AI and privacy
How an insurance broker compares policy conditions and client documents with AI and spots under- or overinsurance, without giving names to the AI.
Yes. An AI assistant can quickly put policy conditions, policy schedules and client documents side by side and point out where the sum insured, an exclusion or a clause does not match what the client owns or does. The condition is that client data does not end up unfiltered with the provider of that assistant.
That last point is exactly the problem. An adviser who pastes a client file into a chat window gives away the entire file: names, addresses, policy numbers, IBANs and sometimes a medical declaration for an income protection policy. The Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, warns that this can lead to a data breach. This article shows how it can be done differently, with a made-up example from the practice of a broker's firm.
An insurance broker works with more personal data than it often realises. A policy schedule names the policyholder, the address of the insured risk and sometimes the number plate. A valuation report describes a home or commercial building in great detail. A claim file contains statements from people, photographs and correspondence.
Income and absence insurance adds information about health. Health data is special category data: the GDPR prohibits its use, except in a limited number of exceptions. The AP explains this on its page about using and sharing health data. What falls within that category is set out in our article on special category data.
In addition, the business data of commercial clients is confidential: turnover, stock, machinery and contracts with customers. That is by no means always personal data, but a client does expect you to handle it with care. And as an adviser you need the information: the Dutch Financial Supervision Act (Wft) requires, in short, that an adviser gathers information about the client and bases the advice on it. So you cannot leave the files out of the analysis. You have to get them into the analysis safely.
PrivaSearch is software that lets your team search its own client files with Claude, ChatGPT or Mistral Vibe, after names, BSNs (Dutch citizen service numbers) and addresses have been replaced by pseudonyms and a colleague has checked that. For an insurance broker, it looks like this in five steps.
You can read how that connection works technically in our article What is an MCP server?. The full approach is on the page How PrivaSearch works.
A made-up but recognisable case. An installation company with its own workshop and a few vans has been a client for years. The business interruption insurance, the contents insurance and the liability insurance were taken out at some point and hardly reviewed since. The company has grown in that time and nowadays also installs heat pumps and solar panels.
The adviser puts the policy documents and the annual report in a project, a colleague approves them, and the adviser then asks his AI assistant: search this project for where the sums insured and conditions differ from what the annual report and the inventory list say about the company. The assistant searches the released passages and comes back with findings, each with its source. The answers show [ORGANISATIE_E349] where the name of the company was, and [PERSOON_B4A1] where the managing director was mentioned.
The assistant sees that the sum insured for contents and stock in the policy schedule is lower than the value stated in the inventory list and the balance sheet of the annual report. It quotes the passage from the conditions on underinsurance and refers to the pages where both amounts appear.
That matters, because in the case of underinsurance an insurer can leave part of the loss for the client's own account. In a 2021 ruling of the Kifid, the Dutch financial services complaints institute, an insurer paid only part of a burglary loss for that reason. The adviser therefore checks whether the policy includes an underinsurance waiver and whether it still applies: in another case the Kifid ruled that such a waiver could lapse under the conditions.
The reverse also happens. The assistant points out that a van which, according to the inventory list, has been sold is still on a policy schedule, and that a machine appears to fall both under the contents insurance and under a separate machinery insurance. Whether that is really double cover depends on the conditions of both policies. The assistant puts the two passages side by side; the adviser decides.
The description of the business activities on the liability policy mentions installation work, but no work on roofs. The annual report mentions installing solar panels as a growing activity. The assistant asks the adviser to check whether that work falls within the cover. That is the kind of question that easily gets left behind in an annual review, because the information is in two different documents.
A good analysis starts with targeted questions. A few examples that work well in practice:
| Question to the assistant | What the assistant uses | What the adviser does next |
|---|---|---|
| Which sums insured differ from the values in the annual report or the valuation? | Policy schedules, annual report, valuation report | Discuss the current value with the client and have the sum adjusted |
| Which business activities are not in the description on the policy? | Policy conditions, annual report, the client's website text | Ask the insurer whether the activity is covered |
| Which items appear on more than one policy? | Policy schedules, inventory list | Put the conditions of both policies side by side |
| Which exclusions affect this client's main risks? | Policy conditions, risk inventory | Assess whether additional cover is needed |
| Which clauses or terms expire soon? | Policy schedules, correspondence | Contact the client in good time |
Ask the assistant to name the passage and the page for every finding. Then you can check every statement in the document itself, and you also see when the assistant draws a connection that is not there.
An honest description of what you can expect belongs here too.
Detection is, moreover, never complete. That is why human review is mandatory. Pseudonymised data remains personal data for your firm: we explain the difference between pseudonymisation and anonymisation in our article Pseudonymise or anonymise?.
Three practical points make the difference between a trial and a way of working.
Connect only a business account. What PrivaSearch returns to the assistant becomes part of the conversation and reaches the provider of that assistant, with pseudonyms instead of names. So connect a business account with a data processing agreement. Which subscription qualifies is set out in our article Which subscription for business use?.
Do not type names into your assistant. If you ask the question in your AI assistant rather than in PrivaSearch, that question goes directly to the provider, bypassing the pseudonymisation. So ask about "this client" or about the pseudonym, not about the name of the managing director.
Work in a separate project per client. Permissions apply per project. That way the assistant only searches the file the question is about, and a colleague who is not in the project sees nothing of that client. Give the project an end date: after that, access is closed and copies, fragments and pseudonyms are cleared away, unless a retention obligation applies.
What is at stake if AI use in the office remains unregulated is explained on our page about AI chatbots and the GDPR. If you want to see how it works with your own policy documents, book a demo.
Not just like that. You then send all the personal data in the file to the provider, which under the GDPR requires a lawful basis, a data processing agreement and an assessment of necessity. Health data is subject to extra strict rules. A safer route is to pseudonymise the data first and have it checked, so that the assistant only sees released passages.
AI can point out where a sum insured differs from a value stated elsewhere in the documents, such as in an annual report or a valuation report. Whether there really is underinsurance depends on the current value and on the conditions of the policy. That assessment is made by the adviser, possibly together with a valuer.
Not if you search via PrivaSearch. The assistant receives passages in which names, addresses, BSNs, IBANs, number plates and dates of birth have been replaced by pseudonyms, and a colleague has checked that. What you type into the assistant yourself does go directly to the provider. So do not put names there.
No. PrivaSearch replaces personal data, not amounts, cover or descriptions of activities. Those are exactly what you need to compare the policy with reality. As a result, they do end up in the conversation with the assistant, and that is why you connect only a business account with a data processing agreement.
No. An assistant searches passages that are limited in size and can miss something or draw a wrong connection. Ask for the passage and the page with every finding, so that you can check it. Use the analysis as a first round that saves you time, not as a replacement for your own check.
Claude, ChatGPT Business and Mistral Vibe, via a secured MCP connection. You connect at your organisation's own address and approve the connection after you have signed in. Searching can also be done entirely within PrivaSearch itself: the connection is an extra, not a requirement.
For each Claude plan: what happens to your input, how long it is kept, whether an agreement is in place and where your data is stored. October 2026.
What you need to arrange for each type of data before using an AI chatbot: from names and email addresses to health, criminal records and BSNs.
Where an uploaded file ends up, how long it stays, who can access it and what deletion means. With four approaches for files containing personal data.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →