When AI takes on part of a task, the work around it changes: who checks, who handles exceptions, what skills matter, and what the day looks like. If people hear about it after the decision, fear fills the gap. Involving the people who do the work, from the start, is the same Lean answer as for any change.
This guide shows how tasks and roles shift, with a worked example of one analyst’s week; sets out six steps for involving people; describes new roles and the skills they need; covers training, trust and over-reliance, and skill loss; and ends with cautions on workplace monitoring and decisions about people.
Employment and monitoring rules vary by place and are changing. This is not legal or HR advice. Example data are illustrative.
Before You Start
AI Changes Jobs Before It Changes Costs
When AI takes on part of a task, the work around it changes: who checks, who handles the exceptions, what skills matter, and what the day looks like. If people hear about it after the decision, they fill the gap with fear, and the people whose judgment the system needs most are the ones who stop trusting it. The Lean answer is the same as for any change: involve the people who do the work, in the design, from the start. See the Change Management for Improvement Guide.
Start With the Task, Not the Job
AI takes over tasks, rarely whole jobs. Describe what changes task by task, with the people who do them.
The Checking Role Is Real Work
Reviewing AI output takes skill and attention. Design it, staff it, and count the time.
Be Honest About Intent
Say what the use is for, what it is not for, and what will happen to the time it frees.
Build Skills Early
People need to know what AI can and cannot do, and how to check it, before they are asked to rely on it.
What Changes in a Role
Riverside Plant, an illustrative 140-employee metal fabrication and assembly plant, put an AI classifier on incoming complaints. The picture shows the plan for one complaint analyst. Reading and coding falls, because the AI does the first pass. New work appears: checking the AI’s work and handling the exceptions it sends up. Time moves to investigating causes and to improvement work, which is where analysts add the most value. The plant promised in writing that the freed time would go to those two things. The numbers are illustrative.
Involve People From the Start
- Say what you are considering and why, before a tool is bought. Include the limits: what the AI will not be used for.
- Bring the people who do the work into the design. They know the awkward cases, and they will be the checkers.
- Map the task changes together, as in the figure, and agree what the new role looks like.
- Run the pilot in shadow mode first, so people can see the tool’s mistakes without relying on it.
- Ask for feedback and act on it. Show what changed because of what people said.
- Review at 30, 90, and 180 days: workload, quality, confidence, and what people need.
New and Changed Roles
| Role | What the person does | Skills to build |
|---|---|---|
| Use owner | Accountable for results of one AI use; reads the monitoring; decides to change, pause, or retire | Process knowledge; statistics basics; risk judgment |
| Reviewer | Checks a sample or all outputs; records errors; flags patterns | Domain expertise; checking habits; resistance to automation bias |
| Exception handler | Handles cases the AI sends up or cannot do | Deeper expertise; judgment; communication |
| Prompt and data steward | Keeps prompts, golden sets, and data definitions current | Clear writing; data discipline; testing |
| AI champion | Helps peers, collects questions, and passes on what works | Teaching; curiosity; credibility with peers |
| Governance lead | Keeps the register, policy, and reviews | Risk management; audit; governance |
Skills and Training
Plan training by role, using the Skills and Training Matrix and the AI Workforce Impact and Training Plan. A short list of what everyone who uses AI at work should know:
- What the tool does and does not do, in plain words, and where it is likely to be wrong.
- Which data may and may not be entered (see the Data Protection and Privacy Guide).
- How to check output against the source and when a check is mandatory.
- How to spot instructions hidden in documents and messages, and what to do.
- How to report a problem, with no blame.
- Who to ask.
Test whether training worked, as in the Competency and Training Effectiveness Guide, and reinforce it as in the Training Reinforcement Guide. Some laws now expect organizations to ensure staff have enough AI literacy; for example, the EU AI Act includes such a duty. Check what applies to you.
Trust, Over-Reliance, and Skill Loss
Two opposite problems appear. Some people distrust the tool and ignore good output. Others trust it too much and stop checking, which is automation bias. Both are managed the same way: show the tool’s measured error rate, show real examples of its mistakes, test reviewers with seeded errors, and keep the tool’s results visible. A third problem is skill loss: when the AI does the routine part, people lose practice at it, and the exception handler who never does the routine work may not learn the basics. Bainbridge described this in 1983 as one of the ironies of automation. Keep some practice in the work, and rotate.
Handle With Care: Monitoring and Decisions About People
- Do not use AI to monitor, score, or rank workers without a clear purpose, a policy, and legal and HR review. Rules on workplace monitoring and automated decisions vary by place and are changing.
- In places with works councils, unions, or consultation duties, involve them early. Check the agreements and laws that apply.
- Keep a person accountable for any decision about hiring, pay, promotion, discipline, or termination, and treat these as high-risk uses (see AI by Function and Industry).
- Make clear how data about work and workers is used, stored, and shared.
Common Mistakes
Announcing, Not Involving
People learn of the change when the tool arrives.
Counting Only the Time Saved
The checking and exception work is left out, and people are overloaded.
Vague Promises About Jobs
Or promises the organization cannot keep.
Training Once
Skills fade and the tool changes.
No Way to Disagree
Overriding the AI is harder than accepting it.
Surprise Monitoring
Using AI to watch people, undisclosed.
Self-Assessment Questions
- Have the people who do the work shaped the design?
- Have we described the task changes and the new roles, and counted the checking time?
- Do we say plainly what the AI is not for, and what happens to freed time?
- Is training matched to roles and tested?
- Do we measure confidence and workload at 30, 90, and 180 days?
Workforce and Change for AI: Frequently Asked Questions
Will AI take my team’s jobs?
AI usually takes over tasks, not whole jobs, and it adds new tasks such as checking and handling exceptions. What happens to the time it frees is a management decision, so say plainly what the intent is and keep the promise. Do not promise what you cannot deliver.
How do we get people to trust the AI tool?
Do not ask for trust; earn it with evidence. Show the measured error rate and real examples of mistakes, let people see the tool in shadow mode, and make it easy to override. Then watch for the opposite problem, over-reliance.
What training do people need?
Training by role. Everyone who uses AI needs the basics: what it does and gets wrong, what data may be entered, how to check output, how to spot injected instructions, and how to report problems. Reviewers and owners need more.
Can we use AI to monitor employees?
Be very careful. Rules on monitoring and automated decisions about people vary by place and are changing, and trust can be lost quickly. Require a clear purpose, a policy, disclosure, and legal and HR review.
What is automation bias?
The tendency to accept an automated suggestion without checking it, even when it is wrong. It is reduced by showing error rates, testing reviewers with seeded errors, and designing the review so that checking is easy.
Sources and Further Reading
- Bainbridge, L., “Ironies of automation,” Automatica 19 (1983), on skill loss and the operator’s role when automation does the routine work.
- Parasuraman, R. and Manzey, D. H., “Complacency and bias in human use of automation,” Human Factors 52 (2010).
- Kotter, J. P., Leading Change, Harvard Business School Press, 1996, and Hiatt, J., ADKAR, Prosci, 2006, as examples of change frameworks.
- Regulation (EU) 2024/1689 (the EU AI Act), Article 4, on AI literacy (check the current status and timelines, which are subject to amendment).
- National Institute of Standards and Technology, AI RMF 1.0 (NIST AI 100-1, 2023), on human oversight and the people involved in AI systems.