← Field Notes

Operations

You don't have an AI problem. You have a process documentation problem.

A pattern I run into constantly: a company wants to automate something with AI, we start scoping it, and within twenty minutes we've stopped talking about AI entirely. We're trying to establish how the work currently gets done — and nobody can tell me.

Not because they're disorganized. Because the process lives in three people's heads, has evolved for six years, and has never been written down.

This is the real bottleneck at most companies between 20 and 200 people, and it's badly underdiagnosed. It gets mistaken for a technology gap, a budget gap, or a talent gap. It's none of those.

Why this stops AI cold

To automate or augment a process, you have to be able to describe it. What triggers it, what steps occur, what decisions get made and on what basis, what the exceptions are, and what "done" looks like.

Most companies can describe the happy path in about a minute. The trouble is that the happy path is often only 60% of the actual volume. The rest is exceptions — the client on a legacy contract, the request that needs the owner's approval, the vendor whose invoices arrive in a different format, the situation Dana handles because Dana has always handled it.

Those exceptions aren't documented anywhere. They're institutional knowledge. And an AI tool implemented against the happy path alone will produce confidently wrong output roughly 40% of the time — which is far worse than no automation, because now someone has to catch the errors.

The tell

You can spot this without any formal analysis. Ask three people who work on the same process to describe it independently.

If you get three different answers, you don't have a process. You have three people improvising toward a similar outcome, which works fine right up until you try to automate it, hire someone new, or lose one of the three.

I've never seen this exercise fail to be illuminating, and I've rarely seen a leadership team predict the result correctly.

Why nobody has fixed it

Documentation is thankless. It takes real time, produces nothing visible this quarter, and the people who could do it best are the ones already busiest — precisely because they're the ones holding the undocumented knowledge.

So it perpetually loses to urgent work. Reasonably, in the short term. It's just that the cost compounds quietly: onboarding takes months, quality varies by who's working, and every improvement effort — AI or otherwise — stalls in the same place.

The good news

You don't need to document everything, and you shouldn't try. Enterprise-style process mapping is expensive, slow, and produces documents nobody reads.

You need to document the two or three processes you're actually trying to improve. That's a bounded, achievable exercise.

And the definition of "documented" is lower than people assume. You need the trigger, the steps in order, the decision points and their criteria, the known exceptions and how they're handled, and what finished looks like. One or two pages per process. Not a flowchart in specialist software — a document a new hire could follow.

The most efficient way to produce it is to have the person who does the work talk through it while someone writes. An hour per process, plus a review pass with a second person who does the same work — which is where you'll discover the disagreements.

The unexpected return

Here's what surprises people. Companies that go through this frequently find that documentation alone recovers meaningful value, before any tool is purchased.

Writing a process down forces you to look at it. You find steps that exist because of a system you replaced in 2019. You find two people doing overlapping work. You find approvals that add three days and no value.

I've watched teams eliminate 20% of the steps in a process purely by looking at it in writing for the first time. That's a return on an afternoon, with no software involved.

Then, when you do bring in AI, you're pointing it at a process you actually understand — which is the difference between an implementation that holds up and a pilot that evaporates on contact with reality.

Where to start

Pick the process that's costing you the most time and irritating people the most. Have someone talk it through while a colleague writes. Review it with a second person who does the same work. Fix the obvious redundancies you find.

Then ask what AI could do for it.

In that order. The reverse order is how the 95% got there.

---