I like to say that businesses are a little like stoves with a thousand burners. At any given moment, one burner has a big fire.
A customer has a problem, someone quits, a report is wrong, sales needs something, operations discovers a bottleneck.
Processes are created to put out the fire and make sure that specific fire never happens again. Then another burner catches fire, and everyone runs over there.
And nobody ever goes back and asks whether the process they invented six months ago (or 6 years ago) still makes any sense.
That is how most business processes are created.
It’s also why I think a lot of companies are approaching AI backwards.
They’re buying tools before they understand the work.
Process work is boring and sometimes exhausting
There’s a reason companies avoid this.
Actually understanding a process is difficult.
Who is involved? What happens at each step? What technology is being used? What information goes in? What is supposed to come out? Where are the exceptions? Where does someone need to make a judgment call? Where does work sit and wait?
It’s laborious.
I bet it’s so laborious that you stopped reading all those questions and skipped down to the next new paragraph.
And the frustrating part is that simply documenting a process doesn’t necessarily transform a company overnight.
Ask a company that just finished documenting all of its SOPs. They’re usually glad they did it, but revenue doesn’t suddenly jump.
Process work is foundational work. Transformations work best when laid on solid foundations, but that doesn’t mean laying a solid foundation is fun or energizing.
Willy-nilly buying AI is nothing new
Companies have been buying the shiny-object-of-the-day technology for much longer than AI has been around.
Now there’s a shiny new tool released every five minutes.
So instead of doing the dirty work, companies start buying.
Sometimes it works, but adding another tool to a bad process doesn’t necessarily make the process better. Sometimes you’ve just added another tab.
There’s another option now: build
Historically, custom software was expensive enough that most companies had a very high bar before considering it. That calculus is changing.
Tools like Claude Code, Cursor, and Codex have dramatically lowered the cost of creating software.
Anecdotally, a buildout that cost $500,000 5 years ago probably costs closer to $50,000 now, and a $50,000 buildout probably costs closer to $20,000. Seriously. There is a limit—usually we don’t see buildouts costing less than around $15,000.
That doesn’t mean everyone should start building every piece of software they use. Most of the time, systems of record have so many requirements that it’s easier to integrate with them than it is to replace them.
A few quarters ago, everyone in Silicon Valley was talking about using AI to replace their CRM or their Slack instance. Not only has that not happened, but Meta recently moved their internal communications onto Slack.
But there are more situations now where a business is paying for a piece of software and twisting its process into unnatural shapes just to make that software work. For example, we’ve seen several instances of Monday.com that are completely unnatural.
That’s where custom automation can be the right fit. Instead of changing the workflow to suit the tool, you can increasingly build the tool to suit the workflow.
Of course, it’s got to be ROI-positive. $500/month for Monday.com is a lot less expensive in the near term than $20,000, so the human must be saving tons of time to make it work. That said, we’re increasingly living in a world where the calculus changes.
The order matters
The sequence I’d use is:
Understand the problem. Understand the process. Identify the bottleneck. Then select or build the technology.
AI makes technology more powerful.
It also makes software cheaper to build.
Neither of those changes the fundamental rule that you need to understand what you’re trying to fix.
The companies that get the most from AI will be the ones that understand their operations well enough to put AI in exactly the right places.