Where human dependency actually hides
Mapping a fragmented operation rarely finds the bottleneck where the process document says it is.
The documented process is not the operation
Every operation has two versions: the one in the process document and the one that actually runs. Mapping work means recording the second, and the reliable way to find the difference is to ask what happens when something is unusual — because the document almost always describes the ordinary case, and the operation is largely shaped by the exceptions.
The documented steps tend to be the ones with a system boundary: a form submitted, a record created, an approval clicked. Those are legible because they left a trace. The undocumented steps are the ones that happened in someone's head between two systems.
What the hidden steps look like
They are acts of interpretation. Someone reads an inbound request and decides which of three categories it belongs to, using judgement they have never had to articulate. Someone notices that a particular account's records are unreliable and silently checks a second source. Someone knows that requests from one team always arrive missing the same field, and fills it in without mentioning it.
None of these appear in a workflow diagram. All of them are load-bearing. And each is a place where automation will either fail outright or, worse, succeed at producing plausible output that is wrong in a way nobody notices for a while.
The tell is a step that takes almost no time but can only be done by particular people. Duration is not the signal; substitutability is. A task that takes an hour and could be done by anyone is not a dependency. A task that takes ninety seconds and only two people can do is the thing holding the operation together.
Why this matters before deployment rather than after
An automated workflow inherits every undocumented judgement it replaced, whether or not anyone noticed. If the judgement was necessary, its absence shows up as a quality problem attributed to the model. If it was compensating for a data problem upstream, removing the person removes the compensation and exposes the original defect — usually some distance downstream, where it is expensive to trace.
This is why mapping precedes design rather than accompanying it. The question is not only where automation could go. It is which human judgements are doing real work, which are compensating for something that should be fixed instead, and which are habits nobody has revisited.
The first category becomes an explicit approval step. The second becomes a defect to repair before any autonomy is introduced. The third is the only one that simply disappears — and it is usually the smallest of the three.