Welcome to AI for Messy Jobs: Weekly field notes from inside $3M+ businesses with complex operations, messy physical realities, and the operators trying to modernize it all. Fast.
I was on a call this week with the owner of an 80-year-old, family-owned equipment company. We were talking about the challenges facing the business, and where AI might help address some of the existential ones.
About half an hour in, Alan looked at me through the screen, eyes twinkling.
“Cris, can I tell you a story? Because I love storytelling.”
When he was in business school “a million years ago,” he watched a new building go up on campus. When it was finished, there was a grand opening, and the students moved into their magnificent new classrooms.
But there were no sidewalks.
Alan thought that was strange. A year later, when sidewalks finally appeared, he asked someone why they’d waited.
“We always wait to see where the paths form before we pour the sidewalks,” he was told. “Otherwise, we spend the next 30 years redoing the landscaping.”
His point: how do you know what’s worth automating before you know what works?
I hear some version of this from business owners every week. Sometimes it’s concern about getting locked into a tool. Sometimes it’s reluctance to invest in something that might be obsolete next year. And sometimes, as with Alan, it’s a more fundamental question:
What actually makes this a better business, and how will we know?
It’s tempting to treat those questions as resistance to change. But Alan wasn’t arguing against sidewalks. He was asking where to put them.
The challenge isn’t just figuring out what AI can do. It’s figuring out which paths are worth making permanent.
I love this question. And frankly, as any business owner knows, a huge part of the job is dealing in ambiguity.
That’s where I turn to product thinking: make the problem specific, separate evidence from assumptions, and run a small test before committing to a permanent solution.
Step 1: Clarify the problem
“Use AI to improve sales” is an ambition, not a clear problem. Before choosing a tool, we need to identify who’s struggling, when, and how that struggle shows up.
I use this template:
[Who] struggles to [do what] when [moment or situation], and we see it in [observable signal].
For Alan’s business, the working problem statement was:
Our sales team struggles to confidently price and describe incoming equipment because they lack consistent, verifiable information. We see this in how they price, position, and ultimately sell (or fail to sell) each piece.
That gives us something concrete to investigate. It also keeps us from assuming AI is the answer before we understand the problem
Step 2: Separate what we know from what we assume
Next, we sort our understanding into three buckets:
What do we know? The team relies heavily on personal experience and gut judgment, with limited supporting information about incoming inventory.
What do we think we know? Historical sales of comparable equipment could help them make better pricing and positioning decisions.
What do we need to learn? Would combining those historical comparisons with current market data actually change their decisions—and improve results?
The distinction matters. “More information would help” sounds reasonable, but it’s still an assumption. The information might be incomplete, hard to interpret, or irrelevant to the decisions that matter.
Before building a system to deliver it, we need to test whether anyone can use it to make a better decision.
Step 3: Turn the riskiest assumption into a test
A useful hypothesis connects a specific change to an observable outcome:
We believe [change] will produce [outcome]. We’ll test it by [small experiment], measure [metric], and consider it promising if [threshold].
For Alan’s business, that could look like:
We believe giving the sales team relevant historical sales and current market comparisons will help them price and position equipment more effectively. We’ll test this in one inventory category by preparing a simple brief for each incoming item and tracking whether, and why, it changes the team’s decisions.
The first test doesn’t need to prove the entire business case. It needs to answer the next important question: does this information meaningfully improve the team’s judgment?
If it does, we can then test whether those decisions translate into better margins, faster sales, or another agreed business outcome. We’d define the success threshold before starting, not after seeing the results.
And we don’t need to automate the whole process to find out. We can assemble the briefs manually, learn what the team actually uses, and build around that.
The goal isn’t to eliminate uncertainty before acting. It’s to make the next commitment small enough to learn from.
Let the paths form. Then pour the sidewalks.
This doesn’t have to take months or become a separate initiative. The goal is to make it habitual: name the problem, question the assumptions, and run a small test before committing to a bigger change, whether that’s a new tool or a new process.
— Cris
P.S. This is how our embedded AI operations partner model works: we work alongside your team to identify problems worth solving, test what helps, and put the useful pieces into practice. Curious what that could look like in your business? Email me.



