AI can take real work out of a Dutch brokerage: advertisement copy, viewing schedules, inbox triage, the first pass at a Wwft file. What it cannot take is the signature. Dutch rules already fix three points where a named person has to be able to show their reasoning, and one of them is written into a valuation regulation that reads like an AI governance policy. Everything outside those three points is negotiable.
What does a week in a Dutch brokerage actually consist of?
Before deciding what to automate, name the systems. Most Dutch makelaarskantoren run Realworks, the CRM that has dominated the sector since 2004, or Kolibri, which has grown as the lighter alternative; both publish to funda. The floor area in that listing is not a free number. NVM and Vastgoed Nederland, the body VBO and Vastgoedpro merged into on 1 January 2025, oblige their members to use the meetinstructie gebruiksoppervlakte woningen, the practical translation of NEN 2580, and have done since 2010. A valid energielabel has to exist and be registered before transfer, and the ILT enforces it after the fact, which is the awkward part: once the deed has passed, you can no longer register a label at all. And where a mortgage is involved, the lender wants a validated report, in practice almost always through the NWWI.
Swap those nouns for an accountancy practice and nothing in this article survives. That is the test. A generic assistant that has never heard of a meetinstructie will cheerfully write you a listing that creates liability, and it will do it in fluent Dutch.
Where is AI already safe in a makelaarskantoor?
Three places, and they share a shape: the model drafts, a person releases.
Advertisement copy is the obvious one, and it is already shipping inside the software you pay for. Kolibri includes AI for listing text; other CRMs connect to the same kind of service. The work is a first draft from structured fields you already hold, and someone still presses publish.
Inbox and viewing correspondence is the higher-volume, lower-risk one. Sorting enquiries, proposing viewing slots, chasing confirmations. Nothing here produces a number that a third party relies on.
Document assembly is the underrated one. Pulling the kadastraal uittreksel, the VvE stukken, the energielabel status and the eigendomsbewijs into one checklist and flagging what is missing is genuinely useful, and the model is only finding gaps. It is not forming a judgement. That distinction is the whole argument of this article, and Dutch valuation supervision has already written it down.
Why is a model value not a valuation?
The NRVT brought hybrid valuations into its rules in July 2021 under the name Bureauwaardering. As of 1 January 2025 the Bureauwaardering lapsed and the Reglement Hybride waardering took its place, approved by the Centrale Raad on 24 October 2024.
The wording is the interesting part. In NRVT’s own description, a hybride waardering does not constitute a valuation by a register-taxateur. It is defined as the combination of a Modelwaarde and the assessment of that model value by a register-taxateur, and the outcome is the value of the home. The taxateur does not value. The taxateur assesses, and can approve or reject. The model supplier carries responsibility for the model value and the references it supplies; the taxateur carries responsibility for the assessment. Use is limited to a home for own occupation, for obtaining, refinancing or adjusting mortgage finance.
Read that as an AI governance document and it is unusually mature. It names the model owner. It names the human reviewer. It separates their liabilities. It restricts the scope of use to one decision type. And it says out loud that the reviewer’s job is a binary accept or reject rather than a quiet edit of the machine’s number. Most AI policies written in 2026 do less than that, in more words.
What did the NRVT disciplinary board actually rule?
In 2023 the NRVT tuchtcollege decided a case that every AI project in the country could learn from. A homeowner had a desktoptaxatie done in May 2022 to adjust his mortgage. It came out at 500,000 euro. A later physical valuation put the market value at 630,000 euro. He complained.
The taxateur explained his method openly: the desktop product generated 25 references, he saw 6 of them, and from those he selected the best 3. The board did not rest its decision on the value itself. It looked at the reasoning. The three references, corrected to 587,000, 565,000 and 550,000 euro, were named in the report but never analysed, and how those figures produced 500,000 was not explained anywhere. Without that analysis the report was not inzichtelijk en navolgbaar, which breached the fundamental principle of care and transparency in Article 12 of the Reglement Gedrags- en Beroepsregels. The complaint was upheld and a warning imposed.
Notice what was not held against him. Not using a model. Not the funnel from 25 to 6 to 3. Not even the 130,000 euro gap, which was not what the decision turned on. What sank the file was that it did not show the step from the comparables to the number. If you cannot reconstruct, months later, why an output was accepted, the fact that a qualified person accepted it will not save you. That is the most transferable lesson in Dutch AI governance available right now, and it did not come from Brussels.
