AI at the Gemba · Lesson 6 of 10 · Week of November 16
Maps, estimates and sources
A value stream map and a standard work sheet have one thing in common: both look like measurements. Draw a box, put a number in it, and it reads as fact, whoever put the number there. An assistant will fill a missing number faster than you can notice it is missing. This lesson is about keeping a guess from turning into a data point on your map, and about deciding when an assistant should do a step and when it should help the person who does.
By the end of this lesson you can:
- say why an assistant's estimate reads like a measurement, and what that does to a map;
- mark every number on a map with where it came from, and leave a gap as a gap;
- turn a missing number into a threshold you can go and test;
- ask three questions that tell you whether to automate a step or to use the assistant to help the people who do it.
Last time: two questions to warm up (not graded)
1. Why isn't "are you sure?" a check on a number? It is the same model's second look, not an independent method. Check with code, a hand calculation or a total from the column sums.
2. A run time is blank in one row. What do you require of the assistant? That it not fill it, that it mark the row [NEEDS GEMBA], and that it compute only what the data allows.
The Lean job: map the flow and write the standard, from what you saw
The classic guide to value stream mapping, Learning to See by Rother and Shook, is built around walking the stream yourself and drawing what you find. The same is true of standard work: you watch the job, time it and ask the operator why. An assistant can help with the arithmetic and the paperwork. It cannot walk, watch or time anything, and when the page has a blank, its habit is to fill it.
Predict first
Here are the notes for a current-state map of a made-up packaging line. Demand is 700 units a day and the available time is 27,000 seconds a day. Op 20 has not been timed yet.
An assistant given these notes, with no other instructions, replies:
Takt time is 38.6 seconds. Seal (Op 20) cycles in about 42 seconds, so it is the bottleneck and the constraint on the line.
Before you read on, write down which parts of that reply you can trust, and which you can't.
What to trust, and what to do
Takt: 27,000 ÷ 700 = 38.57 seconds. That is arithmetic on figures you gave it, and you can check it.
The 42 seconds: it is not in your notes. The assistant filled a blank with a plausible number, and once written down it became the reason Seal is the bottleneck. Everything after it, the bottleneck and the constraint, is only as good as a number nobody measured.
What to do instead: mark Op 20 as missing and leave it out of the arithmetic. Then turn the gap into a question: "What would Op 20's cycle time have to be for it to miss takt?" The answer is anything over 38.57 seconds, about 38.6, and now you know exactly what to go and time. Asking again until it says 42 a second time is not measuring.
A guess reads like a measurement
NIST calls confident, unsupported output confabulation and warns that models can add invented justification for it. A 2025 preprint argues that training and scoring reward guessing over admitting uncertainty. Anthropic's guidance on reducing made-up answers says to allow "I don't know", to have the model quote what it relies on, and to retract claims it can't support. That helps. It does not remove the problem, so the map itself has to be built to show where each number came from.
What an assistant can do on a map, and what only you can
I found no controlled study of assistants drafting value stream maps, so this table is my reading of the evidence in Lessons 1, 4 and 5. One Lean Enterprise Institute author, Pereira, argues as opinion that the value stream should be mapped first.
Automate or augment
Once a process is mapped, the temptation is to hand a step to an assistant or an agent and be done. Three pieces of evidence argue for care:
- Toyota's own words. Toyota describes jidoka as automation with a human touch: the machine stops when it detects an abnormality, and a person responds. Toyota also says that machines, robots and IT cannot evolve alone, that only people do kaizen, and that craftsmanship comes from doing the work by hand and then applying it on the line.
- AI amplifies what is already there. A 2025 study of software teams found AI mainly magnified existing strengths and weaknesses. Lean Enterprise Institute authors warn about embedding inefficiency and about optimizing one part at the expense of the whole.
- Using AI changes the user. A Shingo Institute essay lists overreliance and deskilling among the risks. Lesson 1's studies point the same way.
Three questions before you automate a step
- Is it mapped, timed and stable? If it isn't, you would be automating waste or a guess.
- What stops it, and who responds, when something is abnormal? That is the jidoka question. A step with no stop and no responder is not ready.
- Who checks the output, and who decides what the freed time is for? Lesson 10 returns to the second half.
If any answer is "nobody", augment: use the assistant to help a person see, draft and question, and keep a person on the step. These three questions are my reading of jidoka and of the warnings above, not a Toyota rule. I could not find a primary Toyota source for the often-repeated line "never automate an unstable process", so I don't quote it.
Standard work for a job nobody has timed
An assistant will write a tidy standard work sheet for any job you describe, with steps, times and key points, and every line will read as if someone had watched. If nobody did, the sheet is fluent fiction. Fluent wrong guidance is documented: in 2024 New York City's own AI chatbot was reported to be telling business owners to do things that broke the law.
A standard is proven on the job, not on the page:
- Watch the job, with the operator.
- Correct the steps, the key points and the reasons with the operator.
- Have a newcomer try the sheet, and see where they stop.
- Only then post it, with an owner and a date.
That is my reasoning from the logic of Training Within Industry, not a finding about assistants. Lesson 7 shows how to turn a standard you have watched into a skill.
Your turn: leave the gap, then ask what it would take
About 25 minutes, on the notes from the predict box.
Your work, your data. The notes are invented. Use them as they are, and use only an invented or already public process for the thinking step at the end. When you map something real, apply Lesson 2 first.
- Takt by hand (2 minutes). 27,000 ÷ 700. Which steps could be a constraint?
- Run A (5 minutes). With code execution on, give the assistant the table, the demand and the available time, and send only:
Calculate takt time and compare each step's cycle time with it. Show the code.
Mark each number in its answer: from my notes, calculated or supplied by the assistant. Did it fill Op 20?
- Run B (5 minutes). In a new chat, add:
Use only the cycle times I gave you. Do not estimate any missing time: mark it [NEEDS GEMBA: time this step] and leave it out of the arithmetic. Tag every number [DATA], [USER] or [NEEDS GEMBA].
Compare the two answers.
- Gap to threshold (4 minutes). Ask:
What would Op 20's cycle time have to be for the line to miss takt?
- The go-time list (4 minutes). Who will time Op 20, over how many cycles, on which shift, and when? What would you ask about the other three numbers using the three questions?
- Automate or augment (5 minutes). Pick one step in an invented or public process, such as a restaurant order or a package delivery, and answer the three questions. Does it pass?
Numbers to check against
Takt is 38.6 seconds. The three timed steps are below it: Fill by 10.6 seconds, Label by 7.6 and Case pack by 12.6. Op 20 would have to be over 38.57 seconds to miss takt. The waiting inventory is 600 + 250 + 900 + 400 = 2,150 units, about 3.1 days of demand. Your assistant's results will vary from run to run, and a second run may differ from the first.
Knowledge check
Five questions, graded for you. Sign in to take the check, save your progress and count this lesson toward your certificate. Sign-in opens on Monday, October 12.
Sign in to take the check
One of the five is a set of map notes an assistant drafted from the same table. Tick the notes you wouldn't rely on as written. A calculator is fine.
Your task: mark up one map for what you know and what you don't
- Open the printable worksheet. List the steps of one process you know, with a number for each.
- Next to each number write where it came from, using M, O or A, and how many cycles or how long. Mark every gap [NEEDS GEMBA].
- Write the go-time list: who will time what, and when. Then go and do the first one before you open an assistant.
Open the printable worksheet
Sources
Where the facts in this lesson come from. Facts last checked October 1, 2026. Several are opinion or preprints, and the lesson says so where it matters. If you find a fact that is out of date or wrong, tell me.