Which 150 Jobs Can AI Replace by 2027?

Why the 2027 window, not 2030

Most public AI-displacement work — Goldman's 300-million-jobs figure, McKinsey's 30% of work-hours estimate, the World Economic Forum's Future of Jobs Report 2023 — uses a 2030 horizon or longer. That choice makes the headlines bigger and the analysis softer, because by 2030 nobody can be cleanly held to account for a forecast made in 2026. We picked 2027 deliberately. It is 18 months out from this writing, which is long enough for the next model generation to ship and for firms to restructure entry-level pipelines around it, but short enough that the answers have to be grounded in capabilities that already exist or are credibly imminent.

The shorter window also forces honesty about deployment, not just capability. A model that can technically do 70% of a paralegal's document-review tasks is not the same as a law firm that has rebuilt its entry-level rung around that model. Goldman's 2023 Briggs and Kodnani report is explicit on this distinction — their headline 300-million-jobs exposure is task exposure, not net jobs eliminated, and they expect the displacement to play out over a decade because deployment takes that long. A 2027 horizon prices in the deployment lag in a way a 2030 one does not have to.

Finally, 2027 is the horizon where BLS data can be a real check on us. The May 2024 OES release is the most recent comprehensive U.S. wage snapshot, and the next two annual releases (May 2025 and May 2026) will land before our window closes. If our Tier-A occupations are right, you will see their employment counts in those releases drop or flatten relative to the broader labor market. If they don't, our model is wrong and we should be willing to say so.

How we define 'replace' — tasks vs jobs

The single most important distinction in this entire literature is between task automation and job elimination, and it is the distinction most often elided in news coverage. Frey and Osborne's 2013 "Future of Employment" paper, the one that originally seeded the "47% of U.S. jobs at risk" framing, scored jobs by the share of tasks that were susceptible to computerization. It did not predict that 47% of U.S. jobs would disappear by any specific date, and the actual U.S. employment-to-population ratio in 2024 was higher than in 2013. The mechanism is more subtle: as task content shifts, firms restructure roles, headcount per unit of output falls in some occupations and rises in others, and the wages and entry rungs reprice.

Eloundou et al.'s 2023 "GPTs are GPTs" paper is the methodological backbone of the per-row percentages in our table. They scored 19,000-plus occupational tasks from O*NET for direct LLM exposure (the model alone can do it twice as fast at the same quality) and LLM+software exposure (with additional tooling). Their headline finding was that 80% of U.S. workers had at least 10% of their tasks exposed, and 19% had at least 50% exposed. They were careful — repeatedly — to note this was task exposure, not job loss.

Throughout the table and tier descriptions in this guide, the "% tasks AI can do by 2027" column is task exposure, not headcount reduction. A 70% number for a Tier-A occupation means current frontier models, deployed with reasonable scaffolding, can plausibly perform 70% of the documented tasks of that role at acceptable quality by end of 2027. What firms actually do with that capability — fire workers, reduce hiring, restructure roles, or expand output — is a labor-market and product question, not a model-capability question.

The 4 displacement tiers explained

Tier A is roles where the median worker's task content is dominated by language understanding, structured-format generation, lookup, simple reasoning over documents, and routine voice or text interaction. The capability bar to do most of this is already cleared by 2025-vintage frontier models. The deployment bar — building the eval harness, integrating with the source-of-truth system, getting compliance sign-off — is the rate limiter. We score these roles at 60%+ task automation by 2027 because both the capability and the deployment economics are in place.

Tier B is the messy middle. The models can do a real share of the work but failure modes matter and customers or employers will accept a non-trivial human-in-the-loop cost for the next 18 months. Junior software developers, mid-level copywriters, technical writers, market analysts and recruiters all sit here. McKinsey's generative AI report puts much of this band at 30–50% activity automation in their generative-AI scenario, and our 30–60% range is consistent with theirs.

Tier C is the augmentation band — mid-level designers, senior engineers, sales executives, financial analysts, project managers. The work is too contextual, too political, or too cross-domain for the model to own end-to-end, but a meaningful share of the artefacts these roles produce can be drafted or accelerated. The competitive dynamic for individuals in this band is that productivity per worker rises, so headcount growth slows even if absolute headcount doesn't fall.

Tier D is the resistant band. Trades, hands-on healthcare, teachers, plumbers, social workers, top executives. The work is either physically embodied, deeply human-relational, or strategically ambiguous in ways that frontier models still cannot reliably handle and that humanoid robotics will not commercially deploy in scale by 2027. Wages here will probably rise relative to exposed cognitive work over the window, which is itself a major macroeconomic prediction of this guide.

Tier A: the jobs already shrinking in 2026 BLS data

The most striking thing about the BLS Occupational Outlook Handbook projections published in 2024 and updated in 2025 is which occupations they project to shrink over the next decade. Data entry keyers, word processors and typists, switchboard operators, file clerks, and telemarketers all have negative projected employment growth, and the BLS narrative explicitly attributes the contraction to software automation. These are the canonical Tier-A roles, and they are shrinking in the official forecast even before you layer in a separate generative-AI shock.

