The Real AI Divide: Operators vs. Observers

Fifty-seven percent of enterprises now run AI agents in production. The other forty-three percent are still running pilots, attending demos, and filling Notion boards with "AI strategy" documents that nobody reads.

I lead a team of about 15 people across five businesses. Every single one of them uses AI daily. Not as a novelty. As their operating system. Engineers review AI-generated pull requests before they ship. Our marketing automation platform runs on email infrastructure we built ourselves. Our voice AI assistant ships features that started as Claude Code sessions at 10am and hit staging by lunch.

We call this being an AI Operator. And the gap between operators and observers is widening every week.

The Pilot Trap

Here's what I keep seeing: a founder reads about AI agents, gets excited, runs a proof of concept over a weekend. It works. They post about it on LinkedIn. Then nothing changes in how their business actually runs on Monday morning.

The data backs this up. Langchain's 2026 State of Agent Engineering report found that quality issues block 32% of teams from getting AI into production. Teams without proper evaluation tools get 6x fewer projects live. Without governance, that number drops to 12x fewer.

The pattern is always the same. Excitement, pilot, friction, abandonment. Back to doing things the old way.

The problem isn't the technology. The technology works. The problem is that most teams treat AI as a feature they're adding to their company instead of changing how the company operates.

What Operating Actually Looks Like

I'll give you a real example.

Anthropic published internal data showing their engineering team uses Claude Code in 59% of their daily work. The result: a 67% increase in merged pull requests per engineer per day. But the interesting part isn't the speed gain. Their engineers shifted from using AI for simple tasks (which dropped from 14% to 1%) to complex, multi-step work (which grew from 10% to 37%).

That's the pattern that matters. The tool didn't just make existing work faster. It changed what kind of work was possible.

I see this across our own businesses. A coding session that would have taken a developer two full days gets scaffolded in an hour. But the developer still reviews every line. Still makes architecture decisions. Still catches the edge cases that AI misses. The AI handles the repetitive build work. The human handles the judgment.

Think of it like the Iron Man suit. Tony Stark doesn't become less skilled when he puts on the suit. He becomes someone who can perform at a level that shouldn't be possible. Same human. Wildly different output.

That's the model we run across all five businesses. Same small team. AI on every desk. And it lets us ship at a pace that companies five times our size struggle to match.

The MCP Explosion (and Why It Matters for Small Teams)

If you haven't been paying attention to MCP (Model Context Protocol), start now.

MCP went from a niche developer concept to being on every CIO's agenda in a matter of months. RSA Conference 2026 submissions were packed with MCP-related topics. Enterprises are racing to figure out how to let AI agents actually do things: read data, trigger actions, manage workflows, all through a standardized protocol.

Here's why this matters if you run a small team.

Before MCP, connecting AI agents to your systems required custom engineering for every integration. That favored companies with big dev teams and big budgets. MCP changes that equation. It shifts integration from bespoke code to configuration. Which means a ten-person team can now wire up AI agents across their entire stack in days, not months.

But the flip side is real too. A Stacklok survey of retail companies found that 56% cite data quality as the top obstacle, 49% struggle with legacy system integration, and 45% flag security concerns. Randy Bias noted that financial firms are running full cybersecurity questionnaires before letting MCP servers anywhere near production systems.

The technology is moving fast. Governance is lagging behind. And that gap is where most teams get stuck.

Three Things Operators Do Differently

After running this way for a while, and watching dozens of other founders try and fail to operationalize AI, I've noticed a few patterns that separate the teams that actually ship from the ones that just experiment.

They build process around AI output, not just AI tools. It doesn't matter which model you use if nobody on your team reviews what it produces. We treat every AI-generated pull request the same way we'd treat one from a junior developer: it gets reviewed, tested, and approved by a senior engineer before it ships. The quality comes from the process, not the prompt.

They measure differently. Most founders track whether they're using AI. Operators track what AI changed. Hours reclaimed. Pull requests merged. Features shipped. Support tickets resolved. If you can't point to a number that moved because of AI, you're observing, not operating.

Newfront, an insurance brokerage with about 50-100 people, integrated Claude for contract review, HR processes, and knowledge management. Their result: 60% cost savings and roughly a month reclaimed per staff member per year. Those are operating numbers. "We tried ChatGPT in a meeting" is an observation.

They push AI into uncomfortable territory. Zapier embedded AI agents across 800+ internal workflows. 89% of their employees now use them daily. They didn't stop at the easy wins. They pushed AI into customer success, engineering orchestration, marketing, across every function. The adoption wasn't optional. It was operational.

Simon Høiberg, who runs BamSubscriptions, used a single 30-minute Claude Code session to build a full production system: failed payment detection, dashboard generation, bulk email with magic links, duplicate prevention. 809 lines across 7 files. That went live the same day.

The common thread: these teams didn't wait for AI to be perfect. They built the review process, the evaluation loops, and the governance to use it now.

The Observer Problem

The observer path looks productive on the surface. You attend the AI conference. You subscribe to the newsletters. You try the new model the week it drops. You post a screenshot on X.

But your team still runs the same way it did two years ago.

Gartner forecasts that 40% of enterprise applications will include task-specific AI agents by the end of 2026. Deloitte's data says close to 75% of companies plan agentic AI deployment within two years. But only 21% have actually done it.

That gap (plans vs. execution) is the observer gap. And it's getting expensive.

Every month you delay operationalizing AI across your team is a month your competitors are compounding their advantage. The teams that started six months ago are now shipping features, resolving support tickets, and building infrastructure at a speed that late adopters won't be able to close by just "adopting AI" later. The compound effect of daily AI operations is brutal for anyone still on the sidelines.

How to Start (If You Haven't)

Pick one workflow. Not the most complex one. Pick something your team does every day that involves repetitive decisions or repetitive production.

Wire AI into that workflow. Build the review process around it. Measure what changes after two weeks.

Then pick the next workflow. And the next.

That's how you go from observer to operator. One workflow at a time, with real humans reviewing real output, measuring real results.

It took us months to get AI running smoothly across five businesses. There were plenty of weeks where AI made things worse before it made them better. Where a Claude Code session produced code that looked right but broke in edge cases our engineers had to debug. Where an AI-generated draft needed more editing than writing from scratch would have.

The difference is we kept going. We built the process. We trained the team. And now we operate at a level that shouldn't be possible for fifteen people.

The AI divide isn't about who has access to the best models. Everyone has access. The divide is about who actually changes how they work.

Operators ship. Observers scroll.

Which one are you?