AI and the Future of Work: Which Roles Are Growing, Shrinking, and Transforming

Artificial intelligence is no longer a future consideration—it's a present-tense economic force. The World Economic Forum's 2024 Future of Jobs Report projects that AI and automation will create 69 million new jobs while displacing 83 million by 2027—a net reduction of 14 million roles, or roughly 2% of global employment.

The Jobs That Are Growing

Contrary to popular narrative, AI creates more specialized roles than it eliminates. The fastest-growing job categories through 2027 (WEF data):

1. AI and Machine Learning Specialists — 40% projected growth
2. Sustainability Specialists — 33% projected growth
3. Business Intelligence Analysts — 31% projected growth
4. Information Security Analysts — 31% projected growth
5. Data Engineers — 30% projected growth

LinkedIn's 2024 Jobs on the Rise report corroborates this, noting that "AI" appeared in 3.6x more job postings in 2024 compared to 2022.

The Jobs That Are Shrinking

Roles with high routine cognitive content face the greatest disruption:

• Data Entry Clerks — 26% projected decline (already accelerating with GPT-based automation)
• Administrative Assistants — 18% projected decline
• Bookkeeping and Payroll Clerks — 16% projected decline
• Bank Tellers — 15% projected decline

Goldman Sachs estimates that 300 million full-time jobs globally could be affected by generative AI, with roughly two-thirds of U.S. occupations having at least some tasks that could be automated.

The Jobs That Are Transforming

The most important category is neither growth nor decline—it's transformation. McKinsey Global Institute estimates that 60% of all occupations have at least 30% of their activities that could be automated, but fewer than 5% of occupations can be fully automated. This means the vast majority of jobs will change, not disappear.

Examples of transformation:

• Marketing Managers — Shifting from content creation to AI-augmented strategy, prompt engineering, and performance optimization
• Financial Analysts — Moving from data gathering and modeling to insight interpretation and strategic recommendation
• Software Engineers — Evolving from code writing to system design, AI supervision, and architecture
• Lawyers — Transitioning from document review to complex judgment, client counseling, and AI-assisted legal strategy

The Skills That Matter

The WEF identifies the top 10 skills for 2027:

1. Analytical thinking
2. Creative thinking
3. AI and big data
4. Leadership and social influence
5. Resilience, flexibility, and agility
6. Curiosity and lifelong learning
7. Technological literacy
8. Design and user experience
9. Motivation and self-awareness
10. Empathy and active listening

Notice the balance: half are technical, half are deeply human. The future belongs to professionals who combine both.

What You Should Do Now

• Become AI-literate, not AI-dependent — Understand what AI can and cannot do. Use it as a tool, not a crutch.
• Invest in judgment skills — AI can generate options; humans must evaluate them. Practice critical thinking, ethical reasoning, and complex decision-making.
• Build a T-shaped skill set — Deep expertise in one area plus broad literacy across adjacent domains makes you irreplaceable.
• Use simulations to explore — Job simulations let you test-drive roles in growing fields before committing to a career change.

The European Picture Is Not the Global One

Global aggregates hide the only thing a European candidate can act on: where the demand actually sits, under which regulation, and at what price. Three structural facts reshape the WEF numbers once you land them in Europe.

First, the continent starts from a shortage, not a surplus. Eurostat counted roughly 9.8 million ICT specialists in the EU in 2023 — about 4.8% of total employment — against the Digital Decade target of 20 million by 2030. Doubling a workforce in seven years is not a displacement problem; it is a sourcing problem. In France, the shortage is concentrated in data engineering, industrial software and security operations rather than in model research.

Second, regulation creates jobs before it creates products. The EU AI Act entered into force on 1 August 2024, with obligations for general-purpose AI models applying from 2 August 2025 and the bulk of high-risk system requirements from 2 August 2026. Every high-risk deployer needs risk management, data governance, technical documentation, human oversight design and post-market monitoring. Those are headcount lines: AI compliance officers, model risk analysts, evaluation engineers, technical writers who can document a system to an auditor's standard. This category barely existed in the 2022 job market.

Third, the transformation is industrial, not only digital. France Stratégie and DARES, in their forward-looking work on occupations, point to simultaneous pressure in engineering, maintenance and care occupations — driven by retirements as much as by technology. In practice, the fastest-filling AI roles in France in 2025-2026 are not in pure research labs but inside manufacturers, banks, insurers, energy operators and public agencies that must industrialise models they did not build.

Where the Hiring Actually Happens

Read job flow, not headlines. Four employer archetypes absorb most of the AI-adjacent demand in France and the wider EU, and each buys a different profile.

• Model builders (small, loud, selective). Mistral AI, Hugging Face, LightOn, academic spin-offs and the research arms of large groups. They hire in the tens, expect publications or serious open-source output, and are the least likely route in for a career changer.
• Industrial and financial deployers (large, quiet, volume). BNP Paribas, AXA, Société Générale, Schneider Electric, Airbus, EDF, SNCF, Sanofi. They hire in the hundreds: data engineers, MLOps, integration engineers, product owners who can specify a model in business terms.
• The services layer (highest volume, fastest entry). Capgemini, Sopra Steria, Accenture, Onepoint, Devoteam plus specialist boutiques. Structured graduate intake, apprenticeship pipelines, and the most forgiving door for non-linear backgrounds.
• Public and regulated bodies (growing fastest from a small base). Ministries, ANSSI, CNIL, hospital groups and regional agencies now recruit for AI governance, procurement scrutiny and data protection engineering.

The practical consequence: the growth statistics belong mostly to archetypes two, three and four — the ones almost nobody optimises their application for.

