When a process mixes people and machines, the biggest risk is not that the machine gets it wrong. It is that nobody knows whose job it was. Ambiguity in the split gets paid for in errors nobody catches, work done twice, and people who stop trusting the system.
Three categories, no more
For each step in the process there are only three options, and it is worth writing them down:
- The machine does it. Repetitive tasks with clear rules and limited consequences if they go wrong. Nobody checks them one by one; they are checked by sampling.
- Both do it. The machine prepares, the person decides. This is where most of the value sits and also where the design is most often neglected.
- The person does it. Decisions with real consequences, new cases, difficult conversations, and anything where somebody has to answer for the outcome.
What you cannot have is an implicit fourth category that in practice exists in plenty of companies: “the tool does this and somebody will look at it eventually”. That is not a split, it is a hope.
Four questions to decide
When we are unsure about a particular step, we use these four questions. If any answer is uncomfortable, that step cannot be left to the machine alone.
- What happens if it gets it wrong and nobody notices for a week?
- Could somebody outside — a client, an inspector, a judge — ask us to explain this decision?
- Does it need judgement about something that is not written down anywhere?
- Is there a person who will feel handled by a machine at a delicate moment?
Reviewing is not reading everything
A common trap: it is decided that a person will review everything the machine produces. Because the volume is high, within two weeks that person approves it in bulk without reading it. Now you have the worst of both worlds: the risk of automation and the cost of control.
Review works when it is bounded. You review everything above a threshold, everything the machine itself flags as uncertain, and a random sample of the rest. And you measure how often the reviewer changes something: if they never change anything, the review is unnecessary; if they change half of it, the step was not ready to be automated.
A review that always says yes is not a control. It is a formality that costs money.
Writing it down changes things
We always leave it as a one-page table: step, who does it, what triggers it, who reviews it and who it escalates to when it fails. It looks like bureaucracy and it is the opposite: it is what lets people use the system without asking and without fear.
That table also helps with the difficult conversation. When someone sees in writing which of their tasks move to the machine and which new ones they take on, uncertainty drops a lot. What creates resistance is almost never losing a boring task. It is not knowing what comes next.