Anyone who has been here for awhile knows that my 15+ years in startup life has shown that chaos is very much a stage in the lifecycle of early companies. My argument is that it can be a stage, but it cannot be a strategy.
There is a wild thing happening these days with our access to AI and agents who can amplify all of the angles of our work. We’re choosing to lay AI on top of workflows that make ZERO sense. We want to keep moving fast and breaking things and boy are we cruising.
Sales has to ask CS exactly how to do the new client handoff every time.
Marketing has new sales enablement docs, but no one knows where to share and prep the sales team.
No one knows who owns renewal follow up.
There are important decisions stuck in 10 different Slack channels.
Obviously the natural and ideal next step for this is: LET’S GIVE ALL OF THESE JOBS TO AN AI AGENT!
We’ve made it to this new spot where we aren’t just asking AI to write us emails. Now we’re actually handing them the work. Which in some ways is pretty cool. As a solopreneur, I can truly say AI has done absolute wonders for my ability to make an impact and move fast. However, I’m not layering that agent on top until I’ve got clarity on what the actual F the agent should be doing. What’s that outcome supposed to be?
Your AI agent is the world’s most literal new hire
Imagine hiring a new employee and explaining that when a deal closes, Sales hands it off to CS. Usually. There is technically a process documented in Notion, although you’re not positive it’s the current one. Sarah knows what actually happens, so they should probably ask her. Enterprise customers are different for reasons no one has written down, and Mike knows about those. Billing has its own thing going on entirely.
Then you send this poor person off to work and three weeks later you’re annoyed that they’re not performing. (Not gonna lie, I have seen this exact situation an alarming amount of times.)
We would (or at least should) recognize pretty quickly that the new hire isn’t the problem. We handed them a scavenger hunt and called it onboarding.
Yet we’re starting to do essentially the same thing with AI agents, except our expectations are somehow HIGHER. We’re expecting them to walk into years of tribal knowledge, conflicting documentation, weird exceptions, and half-decided workflows and somehow emerge with a beautifully automated company on the other side.
The technology may be smarter than your average new hire, but it still cannot read Sarah’s mind. Which is unfortunate, because Sarah is apparently holding half the company together.
Humans have been hiding how broken some of this is
I actually think this is one of the more interesting things AI is exposing. Humans are remarkably good at compensating for bad systems. We just kind of remember that one customer gets special pricing. We know that when the founder says something is “approved,” there is still one more unofficial check before it actually goes out. We know which spreadsheet is the real spreadsheet and which one everyone continues updating for reasons lost to history.
None of that feels particularly dramatic when you’re a team of eight people. Everyone talks constantly, the founder is sitting in half the meetings anyway, and when something gets stuck, you yell across Slack until the right person fixes it. The company keeps moving, so you assume the system works.
Except there isn’t really a system. There are a bunch of smart people filling the spaces where one should be.
That distinction becomes much more obvious as the company grows, but AI is giving us another way to see it. The second you try to hand one of these workflows to an agent, you have to explain what actually happens. Suddenly “Sarah knows” isn’t enough. You have to decide what triggers the work, where the agent gets the information it needs, which version of that information it can trust, what it’s allowed to do on its own, and when a human needs to get involved.
Basically, you have to make all of the decisions you’ve been successfully avoiding while everyone was busy moving fast.
Oops.
AI might be the best operations audit we’ve gotten in years
This is where I think things get genuinely useful.
If you’re trying to build an agent and find yourself unable to explain exactly what it should do, I wouldn’t immediately assume you have an AI problem. You may have just discovered an existing operational problem that your humans have been politely working around for the past two years.
An agent forces those shadow things into the open because it needs more than “you know what I mean.” It needs an outcome and enough structure to reliably get there.
This is the part of the AI boom that makes me laugh a little.
For years, the decidedly unsexy side of startup life has been documenting how work happens, creating clear ownership, cleaning up information, and making sure people know what happens next. Startups wanted to talk about growth and innovation and speed. Nobody was breathlessly posting about their newly clarified decision rights.
Now we want fleets of intelligent agents moving work seamlessly through our companies, and what do those agents need?
They need us to know how the company works.
We have somehow built some of the most sophisticated technology humans have ever had access to and automated ourselves directly back to basic operations.
(Gosh, I love a Seinfeld-style full-circle moment.)
I still want the agents
None of this is an argument for being precious about AI or waiting until every process is perfect. That would be equally ridiculous, and very much NOT how startups work. There will always be some mess. There should probably be some mess. A ten-person company does not need to spend six months documenting itself before anyone is allowed to automate an email.
I am extremely bullish on what this technology can do for small teams. I’m already living a baby version of it myself. As one person, I can research, create, organize, analyze, and execute at a level that would have required considerably more time or help even a few years ago. I don’t want to give that back.
I just think we’re getting the order of operations wrong when we expect the AI to create clarity we haven’t created ourselves.
The companies that get the most leverage out of AI probably won’t be the ones with the greatest number of agents running around. They’ll be the ones that understand their work well enough to know exactly where an agent can make it better. They know what outcome they’re trying to create, what information matters, where human judgment still belongs, and what can safely disappear into the background.
That’s when AI gets really exciting to me. You’re no longer asking technology to compensate for the chaos. You’re giving it something good to amplify.
If your Sales-to-CS handoff makes no sense today, giving it to an AI agent doesn’t magically make it a system.
You’ve just given your chaos a company email address.
And let’s be real, it already had enough access.