AI at the Gemba · Lesson 10 of 10 · Week of December 14

Roll it out with respect for people

Everything so far has been about the tool: what it is, what to paste, how to brief it, how to check it, how to guard it. The last lesson is about the people. A Lean team won't use something it fears, and it will hide what it does use. Rolling AI out well means deciding, before anything starts, what the time it frees is for, who is accountable, what you will and won't measure, and how the team will see the results before management does.

By the end of this lesson you can:

say what shadow AI is, and why a ban doesn't make it go away;
write a freed-time commitment in which every line is one you could keep;
name who owns the rule, the tools, the skills library and each output;
run an introduction as a 30-day PDCA pilot with stop criteria;
keep analysis from turning into surveillance, and say where the law starts to apply to decisions about people.

Last time: two questions to warm up (not graded)

1. Which guardrails depend on the model's cooperation, and which don't? Instructions, tags, "no invented numbers" and stop-and-ask depend on it. Scope and permissions, human review and logging don't.

2. What is the Rule of Two? In one session allow at most two of: reading untrusted content, touching private data or systems, and changing things or communicating outside.

The Lean job: introduce a change without losing trust

Toyota's guiding principles call for mutual trust and respect between labor and management. In Lean, respect for people isn't a slogan on a wall. It decides whether a change lasts, because people tell you the truth about a change only when they aren't afraid of it. In my reasoning, a pilot that people fear reports only good news, and the Check step of PDCA goes blind.

I'm not a lawyer, and some of this lesson touches employment law, which differs by place and changes quickly. Where it matters I say so, and I give a date.

Predict first

You find out that several operators have been drafting their shift handover notes in personal AI accounts, on their own phones. Nobody told them they couldn't, and nobody told them they could. Before you read on, write down your first move, and what you think happens if you simply ban it.

What the evidence points to

Ask what job it does for them. Offer an approved tool and a never-paste list, and don't blame anyone. A breach of policy and concealed use are routine, which the next section shows with numbers. A ban is a stopgap: after three reported incidents at Samsung in April 2023, including source code pasted into a public chatbot, the company banned such tools until it had safer ones.

People already use it, and they are afraid of it

Shadow AI

Shadow AI is staff using tools the company never approved, usually personal accounts. As of May 2026, Gallup found that 52% of US employees use AI at work at least a few times a year, yet only 25% say their employer has communicated a clear plan (22,573 people). Globally, in a 2024 and 2025 survey by KPMG and the University of Melbourne, 70% of employees who use AI relied on free public tools, against 42% on tools their employer provided. 44% said they had broken policy, and 61% avoided revealing their use. Production workers stayed at 9% to 11% frequent use over two years, against 27% for white-collar workers. These are self-reports, and the headline depends on the question: Pew found 21% in 2025 for "some of their work done with AI".

Fear makes people hide it

Pew, in October 2024 (5,273 US workers): 52% were worried about the future of AI at work, and 32% expected fewer job opportunities for themselves.
Gallup: 15% of US employees expect AI to eliminate their job within five years, flat since 2023.
A Microsoft and LinkedIn survey, which is a vendor's own claim, found 52% of AI users reluctant to admit they use it for important work.

On surveillance, a 2022 Pew survey (11,004 US adults) found 71% opposed to AI making final hiring decisions and 66% saying they would not apply for a job where it was used. People are watching how AI gets used on them.

The supervisor is the channel

Only 36% of employees strongly agree that their manager supports AI use, and those who do are 1.7 times as likely to use it often. I found no randomized evidence on which AI-literacy training works. In BCG's write-up of its consultant experiment, a short onboarding overview did not protect people once a task was beyond what the tool could do. The EU's AI Act has an AI-literacy duty for staff that stands. Lean Enterprise Institute authors note that people resist when AI removes even tedious tasks. My reasoning from this: train by role, on where the tool fails, and repeat it at the huddle.

Decide what freed time is for, before you start

Time an assistant frees up doesn't assign itself. If you haven't said what it is for, someone will, and people know it.

Toyota in the 2008 and 2009 recession, according to IndustryWeek, kept its regular US team members, used idle time for training and kaizen, and had plant executives give up bonuses.
Womack, in a Lean Enterprise Institute post by Shook, says to stop making employees the shock absorbers for the business.
Byrne, interviewed on LeanBlog in 2012 about the CEO's role in a Lean transformation, argues kaizen is hard without a promise of no layoffs. Graban, writing on LeanBlog, notes that some hospital leaders have promised careers, not specific jobs.
Lean Enterprise Institute (2025) describes the choice AI offers as extraction or amplification.
The tally: Challenger, Gray and Christmas counted 120,136 AI-attributed job cuts announced from January to September 2026, about 21% of announced cuts. Those are stated reasons, not proof of cause.

