Shared definitions, especially like, okay, it sells marketing and rev ops talking about the same thing. Do we mean the same things by what is pipeline or what's a qualified lead [00:14:00] or, uh, you know, things like that. So that's, you know, I talked a little bit about automating ambiguity.
[00:14:04] Like we were asking AI to solve for processes that we hadn't yet like defined ourselves. So this is where you can kind of get wrong in this. Uh, and if you don't have those shared definitions, you know, the AI's just gonna scale that disagreement essentially. Um, second, trusted access. So not. Like we technically have the table or the pile of data, but can the system actually see the right records with the right joins, with enough freshness to make the decision that it needs to for that task.
[00:14:33] Um, third, I think named ownership. Um, so you know when something goes wrong or needs to be fixed or like there's gotta be a human owner, not the model or the data team. Uh, it needs to be a person or, or maybe a. Few named people, um, look for a shared type responsibility for sure. Uh, fourth, you know, I'd say authority boundaries.
[00:14:57] So what's the system allowed to do on its own [00:15:00] and what still needs review or approval, um, from humans like us? Uh, and then fifth is really, you know, an eval path. So before anything gets real authority, you need a way to test whether the output's actually good enough. And, uh, you know what to do about that and, and reverse that if it fails.
[00:15:17] So that's really the heart of it. Uh, not perfect data, you know, trustworthy context plus a clear approval path. So, you know, I'd say for teams wanting to tackle this tomorrow, it's like, take one workflow, you know, score it red, yellow, or green on those five items, and that'll tell you faster than any strategy deck, whether you're actually ready.
[00:15:38] Darrell: Awesome. Awesome. No good insight there, Jason. Um, so
[00:15:42] 4 — Why Marketing Ops Is the Context Architect in an AI-First GTM Stack
---
[00:15:42] Darrell: let's move over to a topic. That's close to my heart, which is marketing ops. Um, I'm fascinated by the overlap between data engineers and MarTech and ops and, you know, it's, it's, uh, it's kind of like a blurry line between who does what. Um, because some of the work that we do is, is pretty technical and does require, you know, typically coding.
[00:16:03] Um, but when it comes to things like system administration. You know, you might not need a computer science degree to, to fix some of those problems. You know, it's more like, kind of like configuration. Um, and I wonder how you think about who tackles what problems when it comes to data. Um, is, is, is ops sort of maybe managing the project and you're looping in an engineer?
[00:16:27] Or how, how do you think about that? Maybe riff on that a little bit for us.
[00:16:30] Jason: Yeah, a hundred percent. Uh, and I think. You know, super important question, and I'm passionate for me as well because I came up in the lifeblood of, of ops. Uh, you know, we talked a little bit about my intelligence background, but my first career. So many moons ago. And, and tech was, uh, you know, on an ops team as a Salesforce administrator, like, uh, like that, that was kind of the lifeblood of coming up and then growing into becoming a data scientist and then into kind of business leadership and, and engineering side.[00:17:00]
[00:17:00] Um, and I think, you know, that's super important, uh, because especially with how we see. Just the, you know, how AI has overtaken the, uh, the ecosystem and how roles are shifting, especially between data engineering ops. Uh, you know, how we need to essentially think about that, uh, and scale up as, as individuals to get ready for kind of the, the next evolution.
[00:17:23] Um, that we're going through. So the role itself is, is getting more important, not less. You know, as like, as these human roles shift towards supervising outcomes, uh, as delegated systems become more common, I think in practice marketing and ops becomes the, the business side context architect. Um, the team closest.
[00:17:41] I mean, they are the team closest to the definitions, to the routing logic, to the handoffs, to the exceptions, and, you know, real world process integrity. Um, so, you know, talking to that, I wouldn't start with like a giant backlog of every bad field, every sync issue, every message table in the company. I think that's how teams get [00:18:00] stuck, uh, become very overwhelming.
[00:18:02] And start with one decision loop, right? Like map the loop end to end. Um, what source systems fed it, you know, what signal or prediction gets produced, uh, where it lands, you know, what action follows and who can stop it if confidence is low, you know, so that really kind of falls in line with some of the, uh, rest of the, the things we've been talking about.
[00:18:22] So that's where marketing obstacle comes incredibly important in my mind. You know, um. They're the team closest, closest to that. Uh, and, you know, in the real world process, that that's what determines whether a, a model actually creates a business value. So, you know, I wouldn't map that entire customer universe first.
[00:18:41] Like I said, start with one workflow, prove it matters, use it as a pattern and a blueprint, um, to kind of move on to the next one.
