AI for insurance brokers: automate the back office, not the advice

November 22, 2026

On 1 January 2020, 7,415 companies in the Netherlands held an AFM licence for financial services. On 1 January 2025 there were 6,551. Over the same years, the market share of firms with more than 25 FTE grew from 53.4% to 59.5%. The market is consolidating, and not slowly.

Meanwhile the insurers automate on their own side. According to Coverager, Nationale-Nederlanden now settles some simple claims, such as a broken windscreen, in about 6 minutes with no person involved. The insurance broker sits in between, with a back office that still does most of its work by hand.

So the question of which AI is best for insurance brokers is the wrong first question. The better question is: where do the hours of your back office go, and which part of them no longer needs a person? The advice stays with the adviser. That is what your clients pay for. The policy changes, the claims intake and the quote requests are a different story.

This article is for owners, directors and operations managers of Dutch insurance brokerages with 10 to 200 people. You have thousands of personal and commercial clients, a back office that carries the work, and growth that stalls on people you cannot find. It is not for you if you are a solo adviser who wants 1 chatbot on the website, if you run an underwriting firm with its own IT department, or if you are looking for a tool that writes your advice report.

In this article you will learn where the hours in a brokerage really go, which part AI does, which part fixed rules do and which part the adviser keeps, a worked example for a firm of 35 people, what you should not automate, how to keep the file in order for the AFM, and what happens after launch.

‍

Key takeaways

• The market is consolidating: 6,551 licence holders in 2025, against 7,415 in 2020, and the large firms are gaining share.

• Large firms do not win on better advice. They win on more policies per person in the back office.

• The hours sit in 4 flows: policy changes, claims intake, renewals of commercial policies and quote requests to several insurers.

• AI reads the email, the form and the policy. Fixed rules check. A staff member approves. The advice stays with the adviser.

• The real work is the connection to your broker system, not the language model.

• In the worked example, the work in those 4 flows drops from about 4,900 to 1,700 hours a year. That is more than 2 FTE.

• Every AI suggestion that lands in a client file is logged with the name of the person who approved it. That answers the AFM question before anyone asks it.

‍

AI for insurance brokers: why the question is not "which chatbot"

Most AI offers for this market start with the client. A chatbot on the website, an assistant that answers questions about cover, a tool that summarises the call. That can be useful. It just does not touch the place where your firm gets stuck.

A brokerage of 30 people usually has 8 to 14 people in the back office. They process the house move, the new car, the claim, the renewal and the quote request. Their work repeats, but it is not simple. Every change needs a look at the policy, a check in the broker system and often a message to the insurer.

When the book grows, that work grows with it. When you acquire a book, it grows overnight. And the people who can do it, who know the policy conditions and the broker system, are not found through a job ad posted on Monday.

In its report on the financial services sector, the AFM calls finding and keeping staff an ongoing challenge. In the same report, the head of supervision asks the question every AI use has to answer: who steers the algorithm, and who checks the outcome? You can read it in the AFM sector report for 2025 (in Dutch).

Those 2 questions make a good blueprint. AI belongs in the back office, where the work repeats and a staff member can check every outcome. It does not belong in the advice, where your adviser carries the responsibility.

‍

Where the hours in a brokerage really go

Ask a back office employee what fills the day, and you rarely hear "advice". You hear email. Email from clients, from insurers, from the repair shop. Behind that email sit 4 flows of work that come back at almost every firm.

1. Policy changes

A client moves house, buys another car, has a child or starts a business. The client emails, calls or fills in a form. Someone reads it, finds the client record, checks which policies are affected, enters the change and asks the insurer for an updated policy.

A simple change takes 10 to 15 minutes. That sounds small. At 6,500 changes a year it is a full job, and it is exactly the work where mistakes creep in: the old address left on the home insurance, the new car on the motor policy but not on the legal cover.

2. Claims intake

A claim arrives as an email with photos, a form from the insurer, an invoice from the repair shop and sometimes a voice message. Someone has to work out which policy it is, whether the damage is likely covered, what is still missing and which insurer it goes to.

The intake often takes 20 to 30 minutes per claim, before anything is settled. Most of that is reading, searching and sending an email back to ask for what is missing.

3. Renewals of commercial policies

For personal policies the renewal often runs by itself. For commercial clients it does not. Before renewal someone should check whether the sum insured still fits, whether the business has changed what it does, whether there were claims and whether the premium is in line with last year.

Done properly, that is 30 to 60 minutes per policy. So many firms do not do it properly, only for their largest clients. The rest renews silently, until a claim shows the sum insured was too low.

