Hoe AI de sectoren IT en zakelijke dienstverlening verandert

In het kort (RaboResearch’s own framing): In de sectoren IT en zakelijke dienstverlening wordt AI gezien als een belangrijke versneller om de productiviteit te verhogen en personeelstekorten op te vangen. Het automatiseringspotentieel is groot: zo’n 86% van IT-taken en 64% van taken in de zakelijke dienstverlening kan (deels) worden geautomatiseerd. AI zal vooral routinematige, analytische en tekstgebaseerde werkzaamheden sterk beïnvloeden. Experimenten tonen aanzienlijke productiviteitswinsten van circa 10% tot soms meer dan 50%, afhankelijk van taak en toepassing. Bedrijven die AI goed integreren kunnen hun productiviteit structureel verhogen en personeelstekorten opvangen; bedrijven die dit niet doen verliezen terrein.

TL;DR

A Dutch-market sector study from RaboResearch (Rabobank’s research arm), published 25 June 2026, applying the ILO occupational-exposure methodology (Gmyrek et al. 2025, ILO Working Paper 140 — not Brynjolfsson; see note below) to Dutch CBS occupation data to estimate AI’s automation and productivity potential for the IT (Informatie & communicatie) and zakelijke dienstverlening (business services) sectors. It is the wiki’s first Dutch national / sectoral source on AI labour and productivity, and its first source to quantify automation potential against the CBS Nederlandse Beroepsclassificatie. The substantive payload is in 13 figures: a task-level automation-potential breakdown (86% IT / 64% business services), a per-business-function decomposition, a 18-study productivity-gain spread that spans +73% down to negative, NL-vs-EU adoption rates, and per-occupation automation rankings. The closing prescription names three conditions for capturing the potential — strategic ownership, a data/IT/process base, and governance/trust — that map cleanly onto the Warner & Wäger digital-transformation primitives.

What was actually ingested

• Full article prose came through the Zotero local-API extract (fulltext_source: zotero-extracted) — the complete running text, the In het kort summary, all section prose, the Onderzoeksverantwoording appendix, and the single literature reference (Gmyrek et al. 2025).
• The 13 figures were image charts dropped by the text extract. Because the figures carry the load-bearing quantitative content, they were fetched from the live page (the rabobank.nl host bot-blocks plain fetchers; rendered in a real browser, the media.rabobank.com CDN serves the chart images directly) and read visually. Figure data below is transcribed from those charts (Figuren 4, 5a, 5b, 6, 7a, 9a, 9b read in full; Figuren 1–3, 7b, 8a, 8b are context/trend charts adequately described by the body prose).
• Honest scoping: this is a ~2,400-word web research article, not a long report. No body chapters are deferred — the whole article is ingested.

Key findings

Automation potential at task level (Figuur 4)

RaboResearch scores each task in the Dutch occupation classification on a 0–1 GenAI-automatability scale (mapping Gmyrek et al.’s ILO/ISCO task scores onto the CBS Nederlandse Beroepsclassificatie), then bins tasks into four categories: hoog potentieel (score > 0.5), potentieel (0.375–0.5), laag potentieel (0.2–0.375), niet mogelijk (< 0.2). “(Hoog) potentieel” = the first two bins combined.

Sector	Hoog	Potentieel	(Hoog) potentieel	Laag	Niet mogelijk
Informatie & communicatie (IT)	49%	37%	86%	11%	3%
Zakelijke dienstverlening	33%	31%	64%	27%	9%
Totale economie	~22%	~22%	~44%	~28%	~28%

IT and business services are among the most exposed sectors in the whole economy (financial services is the only sector with a comparable or higher share); the least-exposed sectors are agriculture, horeca, and construction. The scores are a measure of potential when the technology is further developed and implemented — not current deployment.

Where the potential sits, by business function (Figuren 5a / 5b)

Decomposing each sector’s high-potential tasks into Eurostat/CBS business functions:

• IT (of all tasks in the sector): ICT 41%, Marketing/verkoop 10%, R&D/innovatie 9%, Anders 7%, Productie/serviceprocessen 7%, Administratie 6%, Boekhouding 5%, Logistiek 2% (≈86% total).
• Zakelijke dienstverlening: Boekhouding/financieel beheer 14%, Marketing/verkoop 10%, ICT 9%, Productie/serviceprocessen 9%, Anders 8%, Administratie 7%, R&D 5%, Logistiek 2% (≈64% total).

