Agents in production for a long time, never re-assessed
They were validated when they went live, and have been running since on products, cover and channels that have changed. Nobody checks any more that they route correctly.
Sector · Insurance
AI in insurance is already running. Does anyone know what it returns?
AI in insurance is already here: claims triage, assistants in contact centres, sales support across distribution networks, copilots at head office. Insurers know what these tools cost, rarely what they return. At an insurer, an agent's mistake does not show straight away. It is paid for further down the line, in rework for claims handlers and in delay for the policyholder. We measure real usage from the logs your platforms already produce, and from interviews with your leadership teams.
- 40+AI diagnostics delivered in a year
- 41,200users measured
- 218qualified use cases
- 20 to 40%of paid AI seats unused across the estates analysed (an engagement finding, not a study)
Company markers, across all our engagements: none is attributed to a client, none is a result guaranteed in advance.
In short
In insurance, AI is used mostly in customer relations, claims handling, distribution, head-office functions and IT teams. It is rarely measured: more than half of the AI-equipped companies Idun Group meets measure nothing, or rely only on the vendor's statistics. Idun Group's audit of real usage combines usage logs with interviews, compares what agents do with what happens after them, and says what to industrialise, consolidate or cut.
AI in insurance today: many uses, little measurement
Of the AI-equipped companies Idun Group has met, 93 said how they measure their AI usage: 39 have no measurement at all and 14 rely only on the vendor's statistics. In large companies (5,000 employees and more), 18 of 29 measure nothing. In insurance, usage is established but still young: according to the EIOPA survey, published in February 2026 by the European insurance supervisor and covering 347 undertakings in 25 countries, 65% of the undertakings surveyed already use generative AI; 64% of the use cases they report are still at pilot stage, 32% are in production.
At insurers, generative AI came in first through customer relations: 40% of the 347 undertakings EIOPA surveyed already use it there. Claims handling comes next (32%), then sales and distribution (22%). At head office, Microsoft 365 Copilot, ChatGPT or Claude have often been rolled out to support functions seat by seat. IT teams have their coding assistants.
We do not touch the other uses of artificial intelligence in insurance: pricing, underwriting and risk selection, actuarial models, fraud detection, automated claims decisions.
The working AI already in place is the AI teams use every day: assistants, copilots, agents that prepare files. What it returns remains to be measured. One insurer told us it was after quality, not ROI. Quality can be measured too: the work handlers redo, processing times, the policyholders who give up halfway through a conversation.
Why insurers talk to us
Insurance runs on case files and deadlines, and every exchange with a policyholder puts a contractual promise on the line. Six situations come up again and again when insurers talk to Idun Group.
Invisible human rework
When an agent is slightly wrong, handlers catch it, file by file. That time appears on no dashboard; it shows up in processing times.
An organisation split by entity and by product
Life, property and casualty, health, protection, salaried or intermediated networks: each entity built its own assistants, often for the same use cases.
Seats opened at head office, usage unknown
Microsoft 365 Copilot, ChatGPT or Claude have been rolled out to support functions, often by the group IT department. The data team of a mutual insurer summed it up for us: it does not deploy the tools, it is a "consumer" of them, and it does not try to find out what they return.
Data under constraint
Personal data and GDPR, trade secrets, sector regulation: the risk function, the data protection officer (DPO) and compliance want to know what enters the agents' context before extending anything. The data protection team of one insurer told us it has no access to any logs and relies on what the teams declare.
Proof before scaling
An assistant tested in one contact centre or one region is only rolled out further once its value is measured.
AI use cases in insurance, and what measurement reveals
AI use cases in insurance fall into five territories: claims handling, customer relations and contact centres, sales and distribution, head-office functions, IT teams. For each one, the table gives what usage analysis reveals and the decision it makes possible. It describes what Idun Group encounters in the sector, not work Idun Group has delivered.