What does the Wwft still oblige you to do yourself?
Makelaars and taxateurs in onroerende zaken are Wwft-instellingen, and supervision moved house this year. Since 1 January 2026 it sits with the Dienst Financieel-Economische Integriteit, a directorate of the Ministerie van Financiën that absorbed Bureau Toezicht Wwft and Bureau Economische Handhaving from the Belastingdienst. Open files and investigations carried over unchanged, and so did the duties.
For a makelaar those duties start with risk-based cliëntenonderzoek, and the leidraad written for this sector by Bureau Toezicht Wwft, still the operative guidance after the handover, is worth reading for one detail that catches small offices: your due diligence reaches the wederpartij of your client, not only your client. Unusual transactions go to FIU-Nederland. Staff have to be trained periodically. Results have to be documented. Note that the leidraad does not treat the two roles identically, and in the direction most people guess wrong: the cliëntenonderzoek obligation does not apply to a taxateur at all, while every other duty, the reporting of unusual transactions included, applies to makelaar and taxateur alike.
AI can assemble the file, extract identity data, check it against what you were told, surface a possible PEP or sanctions hit for a person to look at, and keep the training record tidy. It cannot decide that a transaction is unusual and it cannot file the melding. There is also a trap worth naming: the training obligation exists so that your people recognise unusual transactions. Automating the recognising away and then reporting that your staff are trained is a contradiction a supervisor can read in a single visit.
Does the EU AI Act make any of this high-risk?
Mostly no, and the honest answer matters because the sector is being sold the opposite. Annex III does not list property valuation. Point 5(b) covers creditworthiness assessment of natural persons, which is the lender’s system rather than your listing generator. And the high-risk obligations under Article 6(2) and Annex III were pushed to 2 December 2027 by the AI Omnibus, in force since 27 July 2026.
What does apply to a brokerage is smaller and nearer. Article 4 has required AI literacy among the people using these systems since February 2025, and the same Omnibus rewrote it: a deployer no longer has to ensure a sufficient level, but must take measures to support one. That is an obligation of effort rather than result, which is easier to meet and, being an effort standard, is judged on what you can show you did. Article 50 transparency has applied since 2 August 2026: a chatbot answering property questions on your site announces itself, and generated or materially altered images need marking. A virtually staged living room is exactly that case. It is also, separately, a misleading-advertising question that your beroepsvereniging and a disappointed buyer will raise long before any AI regulator does. Our AI Act checklist for Dutch SMEs walks the same ground for firms outside this sector.
What does this cost a four-person brokerage, and when does it pay back?
An illustration with the assumptions stated, not a benchmark. Take a four-person office doing 120 transactions a year. Assume drafting the listing, assembling the Wwft file and handling viewing correspondence take roughly 135 minutes per transaction today. Assume AI removes 45 of those minutes and no more, because the checking does not shrink and the checking is the point. That is 90 hours a year. At an internal loaded cost of 55 euro an hour, about 4,950 euro. Subtract tooling at 40 euro per user per month, or 1,920 euro a year. Net, roughly 3,000 euro.
Three thousand euro a year buys configuration, not construction. A custom build pays back over six or seven years on those numbers, which is not an investment but a hope. For an office of this size the right answer is to switch on what Realworks or Kolibri already ships, integrate nothing, and put the recovered time into the two tasks a model is barred from. Custom work starts to earn its place above roughly 500 transactions a year across several branches, where the same 45 minutes is worth more than 20,000 euro annually and integration cost stops dominating the sum. We would rather write that down than sell the smaller office a project. Our cost breakdown for an AI project sets out where the lines actually fall.
Where should a small brokerage start?
In this order. First, listing drafts inside the CRM you already pay for. Second, inbox triage and viewing scheduling. Third, Wwft file assembly as a checklist that flags gaps while a named person still decides. Do not start with valuation, and do not start with anything that produces a number a third party will rely on.
Then write one sentence for each of the three: who approves this, and what does the file have to show afterwards. That is your AI policy. It is also the literacy duty discharged in a form you can demonstrate, and it is the NRVT ruling applied to work that has no register behind it. If it helps to see the same reasoning in a neighbouring sector, we did this for insurance advice offices and for accountancy practices, and the pattern of connecting a model to the systems you already run is set out under process automation. If you would rather have someone walk the three steps with you, that is what an AI consultant for the Dutch mkb is for.
