What White Collar Jobs Will AI Replace?

Short answer: AI is not replacing whole white collar professions, it is removing specific task clusters, and the roles built almost entirely from those clusters are the ones already shrinking in official projections. The pattern in the newest US Bureau of Labor Statistics data is consistent. Occupations whose core work is structured document handling, routine assessment against rules, or repetitive transaction processing are projected to decline through 2035. Occupations that combine judgement, accountability, physical presence or complex negotiation are projected to grow, several of them faster than the economy overall. This page names the specific occupations on each side, explains why the line falls where it does, and gives you a way to test your own position.

The Data That Actually Answers This

BLS published its 2025-35 projections in August 2026, showing total employment growing 3.5 percent over the decade, from 170.3 million to 176.2 million, which is well below the 10.9 percent recorded over 2015-25. Alongside the projections it introduced AI exposure categories that sort every detailed occupation into Low, Moderate, High or Very high relative exposure, built from five external datasets including observed usage measures derived from Claude and Copilot data.

One line in that publication should shape how you read everything else on this topic. BLS states that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Exposure measures task overlap. It does not predict employment. Most articles on this subject collapse the two, which is why they reach conclusions the underlying data does not support.

Which Office Roles Are Shrinking, and Which Are Growing

The roles already projected to decline

These are not predictions from a consultancy. They are the official 2025-35 projections for occupations that sit squarely in the office economy.

Look at what these share. Each is dominated by taking structured information in, applying an established set of rules, and producing a structured output. A claims adjuster assesses a documented event against policy terms. An underwriter scores an application against criteria. A secretary routes, schedules and formats. That is the exact shape of work current systems handle well.

Notice also the magnitude. The largest projected decline here is 6 percent over ten years. This is erosion, not collapse, and in absolute terms secretaries and administrative assistants remain a 3.4 million person occupation in 2035. The realistic experience is fewer openings and slower hiring rather than mass redundancy.

The roles projected to grow

The contrast is sharper than the decline list.

Software development growing 10 percent deserves attention, because it is the occupation most often named as the first casualty of code generation. The projection points the other way. BLS attributes broad demand in this area to the expansion of AI systems themselves, and the projections release notes that professional, scientific and technical services is the third fastest growing sector at 8.6 percent, adding 926,700 jobs.

There is a second reading of these two tables worth making. The declining roles cluster at the lower end of the pay distribution and the growing roles cluster at the upper end, with two exceptions that prove the point. Claims adjusters and underwriters both pay around $80,000 and are both declining, because pay in those roles reflected the cost of training a person to apply complex rules consistently, not the difficulty of the underlying decision. When applying rules consistently becomes cheap, the wage premium attached to it does not survive.

Accounting is the most instructive case. Bookkeeping and data entry inside accounting are heavily automatable, and the occupation is still projected to grow, because the parts that require a licensed person signing off have grown in volume and complexity.

What to Know Before You Draw Conclusions About Your Own Job

Task exposure is not job loss. An occupation can be highly exposed and still grow, if the exposed tasks were overhead rather than the value. Radiology and accounting have both absorbed heavy automation while employment held.

Job titles hide the real variance. Two people with the same title can have completely different exposure. An executive assistant who books travel and formats documents is exposed. One who manages a leadership team’s priorities, budgets and stakeholder relationships is not.

Erosion arrives through hiring, not firing. Companies rarely announce that a role was automated. They stop backfilling, raise the bar on new openings, and expand responsibilities for people who stay. The effect on you shows up as fewer openings and a harder market, months before any coverage names it.

The mid-level squeeze is the real story. Entry-level work is being absorbed because it was the most structured. Senior work is protected by accountability. The uncomfortable consequence is that the training ladder people used to climb from one to the other is losing its lower rungs.

Industry matters as much as occupation. The same job title carries very different risk in a healthcare system than in a media company, because the sector’s own growth dominates. Healthcare and social assistance is projected to supply about 37 percent of all new jobs through 2035, which lifts administrative and support roles inside it even as the same roles decline elsewhere.

Physical presence and legal responsibility are durable. A licence, a signature, a regulated duty of care, or a requirement to be physically present all create a boundary that better software does not cross.

The Five-Question Exposure Test

Score your own role honestly. Each yes is a point.

1. Could someone learn to do 80 percent of my week from written instructions alone?
2. Is most of my input already digital and structured?
3. Does my output follow a predictable format?
4. Is my work checked mostly for accuracy rather than for judgement?
5. Would a mistake in my work be caught by someone else downstream rather than by me?

Four or five points means the substance of your role is exposed and the sensible move is repositioning within the next year. Two or three means parts of your week will change and the risk is being slow rather than being replaced. Zero or one means your exposure is mostly about tooling fluency rather than about the role itself.

Work through a real example. A mortgage operations analyst reviews submitted files against a checklist, flags missing documents, and passes complete files to an underwriter. Question one: yes, the week is learnable from written instructions. Question two: yes, the inputs are digital. Question three: yes, the output is a standard flag. Question four: yes, accuracy is the measure. Question five: yes, the underwriter catches errors. Five points, and the honest conclusion is that this specific role has perhaps three years of stability rather than ten.

Now change one variable. Give the same person authority to approve exceptions up to a threshold, and question five flips, because now they are the last check. The role has not become technically harder. It has become accountable, and accountability is what organisations do not delegate to software.

The most useful part of this test is question five. Roles where the person is the last line of accountability are structurally protected, because someone has to own the outcome, and organisations do not assign ownership to software.

Where to Move, Depending on Where You Are

• If you scored high and work in a declining occupation: move toward the exception-handling end of your own field rather than leaving it. Claims work is contracting, but complex, disputed and fraud-adjacent claims are not. Underwriting is contracting, but non-standard risk is not.
• If you scored high but work in a growing occupation: the occupation is not your protection. Change what you do inside it. Take the work that involves deciding rather than producing.
• If you scored low: your risk is credibility. Being the person who cannot discuss how these tools work is a career problem in an organisation adopting them, independent of whether your tasks change.
• If you are entering the workforce now: target roles where you will be given decisions early, even at smaller organisations with lower pay. The learning ladder matters more than the starting salary in a market where junior production work is disappearing.

Building the Fluency That Protects the Judgement Half

Every one of these paths ends in the same requirement: being able to use these tools well enough to keep the compressed half of your job cheap, so you can spend your time on the half that is not compressible. That is a practical skill, not a philosophical position, and it is learnable in weeks.

What it looks like in practice is knowing which tasks to delegate, how to specify them so the output is checkable, where the predictable failure modes are, and how to build a review step that catches them. Learning that in a structured sequence beats picking it up ad hoc, because ad hoc learning produces habits while structured learning produces judgement, and a certificate alongside applied projects makes the capability legible to an employer. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.

FAQ

What to Do This Week

Take last week’s calendar and mark every hour as either producing something to a known specification or deciding something that needed a person. The ratio is your real exposure, and it is a more useful number than any occupation-level projection. Then pick the single largest producing block and work out how to hand most of it over, so the hours move to the other column while the choice is still yours to make.