Which Jobs Will AI Change First - and What Should Companies Do About It?

Roles containing substantial document handling, information processing and routine analytical work are sensible places to examine early AI-driven change. For a CEO or HR Director, that means looking at the tasks inside administrative support and professional roles, rather than starting with a list of occupations expected to disappear.

There is no reliable universal timetable for which jobs change first in your company. The sequence depends on the work, available systems, economics and the decisions your organisation makes. A role can be highly exposed to AI capabilities while remaining difficult or inappropriate to automate as a whole.

The practical response is to assess where work may change, prepare people for a tested new process and keep staffing assumptions separate from technology demonstrations.

What the research tells us about exposed work

The ILO’s 2025 global index of generative AI exposure identifies clerical occupations as the most exposed group, with increasing exposure in some highly digitised professional and technical occupations. Its assessment combines task data, worker input, expert judgement and AI-assisted scoring. These are estimates of potential exposure, not observations of every workplace adopting the technology.

The OECD’s May 2026 exposure measure takes a broader capability-based approach, covering cognitive, social and physical domains. It finds AI capabilities closer to standardised information-processing and administrative work than to work requiring contextual judgement and responsibility. The method measures capability gaps; it does not establish which employers will automate or when.

These findings help decide where to investigate. They do not justify declaring all administrative staff replaceable, or assuming that senior professional roles are unaffected. Different studies define and measure exposure differently. A global result also cannot be applied directly to a Dubai workforce with its own industries, roles and operating conditions.

Separate four questions before making a workforce decision

Task exposure asks whether AI capabilities overlap with some of the activities in a role. That overlap may create an opportunity for assistance, automation or a different way of organising the work.

Technical feasibility asks whether a particular system can perform the actual task at the required standard. Real documents, unusual cases, languages and incomplete information matter more here than a polished demonstration.

Adoption asks whether the organisation can and should put that capability into use. It requires accessible data, integration, an economic case, suitable controls, the relevant regulatory assessment and people able to operate the revised process.

Employment outcomes concern what happens to demand for work, responsibilities and staffing after those changes. Removing an activity does not by itself tell you whether the organisation will reduce positions, absorb a backlog, improve service or take on additional work.

The ILO’s April 2026 review of exposure indicators makes this distinction explicit: such measures signal possible change and do not, on their own, predict employment outcomes. Its limitations include static descriptions of tasks and missing economic and adoption constraints. Treat an exposure score as a reason to ask better questions, not as a staffing instruction.

Build a role-analysis sheet with the people doing the work

Choose a role where a meaningful amount of work may be affected. Review it with someone performing the job, their manager and a person who understands the relevant systems. A job description is a starting point; ask about the exceptions, coordination and informal problem-solving it leaves out.

Use the worksheet below for each important task or task cluster. It is a proposed management tool, not a scientifically validated score or an employment decision formula.

Record unknowns plainly. “Estimated time” should not quietly become “verified saving” when the worksheet reaches the board. Capture how the estimate was made and what would need to be observed before using it in a budget.

A hypothetical operations role shows why the details matter

Consider an operations coordinator who gathers updates, prepares a weekly status report, checks inconsistencies and negotiates priorities when teams disagree. The reporting activity appears promising for AI assistance, but the role contains more than assembling text.

The worksheet might show that compiling the first draft is repetitive while resolving conflicting updates requires conversations with colleagues. It might also reveal that different teams use the same status label to mean different things. An automated summary would inherit that ambiguity unless the process is clarified.

A reasonable trial could test report preparation while retaining responsibility for verification and exception handling. The learning plan would cover checking sources, recognising unsupported interpretations and explaining unresolved issues to a manager.

After the trial, assess the whole role. Perhaps reporting takes less time but review takes more. Perhaps the coordinator can address delays earlier, or a backlog remains elsewhere in the process. These are different organisational outcomes, with different staffing implications. None should be inferred merely because a draft report was produced quickly.

The complementary guide to Digital Employees versus human employees helps decide how much authority to delegate to a system. Here, the question is how that delegation changes the work people are expected and equipped to do.

Turn the analysis into a specific learning plan

Avoid assigning the same generic AI course to everyone whose role appears exposed. The person preparing a report, the manager reviewing it and the specialist maintaining the integration need different practice and different evidence of competence.

The OECD’s July 2026 paper Skills in the AI age highlights a mix of foundational, digital and complementary skills, including critical thinking and collaboration. That supports a broader training discussion than learning to operate a particular interface.

For the hypothetical coordinator, I would set an observable exercise: compare a generated report with the underlying records, identify unsupported claims and prepare an explanation of unresolved discrepancies. For their manager, the exercise would be deciding whether the evidence supports action and identifying what requires further investigation.

Protect time for learning and review whether the person can perform the revised task. If the team cannot use the proposed process reliably, revisit the process, training or staffing support before treating the change as complete.

Revisit the organisation after testing the work

Look beyond isolated time savings. Ask whether demand is rising, whether service has improved, whether difficult cases are concentrated on fewer people and whether new review work has a clear owner. Check that junior employees still have opportunities to build the knowledge needed for more demanding responsibilities.

AI can change the need for particular roles; it would be misleading to promise otherwise. Equally, an exposure estimate cannot establish that a position is redundant. Employment decisions need the organisation’s actual evidence, appropriate HR processes and the relevant professional review.

My recommended next step is to bring one completed role-analysis sheet to a discussion involving operations, HR and the accountable executive. Agree a trial, the learning support and the evidence required before changing workforce assumptions. That gives a leadership workshop a real decision to examine, with the people affected included in the process.