A pledge you can't keep is worse than a narrower one you can. That is my reasoning, since broken promises cost trust. So write what freed time is for, and be exact about what you can and can't promise.

A freed-time commitment to adapt

Before the pilot, we write down what freed time is for: problem solving, training, the standard-work backlog.
No one loses a job or hours because of time AI freed during [the period]. We state any exceptions honestly, such as a closure or a collapse in demand. If we can't promise this, we say what we can.
We log the hours freed and what they were used for, and show them at each review.
Checking AI's work counts as work.
Gains are shared as [specify].

This is a template, not a legal document, and it isn't a promise from me to anyone. Whether your company can sign it is a decision for the people who can keep it.

Someone is accountable

Boston's guidelines for city staff say people remain responsible for the outcomes of what they use AI for: fact-check, disclose, don't paste sensitive data, and keep an owner accountable. In 2024 a Canadian tribunal held Air Canada responsible for wrong refund advice its chatbot gave a customer. A review of 41 human-oversight policies found that people do not reliably catch the errors of an algorithm they are supposed to oversee. In the KPMG survey, 66% of AI users had used output without evaluating it, and 56% said they had made mistakes.

So "a human in the loop" is weaker than it sounds. Name the people:

Rule owner: a supervisor or the continuous-improvement lead, who writes and keeps the one-page rule.
Tool owner: who holds the accounts and reads the plan's terms on retention.
Skills-library owner: who keeps the registry from Lesson 9.
An accountable reviewer for each output. Say what they check, give them the time, and audit a sample.

Measure honestly

Gains are real but uneven, and people's reports mislead. A study of 5,179 support agents found 14% more issues resolved an hour, 34% for novices and little for experts. In one small trial of 16 experienced developers, they took 19% longer yet believed they were 20% faster. Time a sample of the task before and after, and count the rework. Don't take "it saves me half an hour a day" as a result.

The one-page rule

A rule people can read in a minute is worth more than a policy no one opens. Each line has a purpose:

Rule line	What it is for
Approved tools: [list]. Anything else, ask first.	A legal path shrinks shadow use.
Never paste: confidential customer or supplier data, personal data, names in safety incidents, passwords.	Prevents leaks (Lesson 2).
You check every output before it leaves the team.	Answers the two-thirds who use it unchecked.
Your name is on it.	Accountability.
Say so when AI helped.	Answers the concealment: 61% avoid revealing their use.
No AI alone on hiring, promotion, discipline or pay.	Legal exposure (next section).
We don't use AI to rank individuals, and we tell you what data we analyze.	Trust, and notice laws.
Report an error or a near miss, with no blame.	The stop-the-line habit.
Review date, and a stop rule: [dates].	PDCA.

You don't have to start from nothing. There are published models: Boston's guidelines, Partnership on AI's seventeen practices for employers (including worker agency in choosing tools, data transparency, opt-out, human recourse, crediting the extra work AI creates and sharing productivity gains), a sample acceptable-use policy from SANS, a UK government self-assessment for small firms, and NIST's voluntary risk-management framework. The AFL-CIO and Microsoft announced a labor-management partnership in December 2023 as an example of bringing workers' voices in.

Analysis can become surveillance. Per-operator scan-rate exports, usage rankings and "productivity scores" turn a tool for seeing the process into one for watching people. Look at the process first, say what data you use, and let a person who knows the operator decide any follow-up.

Where the law starts: decisions about people

As of October 1, 2026. Not legal advice. Check by January 2027, and ask counsel. This area is moving fast. Where I couldn't read the statute itself, my source was a law firm's summary.

An assistant that drafts an A3 is one thing. An assistant that helps decide who is hired, promoted, disciplined or paid is another, and several laws apply to that second case. A map of what I found:

New York City, Local Law 144: automated hiring tools need a bias audit within the past year, a public summary, and notice to candidates ten business days ahead. The state Comptroller judged enforcement ineffective in December 2025.
Colorado: a 2026 act, SB 26-189, signed May 14, 2026, replaces the 2024 law that a federal court had paused. It takes effect January 1, 2027, with notice, an explanation of adverse outcomes, human review and three-year records.
Illinois: from January 1, 2026, an employer may not use AI that has a discriminatory effect in hiring, promotion or discipline, and must give notice. Rules for the notices were reportedly postponed in June 2026.
California: from October 1, 2025, employers answer for the automated tools their agents use, and keep records for four years.
European Union: the employment-related high-risk duties are delayed to December 2, 2027. The AI-literacy duty stands, and workers' representatives are to be informed first.
US federal: the EEOC's AI guidance pages were gone when I looked, but Title VII and the Americans with Disabilities Act still apply to AI-assisted decisions. A December 2025 executive order targets state AI laws without actually overriding them.
Unions: in a union workplace, the duty to bargain over working conditions can apply before a change. Whether an AI tool changes working conditions is a question for counsel, as is supervisors' status.
Monitoring: New York requires written notice of electronic monitoring of employees.
Records: hiring and discharge records must be kept for a year, and longer after a charge is filed. Losing electronic evidence can be sanctioned. It is my reasoning that work chats about a decision could count. Vendors' retention varies by plan.