[00:18:48] Phil: Yeah, it's a cool call out there. I feel like, you know, fixing the data in the marketing ops world, we think of the CRM and the marketing automation platform, the customer engagement platform, depending if you're [00:19:00] B2B or B2C, and those are like systems that we administrate. We have access to them like it, it's kind of on us, but more and more orgs are moving to the data warehouse as the source of truth.
[00:19:11] And so these tools are kind of sitting on top of something that we don't necessarily have the keys to. And that's like the question about the data engineering role versus the marketing ops role. So I'm curious to ask you like, you know, in that world where maybe, you know, there is a data warehouse in the company, uh, hopefully there is.
[00:19:30] Um, and there's a lot of like the fixing the data things that fall out of our world in, in the marketing ops world. Um, like what are your thoughts on, you know, sometimes there is still a lot of stuff that marketing ops and MarTech folks can do themselves without having to. Submit a ticket with the data team and constantly have to like, bug the data team.
[00:19:50] Or sometimes like it falls under the product team. Um, you know, like marketing ops can't necessarily rebuild the warehouse or create new tables all the time on some teams, you know, we can, [00:20:00] 'cause, you know, there's a dedicated data engineer that is servicing the marketing ops team and sometimes like a blend of the two.
[00:20:06] Um, but you know, we're, we're not helpless in, in that world. Maybe like what can someone in this role marketing ops role actually fix, um, you know, without the data team kind of on their own this week, and where exactly does the line fall between, like where they need to go get a ticket and get that prioritized?
[00:20:25] So I kind of like. I came up with like three big areas of things that are like maybe higher level of, of things to, to go and fix this week or, or whatever. I'm curious, like if, if there's any in there that, that are missing, you got a chance to kind of like look at this first, but the first thing that comes up to me is obviously like de-duplication.
[00:20:43] Like we talk about this all the time. This is the most visible one that comes up, especially when you're like going through audit logs or just like. Quick searches and you find like three people with Darrell at Daryl Alfonso comes up seven times in the database. Like, what the heck are we doing here?
[00:20:58] Like, [00:21:00] ddu deduplication is, it's ours. Like it's, it's on us. Um, there's a lot of tools out there that help you. That you kinda like slap on third party tools. It duly clouding go whatever. Um, there even like a lot of maps have like a native D dupe feature. It's not glamorous work, but you know, that's one of the areas.
[00:21:16] The other one was like basic ID resolution. This one is kinda shared between the data team, the marketing ops team, because Id res between product and, you know, pre identification on the website. Like it, it kind of crosses paths with a lot of different folks. Sometimes it's us, especially like on the B2B side, like enrichment tools can kinda come in and help out a lot with that.
[00:21:40] And the third one I had is like the bane of every marketing automation person's existence. Like the sync issues between the CRM and the marketing, uh, automation platform. You know, there's a lot of configuration problems that aren't necessarily architectural problems that don't require a data eng team.
[00:21:57] Wrong field mappings, like seek [00:22:00] filters, um, bidirectional conflicts, all that stuff. What, what are your thoughts on those three? Am I missing any, just riff on that for bit.
[00:22:06] Jason: Yeah, mostly, you know, uh, I'm just reliving the stuff of nightmares and, um.
[00:22:12] You, you know, going through PTSD and trauma right now. Yeah. As you know, as someone who came up in the trenches of, of ops, uh, you know, I've lived that life for the last 15 years. Yeah, a hundred percent. But you know, my answer with that be is marketing ops is way more powerful than people think.
[00:22:29] Uh, you can fix a lot this week without a data engineer. Um, you know, moping ops can absolutely do real readiness work without waiting for huge data engineering project. Um, you know, like we talked about, improve duplication, enrichment rules, uh, field governance, sync mappings, routing, logic, suppression rules, lifecycle stages, qa, the, the real life systems that, that people are actually using.
[00:22:54] Um, you know of sure, maybe a lot of that work is not glamorous. Uh, but it's the lifeblood [00:23:00] actually of what a I sees, uh, and what the business trusts that we're in the background. And that's a lot of, like we talked about. Once the model's good, um, like those are the things that are actually gonna impact, like how successful, um, are these things, uh, within your organization.
[00:23:16] So, you know, where the line usually shifts to data engineering more, you know, the things, the, the other things, you know, the core IDs, warehouse modeling event schemas, you know, semantic consistency across systems, real time freshness or you know, of course those are some of the things that, you know, maybe, you know, there's a lot of things that, that, that, so like, you know, I think that's one.
[00:23:36] Trap I devoid is, you know, more tool access, um, with more readiness. You know, exposing more systems to AI, doesn't necessarily make the workflow trustworthy, you know, better definition, cleaner handoffs, clear controls, like a lot of the things we're talking about that MOPS has direct impact, um, and command and control, uh, too.
