Where AI Actually Helps Process Improvement (and Where It Doesn't)

Every conference I've been to in the last two years has a track on "AI-enabled process improvement," and most of it is the same slide with a different logo. Having actually tried to use these tools inside real transformation work -- and having built this site with an AI coding agent, which is its own small case study -- here's where I've found real value, and where I'd still bet on a person.

Where it genuinely helps

Pattern-finding across data you'd never manually cross-reference. Root-cause work often means looking for correlations across systems that were never designed to talk to each other -- shift schedules against defect logs against maintenance records. That's exactly the kind of wide, shallow pattern-matching AI tools are good at, and it can surface a candidate hypothesis in an afternoon that would have taken a week of manual data-pulling to even attempt.

First-draft structure for recurring artifacts. SOPs, control plans, A3s, and training materials all have a repeatable shape. Having a model produce the first draft of that shape -- which a subject-matter expert then corrects -- is faster than starting from a blank page, and the correction step is where the real domain judgment still has to happen.

Turning messy voice-of-customer or open-text survey data into themes. This used to be hours of manual coding. It's now a first pass that a person reviews and adjusts, which is a much better use of a Black Belt's time than the initial sort.

Where it doesn't

Deciding what the root cause actually is. A model can surface correlations. It cannot walk the floor, watch how the work really happens versus how the SOP says it happens, or notice the thing nobody mentioned because it's normal enough that no one thought to bring it up. That gap between the documented process and the real one is still found by a person paying attention in person.

Stakeholder alignment. No tool gets Operations, Quality, and Finance to agree on a prioritization when they have three different incentives. That's a facilitation problem, not an information problem, and facilitation is still fundamentally human work.

Sustainment. This is the one I'd flag loudest, because it's the one vendors gesture at most confidently. An AI tool can help you design a control plan. It cannot be the leader who asks about the metric three months later and signals to the team that it still matters. If anything, I've noticed the opposite risk -- teams that let a slick AI-generated deliverable substitute for the harder, slower work of building actual ownership.

The honest summary

Treat AI in this work the way you'd treat a very fast, very well-read analyst who has never actually worked a shift on your floor: genuinely useful for the first draft and the wide search, genuinely dangerous if you let it make the judgment calls that require having been there.