Compensation Landscape

Indicative gross annual base salary in France, permanent contracts, variable pay and equity excluded. Île-de-France carries roughly a 10-15% premium over regional packages; consulting adds bonus but usually starts lower on base.

Role	0-2 years	3-6 years	7+ / lead
Data engineer	42 000-52 000 €	55 000-70 000 €	75 000-95 000 €
Machine learning engineer	45 000-55 000 €	60 000-80 000 €	90 000-120 000 €
MLOps / platform engineer	44 000-54 000 €	58 000-75 000 €	85 000-105 000 €
Security / SOC analyst	38 000-46 000 €	50 000-62 000 €	70 000-90 000 €
AI governance / model risk	40 000-48 000 €	55 000-70 000 €	80 000-100 000 €
Business intelligence analyst	36 000-44 000 €	46 000-58 000 €	62 000-78 000 €

Two asymmetries matter more than the absolute numbers. The premium for deployment skills — pipelines, monitoring, cost control, regulatory documentation — has risen faster than the premium for modelling skills, because deployment is where projects fail. And the spread inside a single title now exceeds the spread between titles: two ML engineers with the same job name can sit 35 000 € apart depending on whether they own production systems.

Entry Routes, Ranked by Actual Conversion

1. Apprenticeship or alternance (highest conversion). One to three years inside the employer, with conversion rates to permanent contracts that no external application matches. Structurally under-used by career changers who assume it is only for 20-year-olds.
2. Adjacent internal move. The fastest documented path into an AI role is from a non-AI role at a company already deploying AI. Domain knowledge plus new tooling beats tooling alone.
3. Services and consulting intake. Broad, recurring, tolerant of atypical CVs, and it buys you exposure to five industries in three years.
4. Evidence-first lateral application. A public artefact — a working pipeline, a reproducible evaluation, a documented case — outperforms certificate stacking. Simulations serve the same function: they produce assessable evidence rather than claimed familiarity.
5. Bootcamp or MOOC alone (lowest conversion). Useful as scaffolding, weak as a signal, because it certifies attendance rather than judgement.

Two Counter-Arguments Worth Taking Seriously

"This time is not different." Automation anxiety has a two-century record of over-predicting displacement: ATMs coexisted with rising teller headcount for two decades before it fell. The honest rebuttal is not that AI is magic, but that its target is different. Previous waves automated physical and routine clerical work; this one touches drafting, summarising and first-pass analysis — the tasks that historically trained juniors. The risk is less mass unemployment than a broken apprenticeship ladder.

"The productivity gains will simply fund more hiring." Sometimes true, and the evidence from early enterprise deployments is genuinely mixed: measured gains cluster in support and coding, and evaporate where data quality is poor. But gains accrue to the firm, not automatically to the role. A team that becomes 30% more productive is only expanded if demand for its output is elastic. Ask that question about your own function before assuming the upside lands on your side of the table.

Three Scenarios to 2030, and How to Read Them

Forecasting a single number is a category error. What a professional needs is a small set of futures and a signal that tells them which one is arriving.

1. Diffusion (most likely). AI spreads unevenly through large organisations, gated by data quality and regulation. Net employment barely moves; task content changes almost everywhere. The winners are people who can connect a model to a process — integration, governance, change management. The early signal: growth in job titles containing "platform", "governance" or "operations" rather than "research".
2. Compression. Capable agents automate whole entry-level workflows in support, back office and basic analysis. The junior rung thins, and the classic three-year apprenticeship-by-osmosis disappears with it. The signal: falling graduate intake at the large services firms alongside stable revenue. In this world, structured evidence of capability — simulations, assessed work samples, portfolio artefacts — becomes the only credible substitute for the experience nobody will pay you to acquire.
3. Constraint. Energy cost, compute scarcity, litigation or an enforcement shock slows deployment. Demand concentrates in audit, security and remediation, and the premium shifts from building to proving. The signal: rising spend on assurance relative to development.

All three scenarios reward the same underlying asset: demonstrable judgement over a real workflow. None of them reward a certificate.

The Next 90 Days, Concretely

• Weeks 1-2 — locate yourself. Write down your five most time-consuming tasks and mark each as automatable, augmentable or protected. That map, not a job title, determines your exposure.
• Weeks 3-6 — build one artefact. Take one augmentable task and rebuild it with a model in the loop, end to end, including how you would monitor and document it. One finished, defensible artefact outperforms four courses.
• Weeks 7-10 — get it assessed. Put the work in front of someone who hires. A simulation, an internal proof of concept or a reviewed open-source contribution all convert claimed skill into observed skill.
• Weeks 11-13 — target the second archetype. Apply where volume is: deployers and services firms, with the artefact as the opening line of the application rather than an annex to it.

Method and Limits

What this article is, and is not, so you can weigh it against your own situation.

• The WEF figures are employer expectations, not measurements. The Future of Jobs Report aggregates surveys of large employers about intentions. Net job projections are directional; they have historically been revised.
• Task exposure is not job loss. The Goldman Sachs and McKinsey estimates quantify tasks technically susceptible to automation. Adoption is gated by cost, regulation, data quality and organisational inertia — which is precisely why the transformation category is larger than either growth or decline.
• Salary bands are a synthesis, not a wage survey. They are built from published French job postings, APEC-type market reporting and recruiter ranges, expressed as gross base excluding bonus, profit-sharing and equity. Treat them as negotiation anchors with an error margin of roughly ±10%.
• Geographic scope. Compensation and entry routes describe France and, with adjustment, Benelux and Southern Europe. German and Nordic bands sit higher; Central European bands sit lower.
• What would change our view. A sustained fall in AI-adjacent posting volume across two consecutive quarters, or an EU AI Act enforcement delay, would move the governance roles from growth to hold.