The productivity-gain spread (Figuur 6) — a task-level vs firm-level gap

RaboResearch’s clearest contribution is a single chart collating ~18 productivity studies, which lays the task-level-vs-firm-level gap bare in one view — task-level/occupational experiments cluster high, firm-level studies cluster near zero:

Study (domain)	Reported gain
Ju & Aral 2025 (advertising)	+73%
Peng, Kalliamvakou, Chion & Demirer 2023 (software dev)	+55.8%
Gambacorta, Qiu, Shan & Rees 2024 (software dev)	+55%
Yeverechayou, Mayya & Oestericher-Singer 2024 (software dev)	+37–55%
Noy & Zhang 2023 (writing)	+0.8 SD
Goldman Sachs Research 2026 (firm-level)	+30%
Cui, Demirer, Jaffe, Musolff, Peng & Salz 2025 (software dev)	+26%
Dell’Acqua et al. 2025 (consulting)	+25.1%
Paradis et al. 2024 (software dev)	+21%
METR 2025 (software dev)	+18%
Fang, Yuan, Zheng, Donati & Sarvary 2025 (e-commerce)	+16%
Brynjolfsson, Li & Raymond 2024 (customer service)	+14%
Dell’Acqua et al. 2024 (product dev)	quality+
Bloom, Barrero, Davis et al. 2026 (firm-level)	+1.4%
Otis, Clarke, Delecourt, Holtz & Koning 2023 (entrepreneurship)	~0%
Babina, Fedyk, He & Hodson 2023 (firm-level)	0% TFP
Atlassian Research 2025 (firm-level)	96% report no ROI
Niederhoffer, Teevan & Jaffe 2025 (knowledge work)	−2 units/incident

RaboResearch’s own caveat is explicit: “het is onzeker of deze effecten standhouden buiten experimenten” — implementation, integration and skills determine whether experimental gains survive into daily practice. (Note: the METR row is plotted +18%; METR’s own widely-cited July-2025 RCT found experienced OSS developers were ~19% slower with AI — the sign is worth flagging as a possible RaboResearch reading/metric difference, not taken at face value.)

Netherlands ahead of the EU average, not in the lead group (Figuren 7a/7b)

On a CBS enterprise survey (firms >10 employees), Dutch firms in both sectors use AI above the EU average in every business function, but below the EU leader in every function. For business services: e.g. administratieve processen NL 25% vs EU-leader 33% vs EU-avg 14%; marketing NL 16% vs 31% vs 10%.

Per-occupation automation potential (Figuren 9a/9b)

• Business services (most → least exposed): boekhoudkundig medewerkers (~100% hoog), secretaresses, administratief medewerkers, receptionisten/telefonisten, financieel specialisten, boekhouders, accountants — down to architects and managers (mid), and schoonmakers / beveiligingspersoneel at the bottom (mostly niet mogelijk / laag).
• IT (most → least exposed): callcentermedewerkers outbound (~88% hoog), databank-/netwerkspecialisten (~70%), gebruikersondersteuning ICT (~56%), marketing/PR/sales-adviseurs (~55%), software- en applicatieontwikkelaars (~42% hoog, ~100% (hoog) potentieel) — down to managers ICT, procesoperators, and algemeen directeuren (no hoog, mostly potentieel/laag).

The pattern: routine administrative, financial, and text-based roles are most exposed; physical and top-management roles least. Senior developers shift toward quality assurance, architecture and IT governance (“context waarbinnen AI en agents opereren”); junior/instap functions decline.

The three AI types and the three conditions

RaboResearch distinguishes Predictive AI (decades old — capacity planning, fraud detection), Generative AI (post-2022 — documentation, code generation), and Agentic AI ((semi-)autonomous — resolving incidents, continuously testing software) (Figuur 3). It names three conditions for an organisation to capture the potential:

1. Strategic ownership — leadership positions AI explicitly in the business strategy and actively drives employee adoption, supported by new core skills (data literacy, critical AI use).
2. A data/IT/process base — AI only delivers value when data is reliable, secure and well-governed and applications are embedded in daily workflows; many initiatives fail on poor integration and weak fit to the business context.
3. Governance, risk management and trust — clear frameworks for privacy, ethics and explainability so employees, clients and regulators can trust AI-supported services (see responsible AI).