| Use | What it is | What the analysis reveals | The decision it enables |
|---|---|---|---|
| Claims handling | AI agents that receive, read and route claim notifications: triage by type and urgency, completeness checks, pre-qualification, file preparation for the handler. Always in support of a human decision. | The gap between the routing the agent gives and what the handler then does with the file, read in an export from your claims management system (statuses, reassignments, processing times), not in the content of the files: cases misqualified at the margin, those sent back, those that go back through a human for no reason. | Make the agent reliable where it gets things wrong before extending it to other types of cover, and measure the time handlers really get back. |
| Customer relations and contact centres | Assistants that handle or prepare policyholder requests (emails, tickets, messages) and that equip advisors in real time: suggested reply, policy context, next step. | What policyholders go through, beyond the volume handled. Conversations classified into success, frustration and abandonment, repeated requests left unanswered and the reasons that recur show where the assistant helps and where it makes people lose patience. | Fix the reasons where the assistant fails, industrialise those where it succeeds, and hand business teams the requests nobody had anticipated. |
| Sales and distribution | Assistants for the networks (salaried advisors, tied agents, brokers) and for online journeys: advice support, cover comparison, quotes, routing to the right product or the right contact. | How users really move through the journeys. Routings unrelated to the stated need, steps users work around, points where the conversation stops. These are often inherited process rules that the agent applies faithfully and that nobody has reviewed. | Fix the process before fixing the agent. |
| Head-office functions | Generalist copilots and assistants, such as Microsoft 365 Copilot, ChatGPT or Claude, rolled out to legal, HR, finance, procurement and compliance, often seat by seat and with no defined scope. | Seats used against seats paid for, heavy usage nobody scoped, usage that would justify a dedicated assistant rather than a generalist tool, and unserved demand: what teams look for elsewhere because the internal tool does not offer it. | Right-size at renewal, industrialise the use cases that pay off, know which teams to train, and on what. |
| IT and development teams | Coding assistants for the teams that keep legacy systems and complex regulatory processing chains running, and, more and more, the shared instruction libraries that guide agents. | Adoption per team, the seats that sit idle and the quality of shared instructions (duplicated, contradictory or never called). | Rationalise the shared library, spread the practices of the teams that move fastest, right-size the seats. |
How to measure what AI returns for an insurer
The return on investment (ROI) of an insurer's AI is measured by combining two sources: the usage logs its platforms already produce, and interviews with its leadership teams. Idun Group then sets the cost of each use case against its observed usage, and what the agents do against what happens after them: rework, processing times, conversations that fail. Quality is measured the same way as ROI. Idun Group carries out the audit within 15 days, with at least twelve interviews.
- 01
The logs and exports you already produce
We start from what the platforms you administer produce: internal models, copilots, business assistants, custom agents. For AI agents that prepare files, we match their logs against an export from your case management system (statuses, reassignments, processing times), without the content of the files. We install nothing on workstations, do not instrument your platforms and write no code on your side. Your teams hand us the exports or open a read-only connection for us; we settle which in the scoping meeting.
- 02
What enters the agents' context
We analyse what is sent to the agents as well as what they answer. That reading makes context hygiene possible: serving the agents less data, with answers just as good and less exposure. The risk function and compliance want to know before any extension.
- 03
Interviews, for what the logs do not say
The logs of AI tools say which teams use which tools, how often, at what cost, and which conversations fail or stop. They do not say what people meant to do. Interviews with the IT department and the business units are mandatory: interviews and data, never one in place of the other.
- 04
Entity by entity, then consolidated
We run the exercise entity by entity and product by product, then consolidate. Use cases built several times over surface at this stage, and with them the quickest convergences.
The measurement loop
Three deliverables
The map: usage
What teams do with AI: detected use cases, real adoption per entity and per contact centre, assistants and instructions created, published or kept private.
The account: ROI
What each use case costs and what it returns: paid seats against observed usage, the rework sent back downstream, and the time given back, converted into cost at a rate set with you from your job profiles.
The course: strategy
What to industrialise first, what to cut, which teams to train and on what, over a horizon of six months to two years. And what your teams have to tool up: publish the assets kept private, plug AI into the management systems where the work gets done.
From pilot to scale, three outcomes
The decision to take an AI pilot to full-scale rollout rests on usage signals, not on the impression the pilot left. Once usage is measured, the decision comes down to three outcomes, each tied to a budget action: industrialise a broad, repeated usage that sends little rework downstream, consolidate a use case re-implemented by several entities, cut a marginal usage whose rework exceeds the time it gives back.
| Decision | Observed signal | Action |
|---|---|---|
| Industrialise | Broad and repeated usage, little downstream rework, a visible gain in processing time | Extend to other types of cover or other entities, with a dedicated budget |
| Consolidate | The same use case re-implemented by several entities or on several products | Choose one implementation, align the rules, stop the duplicates |
| Cut | Marginal usage, paid seats unused, rework greater than the time given back | Terminate, or reallocate to a better-measured use case |
What the analysis shows, and what it does not
Idun Group analyses usage, never people or case files: results are aggregated by team, by tool and by entity. Idun Group validates the exact status of the processing with your DPO. The analysis is read-only: nothing is written into your systems, no workstations are scanned. Processing is hosted in France, at Scaleway in Paris, or on your premises. The contractual framework (data processing agreement, data retention, your obligations under DORA) is settled in a meeting with your CISO.
What the data shows
- aggregated usage, by team, by tool and by entity
- cost set against observed usage
- the rework agents send back downstream, read in the statuses and processing times of the management system
- what enters the agents' context
- unserved demand: what teams look for without finding it
What it does not show
- the activity of a named employee
- the content of case files and policyholder data
- pricing, underwriting, actuarial models, claims decisions
- what the exports do not contain: the report says so in its appendix of unknowns
The kind of situation the analysis handles
These three situations are representative of what usage analysis brings to light in insurance. They are illustrative: they describe no particular client and carry no figures.
