Almost every quality tool sits inside one loop:
- Plan. Decide what to change and why.
- Do. Try it.
- Check. See what happened.
- Act. Keep it, adjust it or drop it.
It grew out of Walter Shewhart's work and was spread by W. Edwards Deming. ISO is organized around it. The Lean Enterprise Institute describes Toyota's system improving through PDCA, which it also calls the scientific method.
Deming's Study
Deming preferred Plan-Do-Study-Act. The Deming Institute describes it this way: in Plan you predict what will happen. In Do you try it on a small scale. In Study you compare what happened against your prediction. In Act you adopt, modify or expand the change.
His point was that "check" can mean "did we do it?" "Study" means "what did we learn?"
Why it matters for bots
Bots change fast. You edit a skill, swap a model, add a check. Each change is a little experiment. Most of the time nobody writes down what they expected, so nobody learns anything.
PDSA fixes that with one habit: write the prediction first.
A bot example
Plan. "The report bot misses owners on about a third of actions. If the skill tells it to pull owners from the action log, I predict blanks drop to near zero."
Do. Change the skill. Run it on three weeks of old data, not live.
Study. Blanks dropped from 12 to 2. The two left were actions with no owner in the log either. So the prediction was mostly right, and the remaining problem is upstream.
Act. Adopt the change. Open a new item: actions need an owner when they're logged.
Save all four steps in a short file. That's your record. See Write it down or it didn't happen.
Small scale first
"Do" means try it small. For bots: test on old data, on one routine, or with a tester bot before it touches live work. Then roll it out.
Where bots differ
People often skip Study because it's slow. A bot can run the comparison quickly. But a bot studying its own change is grading its own homework. Keep the measure simple and checkable, and have a person look at the result.
Back to Engineering tools for bots.