The practical rule is the one already in the one-page rule: no AI alone on hiring, promotion, discipline or pay, and ask counsel before any use that touches them or monitors people.

Run it as a 30-day PDCA pilot

Plan. One volunteer team, one Lean job from Lessons 1 to 7, one skill. Name the roles above. Use a business-plan tool, not personal accounts, and read its retention terms. Write the rule and the freed-time commitment. Brief the union representative and HR first, and ask counsel before any use that touches hiring, discipline, pay or monitoring. Set the stop criteria and the review dates now.
Baseline. Time a sample of the task, count rework, and log current AI use without blame.
Do for 30 days. Voluntary, and no per-person usage rankings.
Check at days 10, 20 and 30 in the huddle: measured time, rework, weekly users, a short pulse survey, hours freed and how they were used, and any breaches.
Act. Adopt, adjust or stop, and show the team before management.

With no IT or legal team, the minimum is: a one-page rule, one business-plan account, one named owner, no AI alone on decisions about people, and a dated 30-day review on file.

A worked example, all invented

A 140-person plant. The Line 3 team has one supervisor, eight leads and the continuous-improvement lead. The use is first drafts of A3 background sections and summaries of gemba notes, with no names and no customer data. On day 0, leadership reads the commitment aloud with the union steward present.

Measures: minutes to a first draft (six A3s before, six after); factual corrections per draft; weekly voluntary users; three pulse items scored 1 to 5, "I know the rule", "I can say no to using it" and "I believe freed time will be used as promised"; hours freed and what they were used for; breaches.
Day 10: tune the rule. Day 20: audit five outputs. Day 30: adopt, adjust or stop, and tell the team first.
Stop on a breach of sensitive data (pause for 48 hours), on an error that reaches a customer, or if the third pulse item drops by a full point.

Your turn: write the rule, then try to break it

About 25 minutes, on an invented pilot.

Your work, your data. The pilot is invented: "Line B changeovers run long." Use no names, no customers and no real numbers, and paste into an assistant only text you wrote yourself, under Lesson 2's rule.

Invent the pilot (5 minutes). One team, one job, one skill, and what you would measure.
Draft the one-page rule (10 minutes) from the table in this lesson. Then paste only the rule you wrote into an approved assistant and send:
```
Act as a skeptical operator. List five questions this rule fails to answer.
```
Answer the ones that matter, and change the rule.
Rewrite the freed-time commitment (5 minutes) so that every line is one you could keep. Cross out any you couldn't, and say what you can promise instead.
Write the stop criteria (5 minutes). Three of them, with the day-10, day-20 and day-30 review dates. Test: can a new hire read the whole page in 90 seconds?

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.

When you have passed all ten checks, the final test opens. It has fifteen questions across the whole course.

Your task: put the rule and the commitment on one page and hand it to the team

Open the printable worksheet. Fill in the roles, the one-page rule, the freed-time commitment and the stop criteria for a real pilot.
Read it to the team at a huddle, with whoever represents them in the room, before anything starts. Write down the questions they ask.
Put the day-10, day-20 and day-30 reviews in the calendar. Then take the final test.

Open the printable worksheet

In the toolkit

Daily Management System builds a tier-board structure with KPI selection, escalation logic and a response-to-abnormality protocol. If your 30-day pilot needs a regular check-in and a way to raise a problem, start there.

01Tier boards that nobody looks atDaily ManagementA tiered daily management design with escalationRead the skill →

Sources

Where the facts in this lesson come from. Facts last checked October 1, 2026. Survey figures, laws and vendor terms change quickly, and some sources are law-firm summaries or trade press, which the lesson says where it matters. If you find a fact that is out of date or wrong, tell me.

09Guardrails, and testing what you wrotePrevious lessonLesson 9 → ··All ten lessons, the certificate and how it worksBack to the courseCourse home → ★Fifteen questions. Pass with 12 of 15.The final testFinal test →