[00:23:54] So, you know, I think, you know, people who wanna. Kind of get a, get a [00:24:00] fresh start at that and take a look. Like, do something like make a two column list, like called Ops can fix this now. Uh, and the second list called needs Data engineering. You know, you put dedupe enrichment Sync, logic, field Governance, routing Rules in the first column, all the other shit goes in the other one.
[00:24:16] And then you kind of prioritize and, and see like where to go from there.
[00:24:21] Phil: I love it. Yeah, it's, it's a good, good first step to, to take for sure. Yeah. It's, it's not glamorous work. Um, but, you know, may, maybe earlier in my career, I, I was excited to do it because no one else was excited to do it. And it is important work, like you said, because, you know, like AI is, is only as good as, you know, fixing some of that foundational work.
[00:24:41] And part of that is not. Glamorous stuff. Someone's gotta do it if you want the context to be better. Um, but the kind of related to that, Jason, like,
[00:24:50] 5 — Which Data Problems Block AI Deployment and Which You Can Ignore
---
[00:24:50] Phil: I think a lot of operators are wondering like, what can I ignore for now? Because like maybe they're starting that list and it's actually a pretty long list.
[00:24:59] Even the stuff [00:25:00] that doesn't require data engineering is stuff that's never been prioritized. It's fallen through the cracks because it's not. Directly tied to revenue and the company wants to, you know, quick wins and it's never been one of the quick wins ones. Um, so like. You know, if the answer is like, clean everything, no one's ever gonna start doing that.
[00:25:20] So like where do you draw that line? Like what data problems are ugly but kind of tolerable at, at, at first, like tolerable at first? Like, and and which ones would you say are like not, uh, are disqualifying them? Like, I guess what I'm asking is like there's a lot of readiness work that becomes abstract, like dashboards, governance docs, architecture diagrams.
[00:25:43] If you had to like make this tangible for a team, like what are the first few checks you would run to determine whether they're actually ready enough to do some agentic stuff? Like where does that diagnostic kinda look like for you?
[00:25:56] Jason: Yeah, I, I love that question actually. And. So I think the [00:26:00] first thing, um, I would ask is not necessarily like how clean is the data, I would say, you know, what authority are we delegating and, and what's needed to actually be successful, um, within this workflow or process or whatever. So, you know, if the system is summarizing through things like summarizing, ranking, drafting, you know, we can tolerate more mess.
[00:26:20] Um. don't have to be as strict there, you know, spin cycles on that. You know, if it's changing records, controlling spend, contacting customers on our behalf, making decisions that impact revenue, then yeah, the bar jumps pretty fast, right? Um, so you can, you can ignore some ugliness at first, like, um, you know, messy, free, messy pretax, some like missing non-critical elements.
[00:26:46] Um, sure. Like you talked about, we talked a little bit about duplicates, you know, some low stakes, duplicates. I don't know if anyone's ever like, eradicated 100% of all, uh, things like duplicate records [00:27:00] for some reason in the year here we are in 2026 and we're still plagued by this. I know there's a lot of tools out there that make it a lot easier, but, um, so, you know, I think, but when you, when you can't ignore, you know, undefined business terms, uh, broken, you know.
[00:27:16] Issue resolution still or unknown freshness on like that decision critical data that ai, uh, needs to make those decisions, like unclear approval paths, like, and then, and then no way to like, turn things off or, or wind it back, like if, if something's going wrong in that aspect. So I think my diagnostics for that's pretty simple.
[00:27:34] Like, uh, what exact actions AI gonna take, you know, what fields, uh, drive that action. Can a human reproduce the answer from source symptoms if needed? Like I talked about getting that extra layer down, not just accepting an answer that seems believable, blah, blah, but like, how did you actually, like, where did it come from?
[00:27:54] You know, what's the explainability behind that? Um, you know, that [00:28:00] what kind of threshold matters, um, for, for getting to a good outcome there? What happens if it's wrong and you know who can stop it? So, you know, I think if teams are looking for like a quick exercise, like how do I actually implement this tomorrow?
[00:28:14] Like, take 20 of your historical examples that you have. Workflows like if, if a, uh, run 'em through, you know, if a human can't explain the output, trust it, or identify the owner, then the system doesn't get, you know, more authority to move on yet.
[00:28:29] Sponsor: GrowthLoop
---
[00:28:29]
[00:29:32] Sponsor: AttributionApp
---
[00:29:32]
[00:30:34] Darrell: let's talk a little bit about Fal, false positives. What are common, like most common false positives that you see where a team thinks they're AI ready Because they have things like a warehouse, A-A-C-D-P, or cleaned up dashboard, but then underneath that the foundation is kind of actually not ready.
