Show notes
What's up everyone, today we have the pleasure of sitting down with Jason Dobbs, Head of Marketing and GTM Engineering at Kumo AI.(00:00) - Intro
(01:24) - In This Episode
(01:57) - Sponsor: MoEngage
(02:54) - Sponsor: Knak
(04:35) - How Undefined Data Definitions Make AI Confidently Wrong
(08:18) - Why Context Engineering Replaces Prompt Engineering as the AI Bottleneck
(12:59) - The Five Non-Negotiables for AI Readiness in Marketing Ops
(15:42) - Why Marketing Ops Is the Context Architect in an AI-First GTM Stack
(24:50) - Which Data Problems Block AI Deployment and Which You Can Ignore
(28:29) - Sponsor: GrowthLoop
(29:32) - Sponsor: AttributionApp
(34:24) - What Goes Wrong When Agentic AI Optimizes Directly on Warehouse Correlations
(42:02) - When to Ship AI Before Your Data Is Ready and When to Fix the Foundation First
(48:23) - What GTM Engineering Actually Means When AI Automates the Middle
(50:55) - How Jason Dobbs Decides What Deserves His Energy
(53:08) - What Jason Is Reading: Intelligence History, Mind-Opening Nonfiction, and Dune
Summary: Jason Dobbs spent 7 years assembling intelligence briefings for the President, and he says most AI failures in martech are the same problem he was solving in 2003: teams acting on context they never actually agreed on. In this episode, he breaks down the 5 non-negotiables of minimum viable readiness before you deploy any AI agent, explains why the marketing ops function is becoming more critical as AI takes over execution, and argues that unbounded AI autonomy creates more risk than warehouse data ever will. He also defends GTM engineering as a real discipline rather than a rebrand, and closes with a Dune analogy that lands better than it has any right to. If you think AI readiness is primarily a data engineering problem, this episode will change how you think about your team's role in it.About Jason DobbsJason Dobbs is the Head of Marketing and GTM Engineering at Kumo AI, where he leads go-to-market for KumoRFM, the world's first relational foundation model, which generates accurate, explainable predictions directly from warehouse data. Before Kumo, he served as Global Head of Revenue Marketing at Logitech, where ABM and advanced segmentation drove 40% of B2B sales revenue and 79% YoY ARR growth. He also co-founded Trypp, an autonomous UX research agent for continuous post-ship product monitoring, and has held marketing and analytics leadership roles at Seagate, HTC Vive, Apple, and Google.Jason spent 7 years as a United States Air Force intelligence officer, including work on the President's Daily Intelligence Briefing, an experience that shapes how he thinks about assembling trustworthy context for high-stakes decisions under uncertainty.How Undefined Data Definitions Make AI Confidently WrongEvery marketing ops team has heard the warning: AI is only as good as the data you feed it. You've nodded along. You've probably said it yourself. But the warning leaves out the most important detail, which is what the failure actually looks like when the model is running.Jason Dobbs knows what it looks like. He learned it from a crash. He rides high-speed F1 electric skateboards at 50 to 60 miles an hour, and he's fallen before. He can tell you he's never fallen the same way twice. When he greenlit agentic and predictive workflows at Kumo AI before the data architecture was ready, the failure followed the same logic: unexpected, and avoidable only in hindsight.The model returned results that looked operational. Scores came back precise. Summaries sounded coherent. Recommendations felt grounded. The failure was invisible to anyone who didn't already know what correct should look like.The weakness surfaced when someone pushed. Ask the follow-up question, why did you score this account, what data drove this decision, and the logic fell apart. The definitions feeding the model had never been agreed on across the business. Sales and marketing were not working from the same idea of what a qualified lead meant. The AI had scaled an unresolved internal argument into what looked like a confident answer.Jason traces the failure to a structural problem that predates any model decision. When a system cannot explain its own outputs, and when nobody in the room has standing to say what the correct answer should look like, you have built a very polished way to be wrong. That is dangerous precisely because it passes a surface inspection. People who were not close to the data trusted the output. Nobody pushed back.What he carried out of that experience was a reframe of what marketing ops actually produces. The shared definitions, the trusted data sources, the named owners, the workflow guardrails: that is the product. Every AI initiative sitting on top of unresolved questions about what the business means by its own terms will generate outputs that look credible right up until someone has to act on one. Speed to AI deployment and quality of AI output run in opposite directions for teams that skipped the definition work. The ceiling on any AI system is the clarity of what the business agreed it was optimizing for before anyone touched a model.Key takeaway: Run this diagnostic before signing off on any AI or analytics initiative: can a human reproduce the logic behind the output and explain who owns the decision that follows? If nobody can answer that cleanly, the system is automating an unresolved argument. Start by documenting shared definitions for your 5 most-used business terms (pipeline, qualified lead, active customer, opportunity, churn) and get explicit sign-off from sales, marketing, and ops before any model sees them.Why Context Engineering Replaces Prompt Engineering as the AI Bottleneck"Context engineering" is appearing in every AI strategy conversation right now. Scott Brinker devoted a report to it. Conferences are building entire tracks around it. The framing is right, but for most teams the phrase still points at a feeling rather than a concrete set of decisions.Jason Dobbs's version is more precise. "Fix the data" is the directive most teams have been living under for years, and the structural problem with it is that it makes the work sound like a single epic project with a clear endpoint, a Holy Grail that teams have been questing toward since before the first CRM went live. The warehouse always has gaps. The CRM always has problems. The right question is narrower: what minimum context and control does this specific workflow actually need to produce a trustworthy output?That reframe narrows the scope from an organization-wide data quality initiative to a workflow-specific requirements checklist. For any given AI decision, the context bundle has 6 components: the definitions the system is operating from, the data sources it has access to, the tools it can invoke, any memory it carries between sessions, the guardrails on what it can do autonomously, and the escalation path when confidence runs low. Those requirements are specific to each workflow. They're answered by asking exactly what this workflow needs, not by cleaning the warehouse in general.The shift from prompt engineering to context engineering reflects how the bottleneck has moved as the models matured. A prompt is the last instruction a model receives. Context is everything it's working with before that: the definitions, the data access, the scope of authority, the path back to a human when a decision exceeds what the system should make on its own. Teams tuning prompts while leaving the underlying context undefined are optimizing the most visible variable in the system while the one that actually governs quality sits untouc...