Dynamic-capabilities reading (Warner & Wäger)

The source is tagged against four W&W process-model cells:

• digital-sensing/digital-scenario-planning — the study is a scenario-planning artifact: it scouts AI’s task-level automation signal, interprets a forward scenario for two sectors, and formulates the strategic implication (“groei komt niet langer uit extra mensen, maar uit slimmer werken”).
• digital-sensing/digital-mindset-crafting — condition 1 names exactly this work: leadership establishing AI as part of strategy and promoting a digital/AI mindset across the workforce, underpinned by new core skills.
• digital-transforming/improving-digital-maturity — condition 2 (a reliable, governed data/IT/process base; embedding AI in workflows; leveraging digital knowledge inside the firm) and condition 3 (governance of sensitive data) are digital-maturity work.
• contextual/external-triggers — the study’s macro framing is a set of external pressures: structural personeelsschaarste (high vacancy rates), a consolidation wave squeezing margins, and rising client expectations at lower cost — the triggers making productivity growth a necessity rather than an option.

(The governance/trust condition has no dedicated W&W cell; it is folded into improving-digital-maturity prose above and linked to responsible AI rather than stretching the vocabulary.)

Linked entities and concepts

• Concepts: AI employment effects, automation vs augmentation, enterprise AI adoption, micro-productivity trap, generative AI, AI agents, responsible AI, Warner & Wäger process model.
• Entity (incidental): Brynjolfsson appears only as one row in the productivity table (Brynjolfsson, Li & Raymond 2024, customer service, +14%); the study’s actual methodological anchor is Gmyrek et al. (2025), ILO.
• Dangling (single-source authors, deferred per the second-source promotion rule): Mark van Kampen (Sectormanager Tech, Media & Telecom), Reinder Koelewijn (Sectormanager Commercial Services), Jesse Groenewegen, Floris Jan Sander. RaboResearch / Rabobank and CBS and the ILO / Gmyrek et al. working paper are likewise single-source mentions, not yet promoted to entity pages.

Source-to-source relationships

• supports → [[sources/2026-04-28-brynjolfsson-canaries-coal-mine|Brynjolfsson, Chandar & Chen — Canaries in the Coal Mine?]] — both measure AI’s labour impact through occupational exposure; RaboResearch’s Dutch sector vantage (executing→advisory role shift, declining junior/instap tasks, value concentrating in complex judgment) corroborates the US ADP-payroll entry-level/compositional finding from a different instrument and geography.
• supports → [[sources/2026-04-03-bcg-emerson-kropp-ai-will-reshape-more-jobs-than-it-replaces|BCG — AI Will Reshape More Jobs Than It Replaces]] — same reshape ≫ replace headline at sector altitude: roles transform (accountants, lawyers, consultants, developers move up the value chain) and junior tasks decline, rather than aggregate job elimination.
• supports → [[sources/2026-03-05-massenkoff-mccrory-anthropic-labor-market-impacts-ai|Massenkoff & McCrory / Anthropic — Labor Market Impacts of AI]] — method cousins: both convert task-level capability ratings into an occupational-exposure measure (Anthropic weights Eloundou ratings by AEI usage; RaboResearch maps Gmyrek/ILO task scores onto CBS occupations) and both stress that potential exposure ≫ realised deployment.
• Considered and deferred (no typed edge): [[sources/2025-11-25-yee-mgi-agents-robots-and-us-skill-partnerships|MGI Agents, Robots, and Us]] (>50% of work-hours theoretically automatable, shift-not-elimination) and Evans (lump-of-labor counter-frame) are thematic neighbours but already well-connected through the AI employment effects concept page; no additional source-to-source edge added.

Provenance

Acquired from the local Zotero ai-wiki collection via zotero-acquire (zotero_item_key: 7A3RK8MR, itemType webpage → raw/articles/). The raw stub carries the Zotero text extract; figure data was read from the live rabobank.nl page at Process time. Original article: RaboResearch, 25 June 2026.