[00:30:52] It's too shaky for agents to actually perform well.
[00:30:56] Jason: Yeah. Uh, I think it's a, a fair question, you know, and I think you've had [00:31:00] some other, uh, guests, you know, on the show have had opinions about that. Um, uh, as well, um, a lot of, a lot of great things to unpack there. So I, you know, I think I'd slightly reframe the warehouse question. Um, just being selfishly where I sit, um, you know, I'd say, you know.
[00:31:19] I wouldn't say having that is a false positive. You know, where I sit at Kumo, I think the warehouse is actually a, a hidden strength, um, because it concentrates, you know, the richest relational context in the business. Um, so our, you know, our own products connect around directly connecting warehouse, you know, generating, uh, predictions, you know, explainable outcomes from that relational data structure and pushing those.
[00:31:42] Um, you know, into, into workflows, like I said, with explainability. So, you know, the false positive thinking like warehouse, CDP, clean dashboard plus shiny agent equals uh, readiness, right? I mean, those are ingredients. Those are ingredients at the end of the day, right? So, you know, [00:32:00] what gives me confidence is, is actually a much less, uh, glamorous, um, kind of viewpoint.
[00:32:05] You know, one clear system of record, you know, for the motion documented process named owners. Defined exception handling human validation where, where stakes are high, uh, and narrow use cases with, you know, uh, a bounded blast radius, uh, you know, in case something goes wrong there. Uh, so, you know, at the end of the day, I trust a boring well run system more than a Franken stack.
[00:32:28] Um, so yeah, that's, that's kind of, you know, where I would land on that.
[00:32:33] Darrell: Yeah, it's like simplicity, I think. Um, and not like too complicated processes because I was also gonna ask like the converse of the question, which are like, what are some hidden strengths that a tool, uh, that a team may have that make them more ready or maybe like low hanging fruit or signs that a team can really run fast with, with, um, ai.
[00:32:54] Anything you can think of there that you've seen?
[00:32:57] Jason: Um, yeah, I think I, you could ask one [00:33:00] uncomfortable question in your next meeting, you know, like, um, is this AI workflow, like if it makes the call wrong, call tomorrow. Like, who finds out first? Who decides like, how do we stop it? Like if nobody can answer, that's your readiness gap. Um, so there are tons of like hidden strengths and it's like really a lot of the things that we, we talk about throughout this episode, uh, all the things that MOPS is bringing to the table that were all like. Excellent at, uh, that's the core, like lifeblood of, like we talked about, not model creation. Like eventually we're gonna get to the point we already are at the point where like there aren't gonna be teams, huge teams of data engineers needed anymore to build like all these different models. There's no more training that's gonna be needed.
[00:33:43] Like we already have these, like pretty much one size fits all models or fine tuned models for specific use cases they've been trained on like. Millions of rows of synthetic data. Um, so that whole like needing the data [00:34:00] engineer and the whole team and six months to build a model for this one like use case, like that's gonna go away pretty fast, honestly.
[00:34:07] Um, so what the strengths are is really a lot of these things I were talking about that Miles Springs to the table is once the model's good, like what are the other things that are needed to actually like execute this?
[00:34:19] Phil: I love it. Jason, you talked a lot about the data warehouse throughout the conversation today, and
[00:34:24] 6 — What Goes Wrong When Agentic AI Optimizes Directly on Warehouse Correlations
---
[00:34:24] Phil: there's kinda like this growing trend of letting agentic AI kinda loose on the company's data warehouse. Um, I had a guest on, uh, from Growth Loop that kind of like called this a. Terrible idea. Uh, he said that since a warehouse is really only reflecting a brand's current status quo, which may have nothing to do with causal relationships, what are your thoughts on the risk of accelerating negative dynamics, like aggressively prompting a product that correlates with high LDV, but doesn't actually cause it?
[00:34:56] Like, what are your thoughts about just letting agentic AI loose on the warehouse? I.[00:35:00]
[00:35:00] Jason: Yeah, I have a, you know, I have a unique viewpoint on that as well. I'm happy to share. You know, I, I think I'd partly agree with the warning, but I'd narrow it. You know, the warehouse isn't the problem. I think, um, unbounded autonomy is, like we talked a little bit earlier, like, you know when teams go from like, oh, here's a big pile of data, like AKA, our warehouse, and here's, and here's an agent, and we're just gonna throw it on top and expect it to like. Like, no, the answer answer every possible, like, use case under the sun. Right. So, you know, I think selfishly I have a, a unique viewpoint just because of where I'm at at Kuo. You know, like I talked about our whole like. Uh, thesis is essentially aru, the structured relational data warehouse is oftentimes the highest signal context in the business.
