Your Job Risk · Belgium · Counter-view
What if AI grows the jobs it looks like it's killing?
The rest of this site treats AI exposure as risk: tasks a model can do are tasks a worker might lose. This page flips the frame. If cognitive work gets 10× cheaper, history says total demand for it can rise, not fall. Coal in 1865. ATMs in the 1990s. Code in 2026?
Section 01 · The original paradox
When something gets cheaper, we often use much more of it.
In The Coal Question (1865), William Stanley Jevons observed that as steam engines became more efficient, Britain burned more coal, not less. Cheaper energy per unit of work expanded the set of things worth doing with energy.
The modern restatement: when an input becomes cheaper, total consumption of that input can rise if demand is price-elastic and there is latent unmet demand. Lighting, computing, bandwidth and air travel have all walked this path.
The AI Jevons claim, in one line
If generative AI cuts the unit cost of a cognitive task by an order of magnitude, total output of that task will rise — and so, in many cases, will employment of the humans who steer it.
Historical analogues: ATMs and bank tellers (Bessen, 2015). Spreadsheets and accountants. Desktop publishing and graphic designers. E-discovery and paralegals. In every case the per-task cost collapsed, the "X is dead" headline ran loud, and employment in the occupation kept rising for years.
Section 02 · The three conditions
Three things have to be true for AI to grow a job instead of killing it.
When AI makes a task much cheaper to do, employment in that occupation can grow — but only if all three of these conditions hold at once. If any one fails, you get displacement instead.
Cheaper output has to pull in new buyers and new use-cases that didn't exist at the old price. Software, marketing, translation, legal research, customer support, design and analytics all show large unmet demand once the price drops.
The worker stops producing the artefact by hand and starts directing, reviewing, integrating and signing off on what the AI produces. Each remaining person becomes more productive — and more valuable — not redundant.
Once the model can do the routine task, the constraint moves to judgement, taste, domain expertise, distribution, trust and regulatory sign-off. Those things still sit on people, so hiring shifts toward them.
Section 03 · What to look for, in order
A Jevons effect is visible before it is dominant.
- 01Output per worker jumps measurably (PRs merged, tickets resolved, words translated, contracts reviewed).
- 02Unit price of the output falls — or quality at fixed price rises sharply.
- 03Volume rises faster than price falls. Total spend on the task increases. (The diagnostic Jevons signal.)
- 04Job postings reshape, not just shrink: senior "own the AI workflow" roles rise, narrow producer roles fall, juniors are squeezed.
- 05Adjacent occupations boom (more code → more PMs, data engineers, security, compliance).
- 06Wages of the survivors rise even as headcount in the narrow specialism plateaus.
Section 04 · Where the signal is showing up — in Belgium
Seven Belgian occupation clusters to watch first.
Each one is highly AI-exposed, has elastic demand, and has structural Belgian depth. The badge says how visible the Jevons-shaped signal is here today.
Software & data engineering
Early signalDevs, data engineers, platform teams across Brussels–Antwerp–Leuven, plus Smals and the EU institutions.
25–55% productivity uplifts on bounded coding tasks in published RCTs; total EU developer headcount still growing in 2024–2025; entry-level CS hiring visibly weaker than other graduate segments. Output up, composition reshaping toward senior.
Belgian read
Agoria's annual Digital Skills Barometer continues to flag developer and data roles as structural shortage occupations even after two years of mass Copilot/Claude adoption. VDAB and Le Forem both keep ICT roles in the top knelpuntberoepen / métiers en pénurie lists. Belgian software employers — Odoo, Showpad, Collibra, Itineris, Smals, the EU agencies in Brussels — are still net hiring through 2025–2026 even as per-engineer output has visibly stepped up. The textbook Jevons signature.
Translation & localisation
Early signalEU institutions, NATO, the Belgian federal services, and a dense private localisation cluster (Yamagata Europe, Lionbridge BE, Lexitech).
Per-word prices in steady decline; translated word volumes — especially in multilingual EU work — keep growing. Post-editing, terminology and quality assurance are the new shape of the job.
Belgian read
DG Translation has run eTranslation as a productive tool for over a decade, and the Brussels EU bubble is one of Europe's largest single concentrations of language professionals. The Chambre Belge des Traducteurs et Interprètes (CBTI-BKVT) and the KU Leuven / UCLouvain / UMons translation programmes report a steady shift away from raw translation toward revision, terminology and AI quality control rather than collapse.
Customer support & contact centres
Mixed signalProximus, Telenet, Orange Belgium, KBC, Belfius, ING Belgium, AG Insurance, Engie BE — tier-1 and tier-2 service desks.
Klarna, IKEA, BT and others reported large deflections to AI agents; Klarna later walked back parts of its claims. Total support employment has not collapsed, but the role mix is moving up the value chain. Whether this is steady-state Jevons or a transition is still open.
Belgian read
Proximus has publicly described AI as central to its transformation plan and has already booked job impact in customer-facing roles. KBC's Kate assistant has been live for years and KBC continues to hire customer-experience staff around it. Telenet, Belfius and ING Belgium have all rolled out generative-AI assistants in 2024–2025. The early Belgian pattern is augmentation at the senior end and compression at tier-1 — visible in the layoff strip on the homepage.
Legal, compliance & regulatory
Early signalBrussels Bar associates, in-house counsel and compliance officers in finance and pharma. Magic-circle and Belgian firms: Stibbe, Eubelius, Liedekerke, Loyens & Loeff, Linklaters and Clifford Chance Brussels.
