Design so the mistake is hard to make

Fix the process, not just the person

Managing a team of nine at Vodafone, I more than once had an engineer falling behind and making errors. The easy answer was to manage them out. I treated it as two separate problems instead.

For the person: training, and pairing them with someone already expert in the work. For the process: I checked all of their output for two weeks to confirm real progress rather than assume it, and rewrote the instructions step by step so a step couldn’t simply be missed. That’s poka-yoke, mistake-proofing a process instead of asking someone to be perfect. Nobody was managed out, and the rewritten instructions became the standard reference. That’s Poka-Yoke.

“Human in the loop” is not a control

A lot of AI rollouts rely on one safeguard: a person reviews the output. That’s the same as asking an engineer never to skip a step. It works on a good day. It fails on a busy one, and it fails more as the output gets more convincing.

A better question is the poka-yoke one: how do we make the wrong output hard to produce, or obvious when it is produced?

What that looks like in practice

In Redline, the rule is that a flag can’t be raised without quoting the exact clause behind it. The reviewer doesn’t need to remember to check the source. The source is part of the flag. And when nothing is wrong, Redline says so clearly, rather than staying silent, so the reviewer knows the check ran.

The same idea works on suppliers. When a vendor kept missing commitments, I didn’t escalate harder. I tied payment to validated completion at each site, with mandatory templates checked before sign-off. The correct outcome became the one that got paid. That’s Vendor incentives.

Two habits I’d bring to any AI programme

1. Check everything before you sample. With the engineers, I reviewed every piece of output for two weeks before trusting the process. A new AI workflow deserves the same: full review first, sampling once the evidence supports it.
2. Put the check inside the output. Citations, required fields and explicit “nothing found” states make an error visible without anyone having to go looking for it.

Neither of these is new. They’re how good programmes have handled people and suppliers for years. AI is another part of the system that needs the same discipline.

Fix the person and fix the process, as two separate problems. Most managers only do the first.