[00:35:46] Um, you know, so we're built to connect to that turn relational context into, you know, real life like kind of business outcomes and, and push those into operational workflows, explainability. So overall the warehouse is, in my opinion, like [00:36:00] a context layer. Um, I think where I agree with the critique you talked about is.
[00:36:05] Uh, you know, a warehouse is a record of what happened. It's not necessarily a rule book for what an agent should do next. So, uh, you know, if you let a generic agent optimize directly on historical correlations with unbound authority, yeah, you can absolutely scale the wrong behavior. Uh, a product that correlates with high LTV does not necessarily cause ILTV, um, you know, so my view, I guess, you know, the warehouse is great context layer.
[00:36:32] You know, systems like, you know, where I sit, what we work on, you know, can make it a very strong, uh, next layer. Um, but you know, prediction's not policy. So once you cross into that action, like you still need guardrails, business rules, approvals, evidence to six is actually like driving business income. Um, you know, overall, like if I had to compress, you know, everything we've talked about today into like one idea, I think it'd be this like.
[00:36:59] Most teams [00:37:00] don't have a data problem. You know, they already have more signal, uh, than they think. Uh, what they often lack, you know, are, are the things that we just talked about. Um, and a practical way to test whether, you know, AI should advise draft act. I think that's really what minimum viable readiness is really about.
[00:37:17] So, you know, if you wanted to give like one concrete thing listeners could actually use tomorrow, um, you know, choose one workflow, not 10. You know, write down the exact decision that AI's informing list the five fields or signals driving that decision. You know, name the, the, the source of truth, the owner, the approval boundary.
[00:37:39] Run it on historical examples that you, that you already, you know, have confidence in before giving an authority. Like, that's it. Like there's no, uh, if there's no owner, no source of truth, no rollback, pat, then it's not ready.
[00:37:53] Darrell: Yep. Just the fundamentals. I love like the emphasis on first party data. Two, which is great, and I think it [00:38:00] kind of leads into this question that I have that I wanted to ask. So you are at Kumo, which exists because structured warehouse data is the highest signal context available. And your whole argument, or the main point is that most companies fail to use it properly.
[00:38:16] What, what does that mean? Like most companies fail to use that properly and how do you think about that at Kumo and like what, what does. Your solution due to like make this better.
[00:38:32] Jason: Yeah. Yeah, it's a, it's a good point. So, you know, at Kuo we. We've built, uh, what we call Kumar fm, uh, which is the world's first relational foundation model. So you can think of that like chat GBT, but for structured data. Um, so that's built, you know, connecting directly to your, to your warehouse data wherever that lives, whether that's, you know, snowflake, Databricks, Amazon, like wherever.
[00:38:56] Um, and, you know, generating, you know, predictions from [00:39:00] relational structure and pushing those into workflows with, with the explainability that we talked about. So it's all the common things that, you know, we're working on as, uh, actually there's like so many use cases. Like, I'm not gonna get into like a whole customer testimonials section.
[00:39:14] I, I don't wanna turn this into like a, a plug. Um, but, you know, simple things like, oh, is this transaction fraud? Like. Some of the world's largest like crypto, uh, institutions actually use us on the backend to determine like, is there fraud happening in real time? Uh, when's the next customer gonna turn? Uh, you know, where should we place this ad?
[00:39:32] Uh, what's the demand gonna be next quarter? So a lot of these, like very mission critical business flows. Um, so, you know, the platform's already been pre-trained on billions of relational, you know, patterns across thousands of data sets. You know, delivering a combination of accuracy and speed that would otherwise, you know, be impossible to achieve.
[00:39:51] And that's really what I talked about earlier is, um, I think what we're gonna see going forward is, um, [00:40:00] you know, the, the size and time that it takes for like these engineering teams and data teams to build these models. Like that's gonna continue to shrink and shrink and shrink and go away. Um, and, you know, everyone will, will kind of.
[00:40:12] Uh, grow through that and, you know, the work that MOPS is doing will, will kind of also grow into, you know, we have this convergence of the data engineering world and the operations world, and that's kind of how we set some of these new teams coming up, like GTM engineers, which is really a combination of those two, like skillsets of what was, uh, you know, someone working on marketing automation, for example, in the past, or someone who was also working on like building data models and things like that.
[00:40:38] Um, so that's my own, um, take, uh, on that. Anyways.
[00:40:43] Phil: Yeah. I love it. Jason, we, Darryl and I, were, were chatting before this looking at Una's website and like all the use cases and then like at the bottom of the menu it's just like, see o all other a hundred plus use cases. And I was trying to empathize with. The product marketer on your team who's having to like [00:41:00] write copy for all these different personas and use cases and we're like, man, that must be, must be tough.