What changed between the 2024 projections and what we're forecasting for 2027 is the addition of a layer of roles that pre-LLM automation never threatened: tier-1 customer support, basic copy editing, simple legal document review, basic accounting reconciliation, basic ad-ops, basic transcription, and large parts of marketing operations. These roles previously survived prior waves of automation precisely because they required reading, summarizing, and generating natural language. The 2023–2025 model generation broke that moat. Our Tier-A list includes them not because they will be zero in 2027, but because the per-headcount task share will fall enough that employers stop hiring at the entry rung.

The honest mechanism here is rarely a layoff. It is hiring freezes, redefined roles that fold three former jobs into one, and a steeper ramp-up expectation for any human still hired. Indeed Hiring Lab's tracking of job postings in 2024 and 2025 already shows year-over-year declines in postings for several Tier-A categories even as overall U.S. job postings recovered. If you are in or near a Tier-A role today, the right time to start the transition is now, not 2027 — by 2027 the hiring market for the destination role will be more competitive.

Tier B: where models meet messy reality

Tier B is the most contested band in the literature. Goldman's 2023 work bakes a large share of these roles into the "exposed" category that drives the 300-million headline figure. McKinsey's generative-AI report is roughly in the same zone but more conservative on deployment timing. OECD's Employment Outlook 2023 takes a notably more cautious view and stresses that highly-exposed occupations have so far seen rising, not falling, employment in OECD countries.

Our 30–60% band is wide on purpose. Junior software developer is a good worked example. The model is genuinely capable of writing most of the code a junior dev writes in their first six months, and frontier-lab internal reports plus public Anthropic and OpenAI commentary suggest a large share of code at those companies is now model-assisted. But the firm-level question is whether the senior engineer who reviews the model's output is willing to absorb the throughput of three to five virtual juniors, and whether the team's hiring plan reflects that. Many teams are restructuring; most are not yet.

Mid-level copywriting and technical writing are further along. Several large content shops and SaaS documentation teams publicly restructured in 2024 and 2025 to be model-first with a senior editor on top, and the entry-level job postings for those roles dropped accordingly. Recruiter screening is a similar story — sourcing and first-pass screening can be done end-to-end by current models with a human only on candidate calls and final decisions, and several large recruiting firms have publicly disclosed this shift. The 50–55% numbers we score these roles at reflect both capability and observable deployment, not theoretical capability alone.

Tier C: augmentation, not replacement (yet)

Tier C is where the MIT and NBER literature on AI productivity is most useful. Brynjolfsson, Li and Raymond's NBER paper on a generative-AI deployment at a Fortune 500 customer-support firm found roughly 14% productivity gains on average, much larger for less experienced workers. The Noy-Zhang MIT study on writing tasks found roughly 40% time savings and 18% quality improvement when knowledge workers were given access to ChatGPT. These are augmentation findings — the workers got faster, not replaced.

We score mid-level software engineers, mid-level financial analysts, accountants, project managers, and most managerial roles in the 10–30% band because the available evidence consistently shows productivity gains in this range rather than role elimination. Mid-level designers and sales executives sit here for the same reason — the model accelerates a lot of artefact production, but the relationship work, the cross-functional judgement, and the politics of getting a thing actually shipped or sold are still done by the human. Goldman's report explicitly carves these roles out of the most-exposed category for similar reasons.

The thing to watch in the Tier-C band is whether productivity gains compound into headcount reductions over the 2026–2027 window. Historically they have not — productivity gains in knowledge work have driven output growth, not employment reduction. But the size of the AI productivity gain is unusually large by historical standards, and several large consulting firms and investment banks publicly restructured analyst classes in 2024 and 2025 in ways that suggest they expect Tier-C augmentation to translate into smaller cohorts, not bigger ones.

Tier D: jobs that hold their value through 2030

Tier D is the part of the analysis we feel most confident about. The roles in this band — registered nurses, physicians, dentists, plumbers, electricians, HVAC techs, firefighters, police, social workers, kindergarten and elementary teachers, childcare workers, hairstylists, mechanics — share two properties: a large share of their daily task content is physically embodied, and another large share is deeply human-relational. Neither current frontier LLMs nor the leading humanoid-robotics platforms will commercially deploy at scale to take over these roles by 2027.

The BLS Occupational Outlook Handbook actively projects faster-than-average employment growth in many Tier-D roles — nursing, home health and personal care, mental health counseling, several skilled trades. Demographic demand is the primary driver, and the supply of workers in these roles is the binding constraint, not demand. None of the AI capability gains over the 2026–2027 window are going to change that picture in a serious way.

The interesting Tier-D nuance is wages. If our broader picture is right and Tier-A through Tier-C roles see headcount-per-unit-output fall, the relative wage of Tier-D work should rise, because (a) the absolute supply of Tier-D workers is constrained and (b) some displaced workers from exposed roles will try to retrain into Tier-D — but the retraining cost into trades and clinical roles is high and slow. The likely outcome is a meaningful wage premium in Tier-D occupations over the 2026–2030 window, which is a quietly bullish story for the trades and clinical labor markets that does not get the attention it deserves.

