The head of learning gets a clear request from the executive team: build an AI training programme for the whole organisation, and start before the summer. The first suggestions are predictable. A prompt writing course, an introduction to generative AI, perhaps an online module on responsible use. All useful, and all easy to buy. But when she asks the business which AI skills people will actually need in their new way of working, nobody can give a concrete answer.
That gap is understandable. Most organisations know which AI tools they are introducing, but not how those tools change the daily work of the people who use them. The answer is less technical than many training plans assume.
The skills question is bigger than most plans assume
The World Economic Forum expects nearly 40 percent of the skills required on the job to change by 2030 [1]. In its illustration of a typical group of 100 workers, 59 would need reskilling or upskilling in that period [1]. Those are broad labour market figures, but they match what organisations see when they introduce AI in specific processes: almost every role that works with AI output changes in some way.
Managers are an important part of the picture. Gartner research found that only 8 percent of HR leaders believe their managers have the skills to use AI effectively, and in a survey of 114 HR leaders only 14 percent said their organisation supports managers in integrating generative AI into daily work [2]. If managers cannot judge good use of AI, they also cannot coach their teams on it.
What the Library says people will need
Each use case in the MetroMinds Library describes how work changes for the people involved, and 4,680 of the 4,980 records, or 94 percent, name the skills that the new way of working requires. Across those records there are 192 different skills, with an average of almost four per use case. When you group them into themes, a clear pattern appears.
The two most common themes are checking evidence and sources, and applying policy, thresholds and controls. Each appears in 56 percent of the use cases that name skills. Reading data and signals follows at 40 percent, then judging exceptions and escalation at 35 percent, and recording decisions and overrides at 32 percent. Working with prompts and AI tools comes last, at 19 percent. Prompt skills in the narrow sense appear in just 15 percent.
In other words, the people in these use cases mainly need to judge what the AI gives them. They check whether a source is reliable, decide whether a case falls inside the policy or should go to someone with more authority, and record why they accepted or overrode a suggestion. Those are professional skills that sit close to the job itself, and they are hard to teach in a generic AI course.
The skill mix differs per part of the business
A single organisation-wide programme also misses how much the mix varies. The heatmap below shows, per , how often each skill theme appears.
Some differences stand out. In risk, compliance and internal audit, half of the use cases ask people to record decisions and overrides, which reflects how much evidence those functions must keep. In legal and corporate affairs, 58 percent ask for judgement on exceptions and escalation. In sustainability, ESG and health, half of the use cases require people to interpret scores and scenarios, because much of that work runs on models and forecasts. Technology, data and AI is the area where working with prompts and AI tools matters most, at 35 percent.
For a learning team, this means the content of AI training should follow the work. A compliance analyst and a field engineer may use the same type of AI, but they need different habits to use it well.
Building the skills before go-live
The most practical moment to build these skills is while the use case is being prepared, not after it is live. At that point you know which roles change, which decisions remain human and which mistakes the AI is likely to make. That is enough to design training around real cases from the team's own work.
A closer look at one role
The use case below shows how this works for a technical role outside HR's usual field of view. A maintenance engineer at a telecom operator does not need to understand how the prediction model works. The engineer needs to know when to trust a reading, when to doubt it and how to record what happened.
Questions to ask before you design AI training
- Which use cases in our portfolio go live in the next six months, and which roles do they change?
- For each role, which skills does the use case description name, and which of those does the team already have?
- Which decisions stay with people, and do employees know where their authority ends and escalation begins?
- Can we train on real documents and cases from the team's own work instead of generic examples?
- How will managers learn to judge good use of AI, so they can coach their teams after go-live?
Skills are only half of the people question. The other half is who carries responsibility when AI prepares a decision, which we discuss in who is accountable when AI prepares the decision and a person signs it.
How MetroMinds helps you plan AI skills
Every use case in MetroMinds describes the workforce impact, including the tasks that decline, the new responsibilities and the skills the new way of working requires. When you build a portfolio, the workforce impact view groups those changes per , so you can see which teams need which skills, and when, based on the use cases you have accepted.
That turns AI training from a generic programme into a plan that follows the portfolio. You can start with the teams whose use cases are closest to , and design training around the skills those use cases actually name.
Do you know how ready your people are for the AI you are planning? Take the AI Readiness Snapshot
Sources
- World Economic Forum, "Future of Jobs Report 2025: 78 Million New Job Opportunities by 2030 but Urgent Upskilling Needed to Prepare Workforces", press release, 8 January 2025.
- Gartner, "Gartner Research Finds Only 8% of HR Leaders Believe Their Managers Have the Skills to Effectively Use AI", press release, 8 October 2025.
- MetroMinds Library, 4,980 AI use cases with benchmark ranges based on real-world implementations, August 2026.