[00:41:06] But it is a really cool platform and um, like I first discovered it many years ago when, uh, my former cos and I did like a super deep dive, uh, in like 2023 about like AI and marketing and, uh, at the time the positioning with Kuma was like. Democratizing machine learning pipelines. And I was working at WordPress when we had like a big data eng team and it took us like couple of weeks to create one pipeline for one event in our homegrown CDP.
[00:41:34] And at the time you guys were just like, you know, anyone on marketing can democratize creating ML pipelines. So it's cool seeing the evolution of that and all the use cases and, and how powerful the platform is. Uh, I wanna bring us back full circle though, Jason. We started the conversation talking about.
[00:41:50] You know, your first experience with rolling out Agentic workflows and agents at Kuo and how it kind of like ended up being pretty messy because you didn't fix some [00:42:00] of the foundational things at first.
[00:42:02] 7 — When to Ship AI Before Your Data Is Ready and When to Fix the Foundation First
---
[00:42:02] Phil: I wanna like get your take on the ship first argument here for the folks that are kinda listening and are just like, Hey, we need to move faster than pausing and fixing this like laundry list of data.
[00:42:14] Jason, so. Like I've heard some really smart people argue the opposite, where like ship the ai, let the bad output create the pressure to fix the data as opposed to like, let me get priority on this project that hasn't been prioritized for many quarters to fix the data before we do ai. These people are just like, nah, like.
[00:42:34] Before we fix the, the data, let's ship the ai, let the bad stuff happen, the chaos that you can talked about, and let that kind of be the pressure to actually then get priority to fix the data. The failures kind of become the business case, if you will. Is, is there like any version of that that's actually right, based on what you went through?
[00:42:52] Or is it just kind of rationalization for, for moving fast?
[00:42:56] Jason: Yeah. You know, I think it honestly, um, we [00:43:00] we have these two extremes, right? Where, uh, of course us, you know, who live in the ops world we like, because we live through the pain of like. The problems that, like the bad data causes, and we, we see the impacts of that. Um, so the more we fix that, the easier our life is and the better the outcomes.
[00:43:19] Right? And then you have the, I'm not gonna call 'em the opposition because they're our partners and I'm actually on the business side myself now. Uh, but then you have those on the business side who, who maybe don't always, uh, understand that or wanna learn like. They, they just have a number to hit this quarter, or they, we need to, you know, we need to increase pipeline.
[00:43:38] Uh, and it's like, okay, like, um, I don't care how, like it's your job to go figure out how to do that, but we need to also, we can't wait six months, uh, to, to like, uh, increase pipeline, right? So I think there's always gonna be this like, um, chaotic environment of moving. Fast while also [00:44:00] creating structure and scaling like appropriately.
[00:44:03] Um, so I think, you know, the reality is it's, it's gonna be a combination of both and that's like the kind of the lessons we talked through earlier, you know, which is really to, to break those down. You know, start with one workflow. You use historical examples like not trying to tackle the entire, like data, data ecosystem at once.
[00:44:22] Uh, and that really makes it much more manageable and kind of bite-size because. Um, look at the end of the day. you're, if you're, if you're based on revenue and, and you know, that's not, you're not, you're not driving, that pipeline's not going up. Like that's a problem, right? At the end of the day. But also what, why, what is the causation or what are the problems like behind that?
[00:44:44] There is all the things that we talked about that we have to fix on the back end. So it's, it's, really, I think, fearing out that, uh, that happy marriage, um, and, and the size of your organization matters. Like if you're, if you're extremely lean, uh, and you have a queue of 200 things to get to, like. That's [00:45:00] much harder than a team that has, uh, 200 resources at their disposal and like, you know, so, uh, I think, you know, there's definitely gonna be some of us who suffer more than others through that.
[00:45:10] Um, but that is, I think, you know, the way forward for the most part.
[00:45:15] Darrell: Yep. Totally. Yeah. And it really depends, you know, on what the impact could be. Is this gonna be a really big thing? Like I've always loved the, uh, the two-way door concept where if you go through a door and you. Can't come back, then you have to really take your time and, and, and make a right decision and make sure you're thinking about all the risks.
[00:45:34] But if you can come back pretty easily there, there's not too much risk in just maybe trying it with ai. Um, what. This one's a, a, a cool question. Um, especially now that things are changing so quickly and I feel like at work we're needing to try new things, but also get approval to try those new things.
[00:45:54] So like what's your best tip for getting internal buy-in from leadership? You know, [00:46:00] especially if it's like an unorthodox or like crazy idea.