4. Quote requests to several insurers

A new commercial client wants a package policy. Someone fills in the request with 3 or 4 insurers, each with its own form or portal, then puts the answers side by side for the adviser. The only thing in a brokerage that grows faster than the number of policies is the number of portals.

A commercial request easily takes 2 to 4 hours, most of it retyping. The same data, 4 times, in 4 slightly different fields.

Want to map these 4 flows for your own firm? Our free AI product requirements template helps you record per flow how much work it is, which systems are involved and what a mistake costs. With that document you can also compare quotes from different providers fairly.

‍

Who does what: AI, fixed rules and the adviser

The pattern is the same in all 4 flows. AI reads. Fixed rules check. A person approves. The advice stays with the adviser. You never reverse that order.

AI for insurance brokers: who does what in a brokerage file, AI reads, fixed rules check, a person approves and the adviser keeps the advice

AI is good at reading what a client means. "We moved to Utrecht on 1 December and sold the old Golf" contains an address change, a start date and a cancellation. A language model pulls that out reliably, and it asks the right follow up question when something is missing, such as the registration of the new car.

Fixed rules are good at what must always be the same. Which policies are linked to this address? Does the new postcode fall in a different premium zone? Is the sum insured for the contents still in line with the home? You write those checks in code once, and they do the same thing every time after that. You do not let a language model do that, because it gets it right almost every time. In a policy administration, almost is not enough.

The back office employee then sees a proposal: this is the change, these are the affected policies, this is the message to the insurer. If it is right, it is 1 click. If not, the employee corrects it, and that correction goes back into the rules. The work shifts from retyping to checking.

We have seen the same pattern in other sectors. In a BPM vehicle tax file processor we built, a file went from 4 to 8 hours to under 5 minutes, with a specialist who signs every file. At a subsidy advisory, hours of searching for the right scheme became a matter of seconds. Different trades, the same split of the work.

‍

Worked example: a firm of 35 people

An example, not a client. A brokerage with 35 people, 12 of them in the back office. A mixed book of personal and commercial clients, recently extended with an acquired book. We put the internal cost of a back office hour at 50 euro.

Per year, the work in the 4 flows looks like this:

• Policy changes: 6,500 a year, from 12 to 4 minutes. From 1,300 to about 430 hours.

• Claims intake: 2,400 claims, from 25 to 10 minutes. From 1,000 to 400 hours.

• Commercial renewals: 1,100 policies, from 45 to 15 minutes. From 825 to 275 hours.

• Commercial quote requests: 600 a year, from 3 hours to 1 hour. From 1,800 to 600 hours.

In total the work drops from about 4,925 to 1,705 hours a year. That saves around 3,220 hours, more than 2 FTE. At 50 euro an hour, that is about 161,000 euro a year in capacity.

Against that sit the costs. A Blueprint of 7,500 euro to map and price the flows. Then 3 automations: policy changes 25,000 euro, claims intake 30,000 euro and quote requests 40,000 euro, together 95,000 euro. In this example the renewal check runs on the same connection as the policy changes. Total 102,500 euro. Operations cost 1,250 euro a month per automation, so 45,000 euro a year.

Net, about 116,000 euro a year remains. Payback is around 11 months. If you only get half the savings, 35,500 euro a year remains and payback takes almost 3 years. So always calculate both scenarios, and be honest about which one you believe.

More important than the amount is what you do with those 2 FTE. At most firms nobody leaves. The same team absorbs the acquired book, or next year's growth, without you having to find 2 people who do not exist. And the commercial renewals that now renew silently finally get a real check every year.

‍

The work is in the connection, not the model

The language model is the easy part. The hard part is that the change has to land in your broker system, on the right client, on the right policy, with the right start date. At many Dutch firms that system is ANVA or a similar package. On top of that, messages arrive through the SIVI standards insurers and brokers use to exchange data, and through the insurers' own portals.

For every connection you ask 4 questions. Can we read from the system? Can we write, or only prepare a draft? What happens when the system gets an update? And who, at the supplier or at your firm, manages the access?

The answers decide the route. Sometimes there is a good connection and the automation writes a draft change straight into the system. Sometimes there is only an export and import. And sometimes the most honest route is that the automation prepares a finished proposal the employee copies across in 30 seconds. That still saves most of the time, because the searching and reading are already done.

So always start with a draft, not with direct writes. Only when the drafts have been right for months in a row do you let the automation apply the simplest changes on its own. An address change with no premium effect, yes. A change to the sum insured, no.

‍

What you should not automate

Not everything that can be automated should be. Some things you keep with people on purpose.

