Good morning.

I sat across from an online business operator last month who'd spent $20k across tools, frameworks, and contractors trying to get AI agents built inside his business.

She picked reasonable tools and hired people who knew what they were doing.

After three months, nothing was compounding.

The agents existed and they ran, but the business moved exactly the same way it had before.

The tools worked fine and the people she hired were competent. The thinking underneath all of it, however, was the problem.

The three patterns that got her there are the same three I keep running into at a $1M+ revenue level.

But the patterns seem to be universal, even at lower revenue levels.

— Sam

These three patterns account for most of the failed agent deployments I see at this revenue level. 

My experience so far with smaller revenue businesses (like, around $500k or so), seem to be very similar and sometimes the same. But most of my experience is at minimum $1M+ and beyond, so I prefer to speak from reality than make guesses.

Every week I talk to online business operators who've invested real money in agents and can't figure out why nothing is compounding. 

The answer is almost always one of these. Each one is easy to diagnose and the fix for each is smaller than you'd expect.

The question I hear most often: "Can AI do what Sarah does?" It makes intuitive sense. “Sarah” is a typical team member, costs ~$65k a year.

The thinking is that if an agent can do her job, that's $65k back on the table.

Sarah does 15 different things:

Some of those are repetitive and structured and low-judgment. Perfect for an agent. Others require context only Sarah has, or a level of nuance that still breaks current agents.

When you try to build "the Sarah agent," you end up with something that does 4 of the 15 things passably and the other 11 poorly.

The question you should be asking is, "Which of Sarah's recurring processes can an agent handle end-to-end without human judgment at every step?"

That's a smaller, more boring question. It's also the one that produces agents that actually work, for real.

The agent gets built and it produces output. Somebody checks it the first week, maybe the second.

By week three, nobody reviews it systematically. By week eight, the output has drifted because the market context shifted, the prompts went stale, and the edge cases accumulated. The team stops trusting it and the agent gets abandoned. The operator writes off agents as overhyped.

The operating cadence failed, not the tool.

Every agent deployment needs a review rhythm:

Someone owns that review. They check output quality, flag drift, and decide whether to adjust or expand or shut it down. Without that cadence, even a well-built agent decays.

Most operators have a rhythm for managing their people and their projects and their revenue but almost none have one for managing their AI systems. That gap is where agents go to die.

An operator deploys a content agent. The first few outputs are solid, so someone starts publishing them directly because the whole point was to save time.

Within a few weeks, quality erodes. The voice is off and the specificity fades. The stuff going out the door reads like what it is: machine output with nobody in the loop.

The same thing happens with research agents and reporting agents and client communication agents. The output looks finished, so people treat it as finished. Then they blame the tool when the quality slips.

The fix is simple and unsexy: agents produce drafts, humans refine and approve and publish. The leverage lives in the 80% of the work the agent handles before the human touches it. A research agent that gives you a 90% complete brief in 4 minutes instead of a 2-hour manual build is the win. The remaining 10% of human judgment keeps the output worth publishing.

The teams that get this right treat their agents the way a senior editor treats a junior writer. The junior writer does the legwork and the editor shapes the final product.

Some of this might sound obvious to you. But too many entrepreneurs are stuck consuming AI hype content and hyperbolic junk from overnight experts, thinking “everything has changed!” and then somehow forget the fundamentals of business that haven’t (yet) been eliminated.

All three of these patterns share the same root:

Operators applying a hiring mental model to an infrastructure problem.

When you hire someone, you hand them a role and trust them to figure it out.

When you deploy infrastructure, you scope it to a specific function, monitor its output, and keep a human in the approval chain.

Agents are infrastructure. The operators who treat them that way are the ones whose systems actually compound.

If any of these patterns look familiar, the fix is an operating cadence.

Until next time,
Sam Woods
The Editor

.