[00:46:05] Jason: You know, this is, uh, I, this is a good one that I really like. I think, um, you know, my best tip is don't try to sell the whole idea, right? Uh, try to sell the bottleneck and the reversible next step. So like, uh, a lot of internal buy-in dies because people. Feel like they're being asked to buy your whole worldview at once, right?
[00:46:27] Like, like who's are you, if you're trying to get them to undertake the quest for the Holy Grail with you, like. That's probably a harder sell than, uh, you know, no, we're actually just trying to get to the next town and get a, a, a pouch of water to, uh, you know, fuel us to the next day or whatever. Um, so like I've learned to make it much smaller than that.
[00:46:50] I think. Like, uh, if something's worth doing, I try to frame it like, here's the actual bottleneck, you know, here's the business consequence of leaving it alone. [00:47:00] Um, here's the smallest test that we can do to tell us if we're right in this before we invest too much time or resource. Uh, and, you know, here's the, here's the blast radius if we're wrong in that.
[00:47:09] Again, you know, I think at that point, you know, people don't have to believe in the whole vision. They just have to agree that the problem's real and what's the reasonable next move. And then once you actually start to gain traction, um, and success around that trust grows. Like, of course, like people open up more and start to see, okay, this has actual value.
[00:47:27] Um, and then you're gonna be able to move much faster on the, you know, kind of the rest of the, the things you wanna achieve there.
[00:47:33] Phil: I love it. I, I like how you said unorthodox there, Darrow, because like, I, I feel like if you're just trying to pitch, let's fix all of our data, that's probably falls under the, the crazy thing that you're trying to get buying for, but. I love your advice there, Jason. Um, another fun one, uh, before we get into some of the last ones here, uh, your current title is Head of Marketing slash GTM Engineering and GTM [00:48:00] Engineering is one of the most debated job titles in marketing ops these days.
[00:48:05] Uh, so it'd be fun to get your take on. Like you've previously been head of Revenue marketing, head of marketing ops, head of marketing Analytics and Data science, director of Growth and Lifecycle. Like you've had your kind of taste across the landscape here, not even like touching some of your earlier experience.
[00:48:23] 8 — What GTM Engineering Actually Means When AI Automates the Middle
---
[00:48:23] Phil: What the heck does GTM engineering look like for you? Like is it that different than marketing operations? What, what are the differences for you?
[00:48:32] Jason: Yeah. Uh, you know, I love this question also. Uh, and it is probably is, you know, very, uh, debated. So, um, let's get into it. You know, I think this is exactly bridging, you know, what we were talking about. Um, really the future of. Both marketing operations and data engineering world. You know, as we talked about, systems are, are taking over more and, you know, human's role and supervising those outcomes.
[00:48:58] Um, [00:49:00] I think, you know, the reality is where I'm at, uh, is, you know, we're doing an extreme amount of work with an extremely lean team. Um, and, you know, being an AI first company, you know, one of our, you know, biggest pushes is always like. How can we solve this challenge with ai? Of course. Um, and, and a lot of that is ly actually, you know, implementing agen flows for, um, things that used to require either a data engineer to do or a marketing operations person to do.
[00:49:32] Um, like, and yeah, I still have, you know, a head of rev ops, head of marketing ops that, that work there. So, so there's still definitely humans in the loop. Uh, but I think what we're seeing is machines are now automating a lot of these tasks. Um. Around, you know, all of the flows of marketing, you know, uh, you know, lead generation qualification, um, early, uh, you know, driving early pipeline.
[00:49:57] Uh, and that's really kind of what that, what that is to me. It's [00:50:00] the, it's the, it's the machine and the loop now. Um, so GTM engineering is combining, um, you know, essentially the data folks, the engineers and the, and the, and the marketing operations of rev ops folks. Into kind of one function that is, uh, cyborg in nature, I guess.
[00:50:20] Phil: I love it, Jason. It's actually pretty spicy because I think most people think of just like a technical sales role. When you think of GTM engineering, like I've seen a lot of non-technical sales. BDM people like call themselves GTM engineers,
[00:50:39] so
[00:50:40] Jason: not, I would not agree with that, uh, that assessment probably, but yeah. Yeah.
[00:50:45] Phil: Yeah, no, I, I like, uh, I, I land a lot closer with, uh, where your definition is. It's a, it's a fun time to be in, in marketing ops and, and all the GTM space for sure.
[00:50:55] 9 — How Jason Dobbs Decides What Deserves His Energy
---
[00:50:55] Phil: Uh, we got one last question for you, Jason. You're obviously GTM engineering expert marketing ops leader, but you were also a. Adrenaline junkie, an electric longboard modeler.
