I was helping another team review their document verification automation recently. They'd already built something and wanted a second pair of eyes.
The workflow: AI extracts all the data from two sets of documents using OCR. Lays everything out in a neat table. A person reviews the extracted data. Clicks "next." Then the system matches the two datasets and flags discrepancies.
Honestly, it was good work. But something bugged me.
Why does the human review the extracted data before the matching happens? The whole point is to find mismatches. The machine can compare both datasets in milliseconds. Show me the five mismatches and let me deal with those. Why am I eyeballing 200 line items just to feel involved?
They solved "the documents are scattered" but kept "the human reviews everything": a step that only existed because the old process needed a human to do the comparison manually.
Nobody questioned it. They just made it faster.
I keep running into this. Someone takes a 5-step human workflow, bolts AI onto each step, calls it automation. The steps get faster. The process stays the same.
The original process was designed around human limitations. Humans can't hold two 200-row datasets in their heads, so we built a step where you review one first, then the other, then compare. These steps made sense when humans were doing the work. They make zero sense when machines are.
Elon's 5-step framework
Elon Musk has this 5-step engineering process he talks about publicly. The order goes:
- Question the requirements
- Delete the part or process
- Simplify
- Accelerate cycle time
- Only then automate.
Automate comes last. And step three comes with a warning I think about a lot: the most common error of a smart engineer is to optimize something that should not exist.
He tells a story from the Tesla Model 3 line. They had fiberglass mats on top of the battery pack (fluff, meant to dampen noise). Someone decided to automate placing them. So they built the "flufferbot" complete with machine vision. Robots are terrible at picking up fluff. The line kept breaking down.
Then someone finally asked: does the fluff actually reduce cabin noise? They tested with and without. No difference.
They'd built a sophisticated robot to automate the placement of something that didn't need to be there.
Why people skip steps 1 and 2
Questioning a process means questioning the people who built it. It means telling a team that the workflow they've been running for three years has steps that shouldn't be there.
It also means a smaller, simpler automation. One that looks less impressive in a demo but does more in production.
It's easier to say "we automated your entire process" than "we deleted half your process and automated what's left." I've been in those rooms. The second one is harder to sell. But it's almost always the better thing to build.
My low-tech trick
Before I automate anything, I try to explain the entire process to someone (or something) with no context. AI works well for this, not because it's smart, but because it has zero attachment to how things have always been done. It just asks why.
Half the time the AI doesn't even need to reply. The act of writing out every step (what it does, why it's there) forces me to confront the gaps myself. Rubber duck debugging for business processes.
And when I can't talk myself into why a step exists, I just remove it and see what breaks. Musk's own rule: if you're not adding things back at least 10% of the time, you're not deleting enough.
Most of the time, nothing breaks.
Originally published on LinkedIn. Want to discuss this further? Get in touch.
