Every CEO I talk to right now has an AI strategy. Or at least they say they do. They’ve got Copilot rolling out, a few prompt engineers on the team, maybe a chatbot on the website that’s technically “agentic.” 

It feels like progress. But in lieu of a strategy, it’s a collection of apps with a budget line.

I’ve been inside enough GTM orgs to see the pattern clearly: A handful of individual contributors using AI in their daily work, getting more productive on personal tasks. A couple of execs asking about ROI in QBRs. And zero production systems. Nothing running autonomously. Nothing generating real output without a human pressing a button.

We’re a few years into this thing, and most teams are still tourists — which isn’t a dig! Tourists have a great time. They take photos. They learn stuff. But they don’t build a life there. Eventually they go home. 

The companies pulling ahead right now didn’t invest the most in AI tools. Rather, they stopped treating AI like a tool altogether and started treating it like infrastructure. Like plumbing. Like the way the building works.

That shift is the entire game. And I decided to stop talking about it and actually build it myself.

(Full disclosure: I want to be careful here, because this is where most articles lose people. They either go full sci-fi or full enterprise-speak, and both are useless to an operator trying to figure out what to do on Monday morning.)

Here’s the part that makes my situation different from the typical AI builder story: I’m not doing this at a startup with a blank canvas. I’m doing this inside a company that’s been serving healthcare for over 20 years. We’ve got thousands of customers. Established products. Real revenue. And legacy infrastructure that wasn’t built for an AI-native world.

That’s the reality that most of you are living in, too. You’re not starting from scratch. You’re transforming from the inside out.

What that looked like in practice was building AI-native capabilities that sit on top of our existing products — not ripping and replacing, but evolving. And so, the current platform became the system of record, the data backbone. The AI layer became the brain and the nervous system — intelligence, orchestration, context, memory — that makes it useful in ways a traditional SaaS product never could. And we did it without waiting for some mythical “perfect data architecture” to arrive first.

That same philosophy shaped how I approached our GTM operation. The mental model that helped me most: think of it less as a software stack and more as a parallel organization — one that needs a brain, a team, and a nervous system to actually function.

The brain is a knowledge graph. You can think of it as organizational memory that never degrades, is always learning, and holds everything the system knows: competitive intelligence; product positioning; market research; historical performance data; strategic context from years of accumulated decisions. Any part of the system can pull from it in real time. 

The team is comprised of specialized AI agents organized into departments. Each agent has a defined role and a chain of command. When work comes in, the system routes it to the right agent in the same way you’d assign tasks in a Monday morning standup.

The nervous system is automation connecting everything. Triggers, schedules, event-driven workflows that keep the whole thing running without someone standing over it.

Putting all of this in writing, I recognize that it sounds like a lot. It is a lot. But this isn’t where we started. 

Let me tell you about the moment it stopped being a project and started being a capability: Every week, without any human trigger, the system connects directly to Google Analytics through what’s called a model context protocol (MCP), which allows three different AI agents to interact with live tools and data the same way you or I would interact with a browser.

Agent one pulls all of the raw data you’d normally ask an analyst to export into a spreadsheet: traffic sources, user behavior, conversion funnels, engagement metrics, etc.

Agent two ingests that data and does the actual analysis — not just summarizing numbers, it’s identifying patterns, comparing against previous weeks, flagging anomalies, and building a narrative around what’s happening and why. It then generates specific recommendations that go well beyond the usual, vague “consider optimizing your landing page” kind of platitudes. These are concrete actions tied to what the data actually showed that week.

Agent three formats the output, publishes it to a dashboard, and fires off an email to every stakeholder who needs it.

Total human involvement: zero.

Now, take a moment to think about how this used to work: Someone on the team — usually a marketing ops person or a junior analyst — spends Friday afternoon pulling the GA report. They copy numbers into a slide deck. They add some commentary that may or may not reflect what actually matters. They email it to six people, and three open it. Only one reads past the first chart. By Monday, the insights are stale and nobody acts on them, anyway.

That old process had a leak at every handoff. By the time the insight reached someone who could act on it, half the signal was gone.

In contrast, the new process is faster, and the insights are better because the analysis is consistent, thorough, and completely unburdened by Friday afternoon “just get it done” energy.

I haven’t fully instrumented the ROI across every workflow, currently working down the list. But even rough math is instructive: This one workflow probably recovers 3–4 hours a week in direct labor, plus whatever signal was lost because the insights used to arrive stale, half-formatted, and destined for a deck that three people would open and one would actually read. Annualized, that’s 150–200 hours of recovered capacity to put toward revenue generating activities. 

Mind you, this is one workflow. And I have dozens running at any given time.

(I want to be honest about this part because conference talks always skip it and LinkedIn posts make everything sound inevitable in hindsight.) 

The hardest part of developing this parallel organization within our organization was less of a technical endeavor, and more of a mental shift. I had to stop thinking about individual tasks and start thinking about systems. So, for example, instead of asking “How can AI help me write emails faster?” the better question is: “What’s the full system that produces pipeline? And where in that system are humans doing repetitive work that a well-designed agent could own entirely?”

The resulting data wasn’t perfect. And it’s still not. But I’ve watched too many teams stall out on “data hygiene” projects that become an excuse to never build anything. (Hint: Perfect data hygiene is an unsolvable problem.) The goal should be to build systems that can work with messy reality and that improve the data as a byproduct of running, not as a prerequisite to starting.

Something else: I had to let agents fail. The first version of the analytics workflow was a glorified data dump, which was genuinely embarrassing. But because the system ran every week, it iterated fast. By week four it was better than what any single human on the team had been producing. That’s a plus when it comes to agents: They don’t get defensive when you give them feedback. They just get better.

And I didn’t start with hundreds of them. I started with one workflow, end to end, that was fully autonomous. Then, I added another. Then, another. The architecture emerged from working systems, not from a whiteboard session or a Miro board or a strategy offsite.

The moment I knew this was genuinely different came quietly. It was a Wednesday morning. I was half-awake, scrolling my phone, and there was the analytics report. Clean, insightful, already distributed to the team. With no “hey, can you pull the GA numbers?” prompt from me. It just showed up, like a colleague had done their job overnight.

Because one had.

There’s a massive gap between the AI conversation happening at conferences and what’s actually possible when you sit down and build.

Most of the AI content aimed at GTM leaders right now is either too theoretical to be useful or too tactical to matter. “Here are 10 prompts for your SDR team” is not a strategy. And neither is “AI will transform everything” if it doesn’t show what that transformation actually looks like on any given Tuesday.

I want to show the work. The real architecture. The failures. The wins. The stuff that surprised me. And specifically, what it looks like to do this inside established companies with legacy systems and imperfect data, because that’s where most of you actually live.

If you’re running a GTM org and your AI strategy is still a collection of individual tools and a vague plan to “go agentic,” I want to give you a different vision of what’s possible.

Until then, ask yourself one question (not “Are we using AI?” because everyone already is): What ran this week without anyone touching it?

If the answer is “nothing,” you know where to start. Pick one workflow your team runs every week — the one that’s mostly copy-paste and context-gathering — and make it run without anyone touching it.

Jonathan Moss is an executive with 20+ years of experience in scaling companies from pre-revenue to $15B+. As EVP and GM at Experity ($200M+ ARR healthtech), he built the industry’s first agentic AI platform for patient engagement. He co-founded AI Business Network, co-leads Pavilion’s AI in GTM community, and is Dean of their AI in GTM School. His career focus: transforming legacy industries into scalable, technology-enabled enterprises.

Become a Topline insider by joining our Slack channel.

We want to hear your feedback! Let us know your thoughts.