[00:51:05] You got a ton of stuff going on on this side. One question we ask everyone is how do you decide what deserves your energy at any given moment, and what's your personal system for staying aligned with what actually makes you happy?
[00:51:18] Jason: Uh, I think, you know, I love that question. You know, we talked a little bit about, you know, my, my, my adrenaline junkie side, you know, whether it's riding, you know, F1 longboards or e foiling or snowboarding or free diving somewhere in the world. Um, you know, but happiness is, uh. Is is a very important one, and I think it's different for all of us.
[00:51:40] So I, I'm not gonna say what happiness should be for you, but I'll guess I'll just talk about it from my lens. And I probably have a slightly unorthodox answer because you know, of my background, um, and a lot of my life was shaped by, you know, mission first environments. So I think I don't usually ask like, what makes me [00:52:00] happiest today?
[00:52:01] Um, you know, I kind of ask like, what's my mission today? And that is that aligned with the life that I actually wanna build. Um, so for me, happiness is less about comfort and more about alignment. Um, you know, some seasons are hard. Uh, you know, we talked about where we're at and whether you want to agree we're in a global recession or not, or have we, or, you know, I'm not really here to get into that, but, you know, building anything meaningful always comes with sacrifice.
[00:52:29] Um, but I've also learned not to romanticize the grind and the burnout. Right? Um, so if a mission is. Steadily cost me my health, relationships, or integrity, then it's too expensive, right? Uh, so, you know, I think my filter for happiness is pretty simple. Like, does this matter? Is it mine to carry? Uh, is the contemporary, is the cost temporary or purposeful?
[00:52:51] Uh, and or am I just calling this like burnout discipline, right? Uh, so when those things lined up, you know, I can, I can give something a [00:53:00] lot of energy and, you know, in turn, um, you know, be happy as well.
[00:53:05] Phil: I love it. It's such a cool, uh, filter there.
[00:53:08] 10 — What Jason Is Reading: Intelligence History, Mind-Opening Nonfiction, and Dune
---
[00:53:08] Phil: Uh, last one, Jason, I lied that you said we don't have to ask, uh, you about books that you're reading right now because you're reading a bunch of weird stuff. I'm putting you on the spot here. What's the weird stuff I want to
[00:53:19] Jason: oh. The weird stuff. Oh man, you're gonna put me on stuff with the weird stuff. Um, no, I'm a, I actually, I'm, I read a lot about, uh, you know, as a human, like coming up. I think we all have. Oftentimes are locked by our own, like unconscious bias or like even things like that. So a lot of what I'm reading is like, you know, how to even open up my own mind to like new ideas that maybe wouldn't be, um, possible, um, kind of without this.
[00:53:48] So that crosses into a lot of different realms, whether it be, uh, knowledge or meditation or kind of, you know, other things. Um, but. Also, you know, I'm big into, you [00:54:00] know, of course my background, uh, in the intelligence industry. So I, I, I read, enlist a lot of podcasts on, um, you know, things, things in that environment.
[00:54:09] Um, you know, a lot of kind of CIA, uh, operational things. Uh, those are a lot of things I nerd out on. But then if you want to get into fiction area, you know, we can talk about, I'm a big science fiction nerd. As well. Um, so yeah, talk about a book. Let's, let's talk about dune. Let's talk about dune, I guess, right?
[00:54:27] It's, it's popular. We got the new movie coming up. Maybe not because of sci-fi cannon, but like what's underneath? It's, you know, underneath. I think it's really a story of power incentives, prediction, and the danger of, uh, thinking that seeing patterns means that you control outcomes, right? Uh, and I think that theme has, has aged really well today and kind of the age that we're in also.
[00:54:50] Phil: It's super cool. Yeah. I'm pumped for, uh, the new movie, uh, huge fan of, uh, de ne the, the director. Um, yeah, I Are you reading, have you [00:55:00] read the Dungeon Crawler Carl, uh, books Darrell and I are are, are binging
[00:55:04] Jason: Oh, okay. All right.
[00:55:06] I'm gonna have to put that on the list. I'm gonna put that on the list.
[00:55:09] Phil: Yeah. We'll, we'll share some ideas there. Didn't know you were a big, uh, sci-fi guy.
[00:55:12] That's pretty much all Darrel and I read.
[00:55:14] Jason: Oh, love that man. Yeah, for sure.
[00:55:17] Phil: Awesome. I really appreciate your time, Jason. This is super fun. Thank you so much for joining us.
[00:55:21] Jason: Yeah, it's been amazing. Thanks, Phil. Thanks Daryl. Look forward to, um, you know, seeing you guys again.