Reskilling for AI: Build Change Capability Before the Tools Arrive

This week’s article explores an increasingly contentious topic: many people using AI at work were never taught how to. The Conference Board’s Skilling for AI study, a global survey of nearly 1,300 workers supported by interviews with 35 enterprise leaders, found that 55% of workers use generative AI or AI agents regularly, while only 33% have taken part in employer-provided AI training. Nearly a third, 28%, say that their employer provides no AI training at all.¹ The capability that sits inside most organisations today was assembled privately, from public tools and personal experiment.

That is a more uncomfortable position than it first appears. An organisation that has not trained its people has not thereby avoided the risk of them using AI. It has simply lost sight of how they are using it, on what, and with what standard of care. The work is being done differently already. What is missing is any shared view of how it should be done.

The scale of what that capability now has to become is already clear. IBM’s 2026 CEO Study reports that between 2026 and 2028, executives expect 53% of employees to need upskilling to perform their current role effectively, and a further 29% to need reskilling for a different role altogether.² On their own leaders’ estimate, then, more than four fifths of the workforce will need a deliberate change in capability within two years. Very few organisations have a plan of that size.

What organisations are teaching, and what they are not

Where training does exist, it tends to stop at the foundations. The Conference Board found that organisations concentrate on AI literacy and basic prompting techniques, with far fewer helping workers to develop advanced capabilities such as managing AI agents, integrating AI into workflows, or applying AI to strategic business challenges.¹

The same pattern appears in what people are choosing for themselves. Revelio Labs reported that 31% of all newly recorded certifications on professional profiles in June 2026 were AI certifications, and that 47% of those related to generative AI and large language models.³ Individuals are clearly investing in their own foundational fluency, and at considerable scale, which is welcome. It does, though, produce a workforce that can prompt a model but cannot yet supervise one.

That distinction matters more as agentic systems arrive. Preparing an operating model for agentic AI depends on people who can oversee, question and override a system that acts on its own. Prompting is a skill for using a tool. Supervision is a skill for governing a process, and it is the one that the training is not yet reaching.

There is a practical reason why foundational training is what gets built first. It is easy to procure, it can be delivered to everybody at once, and completion is simple to report. Supervision is none of those things. It has to be taught against the actual work, it differs by function, and it is only really demonstrated when someone declines to accept an output that looked convincing. Training that is easy to measure tends to crowd out training that matters, and the reporting looks healthy either way.

Who the capability is actually for

There is a second question that most programmes answer by default rather than by decision, which is who receives the training at all. The Bipartisan Policy Center, drawing on Lightcast job-postings data to June 2026, found that three-quarters of all AI skill demand remains concentrated in just three occupational groups: computing and mathematics, business and finance, and management.⁴ Demand is also growing quickly outside technology itself, with management consulting postings up 41% over the year, so this is not a pattern that stays inside the IT function.

The temptation is to read that concentration as a map and to train accordingly, building deep capability in a technical group and leaving everyone else with an introductory session. The peer-reviewed evidence points the other way. Research published in Management Science finds that AI and algorithms act as complements to domain expertise, and that they create the most value when algorithmic literacy is diffused broadly among the people who already hold the domain knowledge, rather than concentrated in specialists. The same work finds that markets reward firms’ AI investments more highly when those capabilities are widely dispersed across the workforce.⁵

That is a strategically useful finding, because it reframes what the training is for. The person best placed to notice that an AI output is subtly wrong is the person who knows the underlying work, not the person who knows the model. An organisation that concentrates its capability in a technical group has optimised for building things. An organisation that spreads it among its domain experts has optimised for catching things, which is the harder and more valuable of the two.

Why the training that exists does not land

Provision is only part of the problem. The same study found that only 48.0% of workers agree that their organisation gives them sufficient time during working hours to build AI capability, and that only 47.6% agree they have the tools, access and resources to do so.¹ Training that is offered without the time to take it is a budget line rather than a capability.

This is where reskilling stops being a learning question and becomes an operating-model one. Time to learn has to be taken from somewhere, which means that someone’s delivery expectations have to move, and that is a decision only a line leader can make. Where that decision is left implicit, people quite reasonably conclude that the training is optional and the deadline is not. The programme is then judged to have failed on engagement, when what actually failed was the capacity planning around it.

Ownership is also unsettled. SHRM’s State of AI in HR 2026, a survey of 1,722 HR professionals, found that missing skills were named as a primary barrier to expanding AI, while only 15% of respondents thought that HR should lead on training employees to use AI tools.⁶ The function that sits closest to workforce capability is not widely regarded as the one accountable for building it. When a capability gap is everyone’s concern and no one’s deliverable, it persists.

