Beyond Will AI Replace Jobs

"Will AI replace jobs" is a question about the future. "Which parts of this job should move to AI" is a question with an answer today.

Framing AI's effect on work as a single, binary outcome, replacement or no replacement, makes for an easier headline than it does a useful planning tool. Almost no real job is being replaced wholesale in one step. What is actually happening, inside role after role, is more granular: specific tasks within a job are shifting toward AI assistance or automation, while other tasks within that same job are staying resolutely human, and a third category of tasks is changing shape entirely, becoming about overseeing and checking AI output rather than producing the output directly.

This task-level view matters because it is the level at which decisions are actually made. A contact centre agent's job is not simply automated or not automated. Drafting a first-response message might move substantially to AI. Judging whether an unusual complaint needs escalation almost certainly should not. Deciding how to phrase a sensitive response to a vulnerable customer sits somewhere in between, assisted by AI but reviewed by a person.

Very few organisations have actually mapped their roles at this level of granularity. Most are instead making broad, top-down calls about which functions to "AI-enable," without a clear view of which specific tasks within those functions genuinely benefit from AI assistance and which ones need to stay firmly with people.

That gap, between talking about AI and jobs in the abstract and actually redesigning work at the task level, is where most of the practical difficulty around people, skills and work currently sits.

Leadership ambition for AI adoption is frequently well ahead of what the people expected to deliver on it have actually been equipped to do.

It has become common for senior leaders to set confident, sometimes ambitious targets for AI adoption across their organisation, while the practical, everyday understanding of what AI can and cannot reliably do remains thin across much of the workforce that is expected to deliver on those targets.

This gap shows up in two opposite but equally damaging ways. In some teams, AI is used uncritically: outputs are trusted and acted on without the scrutiny they need, because the people using the tool have not developed a working sense of where it tends to fail. In other teams, AI is barely used at all, not through active resistance but through a lack of confidence about when it is appropriate to rely on, which leaves genuine productivity and quality gains sitting unrealised.

Neither pattern is really about willingness. Both usually trace back to the same root cause: people have been given access to AI tools without being given a genuine, working understanding of how those tools behave, where they are strong, where they are unreliable, and what their own responsibility is when using them.

Closing this gap is not primarily a technology problem. It is a capability-building problem, and it requires leadership to treat AI literacy as seriously as it treats any other core organisational capability, rather than assuming it will develop informally as people simply use the tools more.

The upskilling that changes behaviour is role-specific, grounded in real examples of the work, and revisited as the tools and tasks evolve.

Effective AI upskilling starts from the actual tasks a specific role performs, not from a generic curriculum applied uniformly across the organisation. Training built around real examples drawn from the team's own work, including examples of AI getting something plausibly wrong, teaches judgement in a way that abstract demonstrations rarely manage.

It also needs to be layered rather than delivered as a single event. A useful structure separates foundational literacy, understanding broadly how the tools work and where their limits lie, from role-specific practice, applying that understanding to the particular tasks and judgement calls a given team faces, and from ongoing reinforcement, because both the tools and the tasks keep changing.

Peer learning tends to be underused relative to its value. People closest to a specific task often develop the sharpest intuition for where AI output can be trusted and where it needs scrutiny, well before that intuition is captured in any formal training material. Structured ways of surfacing and sharing that frontline knowledge, rather than relying entirely on top-down training design, accelerate the whole organisation's learning curve.

Measurement matters as much as delivery. Training that is assessed only by attendance or satisfaction tells an organisation almost nothing about whether judgement has actually improved. The more useful signal is whether people are catching AI errors before they cause harm, and whether they know when to escalate rather than proceed.

As AI takes on more of the production work, the value of a role increasingly sits in the oversight around it.

A consistent pattern is emerging across roles that have absorbed meaningful AI assistance: the skills that differentiate strong performers are shifting away from producing a first draft and toward evaluating one. Spotting a plausible-sounding but incorrect claim, noticing when a summary has quietly dropped an important qualification, recognising when a pattern-matched answer does not actually fit an unusual situation, these are becoming core professional skills rather than niche quality-control functions.

Escalation judgement matters just as much. Knowing when a situation has moved outside the boundaries where AI assistance is appropriate, and having both the confidence and the organisational permission to say so, is a skill that needs to be actively built and reinforced, not assumed to exist by default. Cultures where raising a concern about AI output is seen as slowing things down will quietly discourage exactly the behaviour that keeps AI-assisted work safe.

None of this is a lesser version of the skills these roles required before. Evaluating a claim rigorously, recognising a subtle inconsistency, and deciding when to intervene are, if anything, harder skills to develop and harder to assess than the production skills they are increasingly sitting alongside.

Organisations that treat oversight and escalation as first-class skills, worth naming, training and rewarding explicitly, tend to catch AI-related problems earlier and build workforce confidence faster than those that leave this capability to develop informally.

Giving someone an AI tool changes what they have access to. Redesigning their role changes what they actually do with it.

A common and understandable first step is to give a team access to an AI tool and let individuals decide how to use it within their existing role. This is a reasonable way to build early familiarity, but on its own it rarely produces a lasting change in how work gets done, because the role, its targets, its approval steps and its definition of good performance all stay exactly as they were.

Genuine role redesign asks a harder set of questions: if a meaningful share of the drafting work now happens faster, what should the role spend the freed-up time on instead? If quality checking becomes a larger part of the job, does the role's performance measures, training and seniority structure reflect that shift? If a task has moved from being done by a person to being reviewed by one, does the job description, the hiring profile and the career path for that role still make sense as written?

This work is slower and more disruptive than a tool rollout, because it touches job design, performance management and sometimes organisational structure rather than simply granting a licence. It is also where the durable productivity and quality gains tend to come from, because it changes the work itself rather than adding a faster way of doing the same work in the same shape.

Involving the people who actually do the work in this redesign, rather than designing it for them, consistently produces better outcomes. The people closest to a task usually have the clearest view of which parts genuinely benefit from AI assistance and which parts would suffer if handled the same way.

The organisations that get the most from AI will not be the ones that answer "will it replace jobs" first. They will be the ones that answer "how should this job change, starting now" first.

Work is already changing, task by task, in almost every role that has meaningful contact with AI tools. The organisations managing that change well are treating it as a deliberate exercise in role redesign, capability building and leadership behaviour, not as a side effect of rolling out a new tool.

That means mapping which tasks should move and which judgement must stay human, building AI literacy that is specific enough to actually change behaviour, naming oversight and escalation as core skills rather than assuming they will develop on their own, and giving both managers and senior leaders a real, visible role in how the change happens.

None of this is a one-off project with a fixed end date. As the tools continue to evolve, the roles, skills and leadership habits built around them will need to keep evolving too. Organisations that build that adaptive capability now will be far better placed for whatever the technology does next than those still waiting for a definitive answer to whether AI will replace jobs.