An underwriter at an insurer receives an update from the vendor of the underwriting system: a new feature triages applications by risk. That same week a mortgage adviser has ChatGPT make an advice report easier to read, with the client's income and family situation still in it. Two kinds of AI, two kinds of questions, one organisation that must answer both.
That is the core of AI governance in this sector: there are two layers. The AI in the processes themselves, where the AI Act asks whether an application is high-risk. And the daily use of general tools by advisers, underwriters and claims handlers, where the questions are about tools, accounts and client data. Most organisations spend their attention on the first layer and have little sight of the second. A workable approach covers both, and what is AI governance explains why it starts with seeing.
Where does AI sit in the work?
At a mortgage lender or insurer, in more places than the annual report mentions.
| Where in the work | What AI does there | Which data | Which question belongs to it |
|---|---|---|---|
| Application and intake | Read application files, write up meeting notes | Income, IBAN, citizen service number, family situation | Which tool, which account, and what goes into the prompt |
| Underwriting and pricing | Triage applications, estimate risk, set premium or terms | Everything in the file, sometimes health | Is this high-risk under the AI Act, and who oversees it |
| Claims handling | Summarise claims, assess documents, draft letters | Registration plate, policy, for injury medical information | Which data may go in, and is the outcome a decision about a person |
| Client contact | Chatbot on the site, draft replies, summarise email | Name, file, the client's question | Does the client know they are talking to AI, and what does the employee paste |
| Back office and reporting | Analyses, summaries, code | Transaction and portfolio data | Which account type, and what is shared |
The second row is the layer the board already talks about. The other four are where most of the use sits, and where most often nothing has been decided.
When is an application high-risk?
The AI Act lists in Annex III, under access to services such as credit and insurance, two applications that touch this sector directly: systems that evaluate the creditworthiness of natural persons or establish their credit score, with an exception for fraud detection, and systems for risk assessment and pricing in life and health insurance. Mortgage underwriting assesses creditworthiness. A system with that purpose falls under the list in principle.
Two cautions belong with that. The criterion is the intended purpose of the system as the provider has laid it down, and whether your organisation deploys it that way; that is an assessment per system. And Article 6(3) holds a narrow exception for systems that do not materially influence the outcome of a decision, but it does not apply to systems that profile people. What it does and does not cover is in when is an AI application high-risk. Motor or home insurance is not in the list; a system that assesses claims is therefore not automatically high-risk, but it takes decisions about people to which the GDPR applies.
Whoever uses such a system is a deployer under the AI Act. From 2 December 2027 that comes with a list: use the system according to the provider's instructions, assign human oversight to people who can intervene, follow how it performs, and inform those affected. The date was moved by the Digital Omnibus and may move again; what applies now and what later is in the AI Act for employers. The same tool can be low-risk in one deployment and high-risk in another: one AI tool can be both.
What happens in daily use?
The second layer rarely reaches the boardroom, and that is where most use sits. An adviser writes up a meeting note in ChatGPT. A claims handler has a claim summarised. That is usually not high-risk. It does fall under three other things.
The GDPR, because almost every prompt contains a client detail, and in a free account without a data processing agreement that is a data breach. Article 4 of the AI Act, which since 2 February 2025 asks for measures on AI literacy: employees need to know what does and does not belong in a prompt. And the duty of care that supervisors such as the AFM and DNB look at; careful handling of client data is part of that, and a tool nobody has decided on does not fit in. Since 2 August 2026 there is one more thing for the chatbot on the site: the client must know they are talking to AI. How that plays out for a firm under supervision is in AI in financial services.
DORA adds a layer that is often forgotten: the AI services employees use are third-party ICT services, and one nobody knows about is in no register.
What does a workable approach look like?
It starts with knowing what is there, in both layers. For the processes that means: per system, the intended purpose the vendor has laid down, the data that goes in, and an owner. For daily use: which tools are actually opened, how often, and on which account. That second list cannot be made with a survey; it changes every month. Both belong in one AI register, with a column for process AI or tool in daily use.
Then a decision per row. For the underwriting system: high-risk or not, who oversees it, which DPIA, and what the vendor has to supply. For the tools: one of four statuses, with the condition in one sentence for "limited use" and the alternative named for "not allowed". Everything starts at "no decision yet", and that is an honest interim state, not a shortcoming.
The decision then has to reach the workplace at the moment it matters, not in a policy document. Suppose an insurer puts this in place. After three months the underwriting AI has an owner and a DPIA, Perplexity is "limited use" without client data in the query, and the free ChatGPT account is "not allowed" with Copilot as the alternative. What does not happen: nobody has a tool cut off, and nobody is confronted about their use. The figures do the work: the share of use through allowed tools is the yardstick, and growing tools without a decision top the list for the next assessment.
BeeSensible matches a catalog of 865 AI tools against the use observed in the organisation, with no user id in the usage data. An admin sees which tools have no decision yet and records it; the decision appears in the browser when someone opens the tool, and the figures on what happens next sit aggregated in the dashboard. Your vendor's underwriting system it does not see; that enters the register through procurement.
