What 95 FTEs freed looks like: a benchmark of Dutch enterprise AI in 2025

Six Dutch enterprises. 875,000 interactions automated. 80.9% average end-to-end automation rate. 95 FTE equivalents freed. 4.2 million EUR in verified savings. These are not projections or pilot results. They are outcomes from production AI deployments running at scale through 2025.

This post covers what the benchmark data shows, how the results varied across deployments, and what the operational differences were between the highest and lowest performing implementations in the cohort.

The 2025 benchmark cohort: six Dutch enterprises

The cohort spans customer service and accounts payable automation across five sectors. The automation rates reflect end-to-end resolution: interactions or documents handled fully without human involvement, not partial deflection or first-contact resolution.

What 95 FTE freed actually means operationally

The 95 FTE equivalent figure requires context to be useful. It does not mean 95 people were made redundant. It means 95 people-worth of capacity was freed from repetitive, high-volume, low-judgment work and became available for tasks that require human decision-making.

At Novum Bank, returning thousands of hours of analyst capacity meant the credit team spent more time on loan assessments requiring judgment and less time on application status calls that add no value to the decision. At Bitvavo, customer service teams moved away from tier-1 query handling toward exception management and complex account issues. At CitizenM, guest experience staff focused on in-stay issues rather than booking queries that an AI can resolve faster anyway.

The composition of work changed across all six deployments. In each case, the highest-value human work expanded because the lowest-value human work contracted.

What drove variation in automation rates

The range in the cohort runs from roughly 80% to 93%. The factors that explain that variation are consistent.

Use case scope. CitizenM's 93% rate reflects a narrowly defined initial scope: guest communications around booking, check-in, and stay management. Tighter scope means higher automation rates in the first sprint. Broader scope means more edge cases, more exceptions, and lower initial rates, though the gap typically closes over subsequent sprints.

Data quality in source systems. Automation rate is partly a function of how complete and accurate the underlying data is. Where CRM records were complete and current, resolution rates were higher. Where data gaps existed, the AI escalated more frequently.

Escalation design. Deployments where escalation logic was explicitly designed, not defaulted, had cleaner handoffs and fewer repeated escalations. A well-designed escalation path reduces the fraction of interactions that cycle back through the AI after a failed handoff.

The deployment timeline finding

Every deployment in the 2025 cohort went live within 4 weeks of contract signing. The traditional enterprise AI project average is 5 to 9 months to production. That gap is not explained by the complexity of the use cases: Bitvavo's multi-language, compliance-sensitive deployment is not a simple use case.

It is explained by pre-built integrations. A platform with 100-plus native connectors to enterprise systems (Salesforce, SAP, Zendesk, AFAS, Dynamics) does not require months of custom integration work before the first sprint. The integration layer was already built. The sprint was about configuration, not construction.

The compounding consequence: an organisation that deploys in week 4 captures automation ROI in month 2. One that deploys in month 7 captures it in month 8. At the savings rates documented in this cohort, a 5-month difference is material.

FAQ

The Bitvavo case study and CitizenM case study are the most detailed public references from the 2025 cohort. The enterprise deployment timeline post covers what happens in each week of the implementation sprint.