• The advice itself. Which cover fits this client and this risk is your adviser's trade, and so is the responsibility.

• The judgment on a rejection or a doubtful claim. AI can sort the documents, a person decides and calls.

• Complaints. An angry client wants a name and a voice, not a well phrased email.

• A chatbot that answers cover questions on its own. If you let AI talk to clients directly, since 2 August 2026 you also have to make clear they are talking to an AI system. And a wrong answer about cover is exactly the kind of mistake that gets expensive later.

A simple rule of thumb: the bigger the damage of a mistake, the closer a person has to stand to it. A wrong address is annoying. A wrong answer on whether a claim is covered is a dispute.

‍

Keeping the file in order for the AFM

The AFM does not ban AI. It asks what it always asks: that you can show what happened and who was responsible. A good automation makes that easier, not harder.

In practice that means 3 things. Every AI suggestion is stored, together with the documents it was based on. Every approval has a name and a time. And every rule the system applies is written down somewhere, so that when the regulator or a client asks, you can show why a change was made the way it was.

Client data stays in the EU, in your environment, and the language model does not train on your files. You put that in the agreement with your provider, not in a promise on a website.

‍

How to start

You do not have to start with 4 flows at once. Start with 1, and choose it on numbers, not on feel.

• For 4 weeks, count how many policy changes, claims, renewals and quote requests come in.

• For 20 examples per flow, measure how many minutes they really take, from first email to done.

• Write down what a mistake costs per flow: time, money, an angry client or a problem with the regulator.

• Pick the flow with the most hours and the lowest cost of a mistake. That is almost always the policy changes.

Want to work it out yourself? Use the AI product requirements template as your guide. Our piece on how to map the AI opportunities in your workflow shows how to count the hours per step. If you would rather have it done, we do that in the Blueprint: 2,500 to 10,000 euro depending on the number of processes, in 10 to 20 working days. You get the process as it really runs, the bottlenecks in hours and euro, the opportunities ranked, a roadmap with fixed prices per milestone and a first rough prototype in your own repository. It is yours, also if you build on in house afterwards.

‍

Who works on it and what happens after launch

3 questions come up at every firm, and they deserve a direct answer.

Who works on it? 2 engineers by name, from the Blueprint to year 2. Not a team that changes every sprint. You know who to call, and they know your broker system, your insurers and your exceptions.

What happens after launch? Then operations start. A named engineer is on the alert when something breaks, the AI cost per file stays under a fixed cap, and you get a 1 page report every month. When an insurer changes a form or your broker system gets an update, we adjust the connection. The response time is set in the Blueprint.

And if month 1 disappoints? Exit after month 1: pay for what is delivered, keep everything.

Codelevate turns Dutch B2B service companies into software companies: we map the expert process you run by hand, automate it so the same team handles the volume, and build the platform your customers pay for, with 2 engineers by name from the blueprint to year 2.

Want to see first what automation pays back for a company your size? Read our overview of payback times for AI automation.

‍

Send us what you have

An export of your policy changes, a few anonymised claims, or just a description of a day in your back office. Send us what you have. 1 of the 2 engineers reads it with you in 20 minutes, and you get 1 of 3 answers in writing the same day: do it in house, hire, or the Blueprint.

Table of Contents
Share this article

Common questions

What can AI do for an insurance broker?

AI reads client emails, claim forms and policies, and prepares policy changes, claim files and quote requests for a staff member to approve. The advice itself stays with your adviser.

Is it allowed under AFM rules to use AI in a brokerage?

Yes, as long as you can show what happened and who was responsible, so store every AI suggestion with its source documents and log every approval with a name and a time.

What does it cost to automate a brokerage back office?

With us a major process costs 18,000 to 45,000 euro fixed per automation, plus operations from 1,250 euro a month, usually after a Blueprint of 2,500 to 10,000 euro.

Does AI work with ANVA and other broker systems?

Usually yes, but the connection is the real project: per system we check whether we can read, write or only prepare a draft, and we always start with drafts.

Which back office task should a brokerage automate first?

Almost always the policy changes, because they have the most hours and the lowest cost of a mistake, so measure 4 weeks of volume before you choose.

Should we put an AI chatbot in front of our clients?

Usually not first: since 2 August 2026 you must tell clients they are talking to AI, and a wrong answer about cover costs far more than the hours a chatbot saves.

Get started with
an intro call

This will help you get a feel for our team, learn about our process, and see if we’re the right fit for your project. Whether you’re starting from scratch or improving an existing software application, we’re here to help you succeed.