What the macro studies actually predict (Goldman, McKinsey, OECD, MIT, NBER)

The macro studies disagree, sometimes by enormous margins, and any guide that pretends they don't is selling something. Goldman Sachs's 2023 Briggs and Kodnani report estimates that generative AI could expose the equivalent of 300 million full-time jobs to automation globally and lift global GDP by 7% over a decade. McKinsey's 2023 generative-AI report estimates 60–70% of employee time across all activities is theoretically automatable when generative AI is added to prior automation tech, with a midpoint scenario in which generative AI adds the equivalent of $2.6–4.4 trillion annually in productivity.

OECD's 2023 Employment Outlook is the most cautious of the major institutional studies, noting that highly-exposed occupations in OECD countries have so far seen employment rise rather than fall and stressing that policy and worker bargaining power will determine outcomes as much as capability. Frey and Osborne's original 2013 paper estimated 47% of U.S. jobs were at high risk of computerization — a number that did not materialize as job loss over the subsequent decade. The IMF's 2024 Gen-AI and the Future of Work paper finds roughly 40% of global employment exposed, with higher exposure in advanced economies.

The micro-level academic literature — Eloundou et al. at OpenAI/Penn, Brynjolfsson and co-authors at NBER, Noy and Zhang at MIT — converges more tightly. Task-exposure shares in the 30–60% range for white-collar knowledge work are consistent across these studies, and the productivity-gain magnitudes when workers have access to LLMs are consistently in the 14–40% range depending on task type. We weighted our per-row percentages toward this micro-level consensus, with the macro studies acting as sanity-check on the overall tier shape rather than per-occupation numbers.

The honest summary is: every credible study agrees that exposure is large and concentrated in cognitive routine work; the studies disagree on whether exposure translates into net job loss versus productivity-driven employment growth. Stanford's HAI AI Index 2024 is a useful neutral aggregator if you want to read the individual studies yourself.

Geographic and demographic disparities in exposure

Exposure is not evenly distributed across the U.S. or globally. Goldman's 2023 work, the IMF's 2024 paper, and the OECD's 2023 outlook all converge on the finding that high-income economies and high-income workers within them face disproportionately high exposure, precisely because their work is more cognitive and language-heavy. The U.S., U.K., Canada, Australia and parts of continental Europe top global exposure rankings; low-income economies, where a larger share of work is physical and informal, face lower direct exposure but other indirect effects.

Within the U.S., the exposure pattern follows the geography of administrative, financial-services, marketing, paralegal, and customer-support work — the major metros on both coasts plus Chicago, Atlanta, and Dallas. The BLS State Occupational Employment data shows these metros have disproportionate concentrations of Tier-A and Tier-B occupations from our table. Trades-heavy regions face the opposite picture: lower direct exposure, and a likely relative wage gain over the window.

Demographically, the exposure also skews. Women hold a disproportionate share of administrative, customer-support, and several Tier-A roles, which means displacement risk is gendered in the same way prior service-economy shocks have been. Younger workers face higher entry-rung risk because Tier-A and Tier-B roles are disproportionately the ones early-career workers start in. The World Economic Forum's Future of Jobs Report 2023 and the LinkedIn Workforce Reports both flag these patterns, and the policy implication is uncomfortable: the workers least equipped to retrain are also the ones most exposed.

What to do if your job is on the list

If your role is in Tier A or upper Tier B, the right action is to move within the next 12–18 months, not to wait for the headcount conversation to find you. The cheapest move is laterally into a Tier-C or Tier-D role that uses adjacent skills. A Tier-A paralegal moving into a more strategic legal-operations or compliance role inside the same firm has a much shorter retraining path than a paralegal trying to switch industries. A Tier-A copy editor moving into developmental editing or content strategy has the same advantage. The general pattern is to climb the value-of-judgement curve inside your existing domain.

If your role is Tier C, the right action is to become unusually fluent with the tools that augment your role and to push to be the person on your team who designs how the tools are deployed. The Brynjolfsson NBER paper and the Noy-Zhang MIT paper both find that the largest productivity gains accrue to less-experienced workers, which is great for the worker but a problem for the senior — your relative advantage compresses unless you become the one steering the deployment. Our freelance rate with vs without AI calculator is a useful pricing tool if you bill hourly or by project and need to reprice for AI-augmented output, and the contractor rate cards for AI-augmented work guide has the market-rate context.

If your role is Tier D, the right action is mostly to stay and to negotiate. The macro picture suggests Tier-D wages will rise relative to exposed cognitive work over the window. There is no urgency to switch out, and there is a reasonable case for switching in if you are in a Tier-A role with a credible retraining path into a trade or clinical occupation. Finally, if you manage hiring at any level, take the cost of replacing a junior dev with AI calculator seriously before pulling the trigger on a junior cohort — the failure mode of skipping a generation of junior hires is that you have no senior bench in five years.

Frequently asked questions