The 2nd-order effects of generative AI on marketing and martech, 3 years later

Three years ago, I published a post exploring the 2nd-order effects of generative AI in marketing and martech. ChatGPT was barely five months old. “Agentic” wasn’t yet a word that appeared in every vendor keynote, board deck, and bar napkin in martech. “Prompt engineer” was being breathlessly reported as the hot new $300K career — a profession that has since enjoyed roughly the lifespan of a Tamagotchi.

Ah, simpler times.

The premise of the post was simple. The triggers of generative AI were the things these AI engines were directly able — or expected — to do:

Given those capabilities, it was easy to identify the zero-order effects of what would be produced:

These were the oft-quoted greatest hits of AI’s promised nirvana, but honestly kind of vague. What would being able to do all that mean? This is where things got interesting. It was relatively straightforward to extrapolate the first-order effects that would follow. To pick a few from the table above:

I’d say most of those are playing out as written. But you didn’t need to be Nostradamus to see that future. Many others in our industry were making those same predictions.

What I wondered was what would happen as a result of all that? Newton’s Third Law of Motion: every action has an equal and opposite reaction. In a derivative Brinker’s Third Law of Martech — I’m totally making that up — the reaction might not be exactly equal, but there will be a reaction. Those reactions are the second-order effects.

Second-order effects aren’t as immediately obvious, but they still follow a logical chain. And while everyone else is watching the splash, there’s opportunity in the ripples.

So let’s revisit these second-order effects now three years later…

The first-order effects of AI content are about production. The second-order effects are about consumption. To paraphrase General Jim Mattis, the buyer gets a vote. And they get a vote with their own AI.

Email clients are increasingly sorting, filtering, summarizing, and protecting inboxes. The iPhone is happy to agentically screen unknown callers before they interrupt you. Social feeds ruthlessly compress attention around what their algorithms believe will earn a glance.

Across channels, the buyer’s environment is becoming more defensive.

Meanwhile, the supply side is getting noisier. AI-written outbound emails, AI-personalized LinkedIn messages, AI BDR sequences, AI-generated nurture streams — all of it raises the volume. Some of it works when it is grounded in real intent, relevance, timing, and judgment. But generic AI mass personalization has quickly become coded as spam. “I noticed your recent post…” is the new “Dear {FirstName}.”

It’s a kind of negative flywheel. More seller-driven noise pushes buyers to be more aggressive in tuning it out. That reduces its efficacy, which naively leads sellers to try harder to break through, further increasing the noise, and so on. The escape hatch from this game-theoretic journey to hell is buyers asymptotically going full-pull.

The old game was to publish marketer-controlled content, optimize it for Google, and force guide buyers through your carefully arranged journey, form-gated and oh-so automatically nurturing. Now, more and more buyers ask ChatGPT, Claude, Gemini, or Perplexity for the answer they want, in the format they want, at the moment they want it. Summarize this. Compare these vendors. Give me the tradeoffs. Draft the RFP questions. Tell me what the marketing copy is not saying.

Buyer-side AI agents are still in their infancy today. They mostly help search, summarize, and compare. Tomorrow, they will intervene more actively: filtering outreach, interrogating claims, checking references, negotiating terms, monitoring usage, and recommending next-best actions on the buyer’s behalf.

The most effective counter sellers can intentionally build is trust. Human credibility matters more. Distinctive points of view matter more. Communities, ecosystems, peer networks, and voices from practitioners with actual scars matter more. (I’d add respected industry analysts who write snappy newsletters, but that might come across as self-serving.) This is harder to build than a clever prompt. But that’s its power.

The first-order effects of AI generating code are about creation. The second-order effects are about the changing lifecycle of software.

For most of software history, creating an app was expensive enough that, once you had one, you were kind of stuck with it. Even mediocre apps accumulated gravity. They got patched, extended, integrated, documented, defended, and maintained long past their sell-by date because replacing them felt harder than living with them. Hello, tech debt.

But the abundance unleashed by AI-generated code flips that script.

Whether we call it AI-assisted development, vibe coding, or just “I made a thing because I needed it this afternoon,” software is becoming radically easier to spin up. Not just by developers, but by marketers, ops teams, analysts, agencies, and anyone with a problem or idea specific enough that no commercial SaaS app quite fits.

This produces a hypertail of custom software: thousands of tiny, situational, purpose-built agents, tools, scripts, skills, automations, dashboards, etc. Most won’t become “apps” in the formal sense. They are closer to disposable mechanisms. Useful for a moment, a campaign, a team, a workflow, a customer segment, a quarter… and then thrown away.

Some might shudder, imagining a tsunami of tech debt from this hypertail of a million micro apps, agents, and automations. But ironically, disposable software might be the antidote to tech debt. When it’s easy to spin up something new, tailored exactly to what you want in the present, then why waste time wrestling with something from the past?

