Essay · August 2026

The Kaizen of Knowledge Work

Speeding up one step gets you the old process, a little quicker.

When Toyota changed manufacturing, it went after the whole line. The engineers hunted the waste between the stations, the waiting, the rework, the parts piling up half-finished, and they designed it out. A faster machine in the middle of a wasteful line only builds the pile in front of the next bottleneck a little quicker. The gain was in the flow, never in any single station.

Knowledge work is about to learn the same lesson, and most companies are learning it the expensive way. They take the process they already run, find the one step that is slow, and point AI at it. The report that took days now drafts in minutes. Everyone is pleased. And the process still takes about as long as it ever did, because the drafting was never the part that was slow.

If you automate a wasteful process, you get a faster wasteful process. The waste does not leave. It just runs at machine speed now. That is the trap of bolting AI onto the pieces: you speed up the one step you can see, the waiting and the handoffs and the review cycles around it stay exactly where they were, and you have spent real money to move a bottleneck six inches down the line.

The transformation is somewhere else entirely. It is in asking what the whole process would look like if you designed it today, from nothing, with these tools in hand. It means asking, of each step, why it exists at all, which is a very different question from how to make it run faster.

I have watched this inside my own function. A risk assessment that used to take months had a familiar shape: gather the data, wait, sample it, wait, test the sample, wait, write it up, review, revise, report. Most of that timeline was not work. It was waiting and handoffs, the kind of waste a lean engineer would spot in a second. When we rebuilt the process around what AI could now do, the steps that used to run one after another began running at the same time. Gathering stopped being a phase, because the data was already connected and current. Sampling disappeared, because we could test the whole population. What had taken a quarter took a matter of days, and the days that remained went to judgment, which was the only part that was ever really the work.

That is what people miss when they measure AI by how much faster a single task got. Faster is the smallest thing AI does. Underneath the speed sit two larger gains. The first is insight: reach across an entire data set instead of a sample, and you stop asking what went wrong in one place and start seeing the patterns forming across everything. The second is capability you did not have at all before, work that was impossible at any budget, suddenly within reach. A company that stops at "the report is faster" has taken the smallest of the three and left the other two sitting on the table.

None of this happens by accident, and it does not happen one clever prompt at a time. It is kaizen, the discipline Toyota built its house on: keep asking, at every step, what the work is actually for, what in it is waste, and what becomes possible now that the old constraint is gone. The tools are new. The discipline is a century old.

So before you point AI at the slow step, step back and look at the whole line. Ask why the process runs the way it does, which parts are real work and which are just waiting, and what you would build if you started today. Then point AI at that. Speed alone gets you the old process, a little quicker. Redesigning the process is where the transformation has been hiding the whole time.