Upskilling is not reskilling

The Conference Board’s central finding is that training investment remains heavily concentrated on upskilling people within their current roles, and that most organisations are not yet preparing for large-scale reskilling.¹ The two are quite different undertakings. Upskilling assumes that the role survives, and that the person needs more capability within it. Reskilling assumes that the role does not survive in its present form, and that the person needs a route to a different one.

IBM’s figures place roughly three in ten employees in that second category by 2028.² Reskilling on that scale takes longer than upskilling, it requires redeployment paths that are designed well in advance, and it depends on people believing that the destination is real. An organisation that begins it in the quarter that the roles change has begun too late.

The credibility point deserves particular attention, because it is the one most often underestimated. Reskilling asks a person to invest months of effort on the strength of a promise about a role that does not yet exist. If the organisation has previously announced a redeployment programme and then made the same people redundant, that promise carries no weight, and no amount of course design will fix it. Reskilling runs on institutional trust, which is built well before it is needed and cannot be assembled on demand.

The South African position

The local starting point sits below the global one. A survey of 426 Western Cape businesses, published in Development Southern Africa, found that more than 70% were struggling to find workers with the right skills. At least 55% judged their workforce to have only basic digital skills, 70% rated their workforce’s technology skills as poor or very poor, and 52% said that they found it difficult to recruit people who were prepared for digital transformation.⁷

The same study found that most firms were spending only 1% to 4% of their wage bill on staff training.⁷ A thinner skills base and a thinner training investment together mean that South African organisations have further to travel, and less momentum with which to travel it. That is an argument for starting earlier rather than waiting until the tools force the question.

It also argues for building capability internally rather than expecting to buy it. If more than two thirds of local employers already cannot recruit the skills they need, then a reskilling strategy that depends on hiring people who are ready is competing for a supply that demonstrably is not there. The organisations that will do well here are the ones that treat their existing workforce, with its domain knowledge and its institutional memory, as the asset to develop.

What the capability actually has to cover

Fluency with a model is the entry requirement rather than the objective. Capability for an AI-enabled operating model has to reach a good deal further:

• Supervision. Reviewing what a system produces, recognising when it is wrong, and holding the authority to act on that judgement. The authority matters as much as the recognition, because a person who spots an error and does not believe they may act on it has not been given a capability at all.
• Verification. Knowing how to check an output against something other than its own confidence. Fluent, well-structured and wrong is the failure mode that these systems produce most readily, and it is the one that reads as competent.
• Escalation. Recognising the boundary of one’s own competence, and the point at which a decision belongs to someone else. This has to be defined by the organisation rather than left to individual discretion, or it will be applied inconsistently and usually too late.
• Domain judgement. Knowing the work well enough to notice when a plausible answer is a wrong one. This is a capability that deepens in value as automation spreads, and the Management Science evidence suggests that it is precisely where algorithmic literacy pays off best.⁵
• Learning capacity. The tools will change again. Capability that is built around today’s interface will need rebuilding, whereas the ability to absorb a new way of working will hold.

Where to start

Organisations that are further ahead on this tend to have done four things in a recognisable order. They have established honestly where capability actually sits today, rather than assuming it from job titles. They have decided which roles are being upskilled and which are being reskilled, and said so. They have allocated the working time that the learning requires, and adjusted delivery expectations to match. And they have named an owner who is accountable for the capability outcome rather than for course completions.

None of those four is a procurement decision, and none of them can be delegated to a training provider. They are governance and operating-model choices, which is why capability programmes so often stall in organisations that treat them as a learning-and-development matter.

Building it before the tools arrive

This is the work that CCG does, and the sequence of it matters as much as the volume:

• Build the judgement, not only the fluency. CCG Learn develops the confidence and capability that people need in order to supervise, question and override these systems, which is precisely the capability that the survey evidence shows to be missing.
• Establish where you actually stand. CCG Analytics measures readiness, adoption and sentiment, so that capability decisions rest on evidence rather than on assumptions about who can already do what.
• Design the roles that the capability serves. CCG Consult translates an AI ambition into changed processes, escalation paths and defined human roles, so that the training has a destination.
• Settle the governance. CCG Advisory works with Boards and executives on the oversight and accountability model, which determines what people must be capable of deciding.

Reskilling is the half of an AI programme that cannot be bought in the quarter that it is needed. Tools arrive on a procurement timetable and capability arrives on a human one, and the second of those is a good deal slower. Organisations that build it before deployment will have people who are ready to supervise the systems they are given. Those that begin afterwards will be training their people to catch up with decisions that the technology is already making.

If you are building change capability ahead of an AI programme, we would welcome the conversation. Get in touch with CCG.