Why AI Can't Pick Your Tech Stack | Scott Brinker

If Part 1 was about the state of the landscape, Part 2 is about what happens when you actually try to operate inside it.

Scott Brinker returned to Making Sense of Martech to pick up where the cliffhanger left off: the concept of synthetic certainty, AI outputs that feel authoritative but can't actually be validated. What followed was the most direct conversation about AI's real limits in marketing published this year, from the person who has spent 17 years tracking what actually gets deployed versus what gets hyped.

The headline: most of marketing is not a verifiable domain, and that is not a weakness. It is the argument for why human judgment isn't optional in the stack. It's the entire point.

The setup: a client asks Jacqueline to use AI to evaluate six CDPs across 90 criteria. The question was why can't AI just do this. Brinker's answer was one of the most clarifying things said about AI in enterprise marketing in a long time.

The problem isn't that AI gets the wrong answer. The problem is that it gets an answer that feels like the right one. Brinker draws a direct line to attribution modeling: "That is absolutely delusional. You have a subset of data. You do not have all the data, you do not have all the context."

The same logic applies to vendor evaluation. Even if the data exists, it isn't uniformly structured, quantified, or current. Vendors sugarcoat. Websites go stale. And above the data layer sits something AI genuinely can't access: the context of what actually matters to a specific organization, team, or use case. "What isn't in the data that would actually, as a human, really matter to you, in which experience you would choose" — that layer, Brinker says, is not where AI is right now.

The practical implication is not that AI is useless in vendor evaluation. It's that there's a meaningful difference between AI as a research assistant and AI as the decision-maker. The organizations getting this wrong are the ones collapsing that distinction.

The most under-adopted AI use case in the entire 2026 survey, out of 70 measured, is using AI for vendor selection and stack management. 88 of 130 respondents aren't doing it. Brinker calls it the cobbler's children problem. The martech industry builds tools for everyone except itself.

But the more important point is the one underneath: there's a meaningful difference between using AI as part of a use case and outsourcing the entire use case to AI. Brinker is explicit. "I want an AI agent that's going to track all the major vendors I have in my existing stack or under consideration and alert me to news and announcements." That is a legitimate use of AI. "The AI just picks our tech stack for us, and I haven't looked at it in ages" is something else entirely.

The cobbler's children aren't just barefoot. They're walking on gravel and telling everyone the stones feel great.

103 of 163 respondents in the 2026 survey are doing nothing to verify the AI-generated content they're producing at scale. For many people, that number reads as negligence. Brinker offers a more generous and more useful interpretation.

Right now, the verification mechanism is still a human. "A human that's going to review it and decide, do we want to do this or not?" The gap isn't the absence of verification. It's absence of scalable verification. That distinction matters because as personalization at scale becomes the expectation rather than the aspiration, the human-in-the-loop model stops working. "You can't review all those." The tooling to govern and monitor AI content dynamically isn't deployed yet because the urgency hasn't fully arrived.

But 73% of organizations now have a formal AI policy, up from 52% in 2024. Progress on paper. SAS found that only 8% have full confidence in their governance readiness. The gap between having a policy and being able to enforce it is not a technology problem. "The technology is never the barrier. It's always the change management." Policies get written. Culture changes more slowly. And there is, Brinker concedes, a limit to how fast you can accelerate change management before it simply stops moving faster.

The jobs conversation in martech tends to swing between two poles: AI will eat everything, and marketing is safe because it requires creativity. Brinker cuts through both with a framework he's been using in talks for the past year.

Marketing work falls into three buckets. Strategy and creative — what everyone thinks marketing is and what has historically received the most glory and the least calories. Production and analysis — the vast majority of actual working hours, less valued, and now being directly consumed by AI. And marketing ops and martech: underfunded, underappreciated, and increasingly the foundation for everything else.

The unwise executive response to AI eating the production and analysis bucket is to cut headcount and pocket the savings. "That efficiency crunch has absolutely zero competitive differentiation. Every single company on the planet is going to get the benefit of that." The wise response is reallocation: toward strategy and creative, where AI becomes a multiplier, and toward marketing ops and context engineering, where the infrastructure that makes any of it reliable gets built.

The risk, Brinker says plainly, falls on people who define their role by the task rather than the function. "If you're defining your role based on the task, then you are at risk." The ones who are concerned but aren't learning are the ones who should be most concerned. "The only way out is through."

The most dangerous assumption martech operators are making right now. Brinker doesn't hedge: the assumption that full vendor consolidation is the goal.

"I always hear marketers say, ' Can't we just have this stuff consolidated? ' And I get it." The fragmentation is real. The cognitive load is real. A certain amount of consolidation is genuinely good. But the logical endpoint of that desire is an organization that is 100% dependent on a single vendor. And what happens at the next renewal cycle? "Your prices are going up 20%. And the renewal cycle after: another 50%. Because hey, we gotta eat. And you are just completely locked in."

Brinker put it plainly: do you really want to be that vendor's b*tch? And then answered his own question. The IT world went through this in the 80s and 90s. Marketers are watching the same movie for the first time and treating the shining city on the hill as a destination rather than a cautionary tale. Optionality is leverage. Modularity is negotiating power. Strategic consolidation is smart. Complete dependency is surrender.

The question nobody asks the analyst: the report is sponsored by seven vendors, each with a stake in its conclusions. How does editorial independence hold?

Brinker is transparent about the mechanics. Sponsors have no say in the core editorial content. They don't see it until the day before publication. What they do get is branding, a lead-generation webinar, and a sponsored interview with one of their executives. A place where the sponsor has narrative control over their own perspective, distinct from the rest of the report. "We try to advise them: if you want things that people are actually going to read and enjoy, don't make it a sales pitch."

The structural protection is what the report doesn't do. "There is nothing we do in these reports that is in any way trying to rank vendors, say this vendor is better than that vendor." No vendor heuristics. No evaluative frameworks that could be gamed. The analysis operates at a level above the vendor layer, which means even if sponsor conversations introduce ideas, those ideas land in a space where they don't change the narrative. "I'm happy to be transparent that I think it's a very messy thing in the market." It is. At least this version of it is honest about the mess.

Missed Part 1? Check it out here.

Special thanks to Claude for helping to summarize this conversation.