Tools like Harvey, CoCounsel and Lexis+ AI are cutting research and contract-review time sharply. Headcount is roughly flat to up; volume of matters reviewed is sharply up. The bottleneck has moved to partner sign-off and to compliance — both human.
Belgian read
Belgium sits inside three regulatory waves at once — the EU AI Act, DORA for financial services, and the post-GDPR enforcement build-out by the APD/GBA. Each one expands work that must be signed off by a human lawyer or compliance officer. The Brussels EU practice is one of the largest single concentrations of regulatory work in Europe; AI cuts the research time, but the regulator still wants a name on the memo.
Marketing, brand & content
Early signalIn-house marketing teams in retail (Colruyt, Delhaize, AB InBev), telecom (Proximus, Telenet, Orange BE), FMCG and B2B HQs, plus Belgian agencies (mortierbrigade, Famous, BBDO Brussels, TBWA).
Output per marketer has exploded; spend on paid distribution and on senior marketers who can direct AI pipelines is rising. Junior copy roles and content-mill freelancers are the casualty.
Belgian read
UBA member surveys and BAM's annual marketing barometer describe a consistent pattern across Belgian advertisers in 2025: more campaigns, more variants, more NL/FR/EN localisation per market — produced by teams that grew at the senior end and shrank at the junior end. Multilingual delivery, historically a Belgian cost disadvantage, is becoming a Belgian Jevons advantage.
Audit, finance & controlling
Mixed signalBig Four audit (PwC, Deloitte, EY, KPMG Belgium), corporate controllers and management accountants across BEL 20 firms.
Spreadsheets did not kill accountants; AI may repeat the trick at a higher altitude. Per-engagement effort is falling; total volume of assurance and analysis work continues to grow under tightening reporting standards.
Belgian read
CSRD and the Belgian transposition continue to expand the assurance perimeter (sustainability data, supplier reporting). The IBR-IRE (auditors institute) and ITAA (accountants and tax advisers) both report ongoing shortages, not surpluses, of qualified professionals through 2025–2026 — with AI tooling spreading on top of that demand, not in place of it.
Healthcare professionals using AI
Not yet visibleRadiology, pathology, primary care GPs, nursing leads across the UZ Leuven, UZ Brussel, UZA, CHU Liège and Cliniques Saint-Luc networks.
Per-scan and per-record AI assistance is real and growing. But Belgium's aging demographics and the structural shortage of GPs and nurses push demand up faster than any conceivable efficiency gain pushes it down.
Belgian read
RIZIV/INAMI data and the Federal Knowledge Centre (KCE) continue to project shortages, not surpluses, across most clinical roles through 2030. Domus Medica and the SSMG flag a worsening GP shortage, especially in Wallonia. AI here is a relief valve on workload, not a substitute for headcount — the textbook case of Jevons being dominated by underlying demand growth.
Caveat: at the macro level, neither Belgium's 2025 LFS nor the OECD Employment Outlook 2025 yet show AI as net job loss at the occupation-group level. That is consistent with a slow Jevons unwind and with delayed displacement. The sector-level signals above are the place this will become distinguishable first.
Section 05 · Where the Jevons argument breaks
Four honest objections.
Demand may be satiable
Elastic demand is an empirical question, not a theorem. If the world doesn't actually want 10× more legal memos, marketing copy or translated PDFs, lower unit cost just shrinks the wage bill.
Complementarity may not survive the next model
If the AI gets good enough to also do the review and integration step, the human bottleneck moves out of the occupation entirely. This is the agentic scenario and it is the strongest argument against a Jevons reading.
Distributional pain is real even if totals grow
Junior cohorts, narrow specialists, and workers without the tooling, language or training to ride the new workflow can lose badly while aggregates look healthy. A macro Jevons effect is cold comfort to a 24-year-old whose entry-level role just disappeared.
Adjustment is slow
Even in the optimistic reading, the gap between layoff announcement and the new hiring wave is years, not quarters. The Belgian social model makes that gap more survivable than most — but it does not erase it.
Section 06 · Synthesis
Risk and Jevons aren't opposites. They're two phases of the same transition.
Phase 01 · Today
The visible, painful one.
Firms book the productivity gain as cost reduction. Layoffs land in the high-AIOE roles. Juniors are first to be cut. The news cycle is dominated by displacement stories.
Belgium is here today. The layoff strip on the homepage is what phase one looks like.
Phase 02 · The Jevons phase
Prices fall. Volumes explode. Roles rebuild.
Headcount in the affected occupations stabilises or grows, but composition shifts permanently toward senior, integrative, accountability-bearing work.
Whether Belgium spends three years or fifteen between the two phases — and how many workers fall through the gap — isn't a property of the technology. It's a policy and institutional choice.
Section 07 · Further reading
The papers and primary sources behind this page.
- Jevons (1865) — The Coal Question
- Bessen (2015) — Learning by Doing
- Brynjolfsson, Li & Raymond (2023) — Generative AI at Work
- Peng et al. (2023) — Copilot RCT
- Anthropic (2025) — Economic Index
- Imas (2026) — What will be scarce?
- Handa et al. (2025) — Claude task-use evidence
- Acemoglu (2024/2025) — Simple Macroeconomics of AI
- OECD Employment Outlook 2025
- ILO (2025) — GenAI & jobs: refined global index
Belgian institutional data