I know, I know. There’s an obvious are-you-nuts objection to reinventing the wheel every morning before breakfast. But this is where the distinction between infrastructure and apps/agents — see my previous articles here and here — matters. Infrastructure is the solid, stable foundation that isn’t vibe coded on a whim. But it enables much more fluid apps and agents to be dynamically spun up, while inheriting the core structure and services of the business that shouldn’t be rebuilt every other day.

The bigger challenge, which we have yet to solve, is the governance and orchestration of these myriad of apps and agents running throughout the org. Gartner predicts the average Fortune 500 enterprise will run more than 150,000 AI agents by 2028 — up from fewer than 15 in 2025. While AI is closing the capability gap of what people always wanted to be able to do, it’s creating an immense visibility gap in its wake. This is how I illustrated it in that post three years ago:

But as many aspiring entrepreneurs and the VCs who love them will note, a Big Hairy Problem like this is also an incredible opportunity. A massive platform play awaits. If “context” is martech’s word of the year, the runner-up will be “coherence.”

Speaking of context. The first-order effects of AI absorbing content and data are about delivering answers. The second-order effects are about the supply chain behind them.

When buyers ask AI instead of you, what the AI knows about you is effectively part of your marketing. So feeding AI answer engines accurate, structured, current, machine-legible facts about your company and products, while orchestrating legitimate trust signals from your communities and ecosystem at large, is becoming a discipline of its own. (Can we please settle on AEO or GEO? I like AEO because the other sounds like internationalization. But honestly I don’t care which, as long as it’s just one.)

But the bigger opportunity is inside your own go-to-market operating system.

Every copilot, agent, and automated workflow in your GTM operating system is only as good as the data and context it can draw on. Point a brilliant agent at fragmented, stale, contradictory data and you get fluent nonsense, delivered with impressive confidence. Your data foundation and the context engineering (here’s to context-as-a-service!) to wield it determine the effectiveness of everything downstream. Garbage in, garbage out hasn’t been repealed. It’s just that the garbage is now beautifully written.

This is the case I made in The New Martech “Stack” for the AI Age for a composable canvas of apps and agents, built on a universal data foundation, with a strong semantic layer binding it together. The semantic layer is the unglamorous but epic hero of that architecture, ensuring a thousand agents don't answer with a thousand interpretations. (Coherence, again.)

Massively democratized analytics in plain English have arrived. Plainly agreed upon semantics have yet to catch up in most companies. I’d humbly suggest that defining many of these semantics is marketing, as much as any brand guidelines.

Note that this blossoming data marketing channel doesn’t stop at your four walls. The data and context you exchange with your ecosystem partners increasingly shapes how well the AI in each of your environments serves the customer journey you’re both a part of.

The first-order effects of AI executing digital tasks are about delegation. The second-order effects are about distribution.

In my original post, I suggested that API services would become a first-class marketing channel, serving the AI agents who are serving their humans. What’s remarkable is just how quickly the industry converged on a standard for it. MCP went from Anthropic side project to the lingua franca of agentic AI in barely 18 months.

The biggest martech platforms have read the room. Salesforce went fully headless, exposing data, workflows, and business logic to agents with no UI in the loop. HubSpot committed to full API parity — anything a human can do in the app, an agent can do through the API. Dozens of others have shipped MCP interfaces of their own. I covered the competitive dynamics of this in the martech infrastructure wars, so I'll spare you the rerun.

The point here is that a marketing platform’s capabilities are no longer confined to its own interface. They can be distributed through whatever agent the customer prefers to work with. Distribution of functionality, unbundled from UI. The most forward-thinking companies are actually using MCP as a channel.

Bot commerce, admittedly, hasn’t arrived yet. Buyers are delegating research to agents, but the purchase still mostly ends with a human at a checkout flow designed for one. I think it’s coming, though, because this is the natural trust ladder of delegation. First we let agents gather, then recommend, then transact. The rails are being laid ahead of the trains. And in B2B, where “commerce” means evaluations, negotiations, procurement, renewals, etc., we’re arguably already further along with buyer-side agents in the mix.

This is also where all four second-order effects collide in Big Ops (a much bigger deal than Big Data).

AI-generated content creates the trust problem. AI-generated code creates the visibility problem. AI-generated answers create the context problem. AI-executed tasks create the authority problem. Who sees what? Who can do what? Which data is canonical? Which agent is acting on whose behalf? What gets logged, approved, escalated, rolled back, or retired? Big Ops is the operating discipline for coherence across all of that.

Left as an exercise for the reader: what do you think the third-order effects will be?

Scott

P.S. Please check out this cool diagnostic on leaks in your acquisition spend from my friends at BlueConic, who kindly sponsored this week’s newsletter.