"The triage agent was wrong at the margin. Nobody saw it."
An agent has been routing claim notifications for months, and its results look good. Set against what handlers then do with the files, its logs show that a share of notifications is misqualified, one way or the other: handlers rework them, or the urgency of the file is misjudged. None of this showed up in the agent's indicators. An agent's performance is read in what happens after it.
"Reading what policyholders go through, not just what they ask for."
An assistant handles incoming requests. The response rate is high; satisfaction is not measured. Classifying real conversations into success, frustration and abandonment, then analysing the reasons that recur, brings out the situations where the assistant answers without resolving anything, and those where the policyholder gives up.
"Data that had no business being in the context."
Analysing what is sent to the agents shows personal data circulating there for no reason. It does not serve the answer, it weighs down the context and it creates an exposure nobody had measured. The finding leads to cleaning up what goes into the agents: less data, answers just as good, reduced exposure.
What we have already measured
Idun Group is an independent consulting firm. That independence gives the analysis its value: we have to be able to conclude against our own interests. None of Idun Group's publishable references is an insurer.
Consulting and training firm
At a consulting and training firm, the analysis of real Claude usage established 85% real activation across around forty seats, false starts included. With the seats that left no trace at all, 31% of seats, nearly one in three, were to be reassigned or activated. At the firm, the three heaviest users' consumption is worth, at API prices, 12 to 25 times the price of a premium seat.
The limit The limit, written at the top and never in an appendix: the platform export had a rolling thirty-day retention while the rollout had been running for five months. One month out of five was measured.
The sector is not insurance. What carries over is the reading: active seats against paid seats, the seats that leave no trace.
What leadership teams ask us
What are the uses of AI in insurance?
In insurance, the AI that teams use is concentrated in five territories: customer relations in contact centres, claims handling, advice support in distribution networks, copilots in head-office functions and coding assistants in IT teams. According to EIOPA, 40% of the insurance undertakings surveyed already use generative AI in customer relations, 32% in claims handling and 22% in sales and distribution. Pricing, underwriting, risk selection, actuarial models, fraud detection and claims decisions also use AI, but fall outside Idun Group's scope.
How far have insurers gone with generative AI?
According to the EIOPA survey published in February 2026, covering 347 insurance undertakings in 25 countries, 65% of the undertakings surveyed already use generative AI. Of the use cases they report, 64% are still at pilot stage and 32% are in production.
How do you measure the ROI of AI in insurance?
The ROI of AI in insurance is measured on observed usage, not only on the time savings teams declare: platform usage logs are combined with interviews with leadership teams. The cost of each use case is set against its observed usage, then against what happens after it: time given back to handlers, rework, processing times. A use case whose rework exceeds the time it gives back should be cut. Idun Group puts no figure on the gain in advance: it is measured in your organisation.
How do you move from a pilot to a full-scale rollout?
An AI pilot moves to scale on usage signals, not on the impression it leaves. At an insurer, a pilot deserves to be extended when its usage is broad and repeated, it sends little work downstream and processing times fall. Without those signals, you fix or cut before extending.
How do we know whether our claims agents triage correctly?
A claims triage agent is judged by comparing the routing it gives with what the handler then does with the file, using an export from the claims management system (statuses, reassignments, processing times), without reading the content of the files. That gap reveals the cases misqualified at the margin and the share of work the agent sends back without anyone noticing.
How do you measure real AI adoption at head office?
Real AI adoption at head office is measured by seats used against seats paid for, and by retention over time, per team and per tool: Microsoft 365 Copilot, ChatGPT, Claude. Across the estates Idun Group analyses, 20 to 40% of paid AI seats go unused: an engagement finding, not a study.
What data enters our agents' context?
To find out what data enters an AI agent's context, Idun Group analyses what is sent to the agent as well as what it answers. This analysis brings out the data served for no reason. The findings lead to recommendations on context hygiene and pseudonymisation.
Our vendors already provide dashboards. What do you add?
A vendor dashboard cannot see outside its own tool. It misses the usage that has moved to other tools. Idun Group brings your tools together in a single view, to compare the same use case from one tool to another, and conducts the interviews that explain the gaps.
Do you process policyholder data or health data?
Idun Group reads neither the content of case files nor policyholder databases. The logs of AI tools can still contain what teams pasted into them, personal data and sometimes health data included, and the context analysis flags it among its findings. When the analysis detects personal data, it stops at the finding; what happens next is decided with you and your DPO (scope, purpose, informing the people concerned). Results are aggregated by team and by tool, and processing is hosted in France, at Scaleway in Paris, or on your premises.
Do your agents still do what everyone believes?
Thirty minutes to choose where to start (a claims triage agent, a contact centre, head-office seats), the platforms to read and what you expect from the read-out. You leave with our view on whether the exercise makes sense for you.
Read-only. We analyse usage, never people.
Supported by
