AI in insurance: the use cases that hold up in production

In insurance, artificial intelligence no longer stumbles on technology. It stumbles on two questions the sector handles badly. Where to place the first flow, and how to prove what it returns. The pilot phase is over, and the difficulty has changed in nature.

Tandem works with more than 40 companies on AI applied to operations, including five assignments in insurance and brokerage. This article gives the grid we use on those assignments. It ranks eleven projects by value, feasibility and regulatory risk, and its result often displeases the executive committee.

Where does AI actually stand in insurance?

The figure that matters is not the adoption rate, it is the measurement rate. The Datos Insights study "AI Implementations in Insurance: The Pilot Phase Is Over" (2026) gives the figure. 62 % of surveyed insurers have AI in production, against 37 % a year earlier. But only 22 % measure a positive return on investment. Read those numbers with care, they describe the American market, which is markedly more advanced than the French one.

Europe is a notch behind. McKinsey surveyed more than fifty executives at the main European insurance groups. One in three reports first generative AI use cases in production, and 20 % rate their maturity as advanced. France therefore still has the window the American market has just closed.

• Where AI is already deployed. IT leads at 62 %, then underwriting at 57 % and claims at 50 %. Distribution stays at 24 %, product development at 14 %.
• What actually blocks it. Talent for 41 % of respondents, data quality for 38 %, legacy integration for 31 % against 16 % a year earlier. Compliance and budget come last at 17 %.

In France, AI is the leading transformation driver for insurance jobs. That is the finding of the forward-looking barometer of the Observatory for Insurance Careers, published by France Assureurs in June 2026. The observatory places usage on data analysis, fraud detection and task automation. It adds a nuance teams remember, AI reshapes the content of jobs more than it removes them.

Two announcements from the summer of 2026 show the sector moving on two distinct fronts. In July 2026 AXA announced a large-scale rollout of Microsoft 365 Copilot to its staff, three years after launching its internal Secure GPT service. That move is about internal productivity.

MAIF went the other way. In June 2026 the mutual insurer opened an app inside ChatGPT. It rates a home's exposure to climate risk from an address, and estimates the price of bicycle cover (La Tribune de l'Assurance). That is not productivity, it is distribution. The insurer meets the prospect where the question is asked.

Augmentation or automation, two projects and two calendars

Every use case belongs to one or the other, and confusing them is the most frequent cause of overrun. In augmentation, AI prepares and the human decides. The handler stays in charge of the file, nothing goes out without validation. In automation, the flow runs and the team handles only the exception.

Which AI projects should insurance open first?

Three scores from 1 to 5 are enough to decide. Value, which adds addressed capacity and value beyond time. Feasibility, which measures data access and whether rules are written down. Regulatory risk, which runs from internal data with no customer impact to a decision affecting the customer. The index is value times feasibility, divided by risk.

The result displeases, and that is what makes it useful. The topic most discussed in executive committees, the customer-facing voice agent, comes tenth out of eleven. The one nobody talks about, tidying up the document corpus, comes first and conditions half the others.

1. Corpus and internal knowledge, index 16. A single, up-to-date and versioned reference base that every later case relies on.
2. Accounts payable and expense claims, index 15. No customer impact, written rules, maximum feasibility.
3. Internal control and compliance, index 12. Internal data, immediate coverage gain.
4. Cancellations, endorsements, renewals, index 10. The highest value in the table, legal deadlines, written rules.
5. Inbound mail and post-call work, index 8. Sorting the generic inbox and writing the call summary, with no reputational risk.

This hierarchy is not specific to insurance. We apply the same grid in accounting firms, where document flow also comes first. See our analysis of seven AI use cases for accountants, which shares the legal deadline and supporting document constraints.

Fraud detection deserves a separate mention. The French insurance fraud agency ALFA recorded 946 million euros of fraud detected by its members in 2025. Fire, accident and miscellaneous risk lines account for 627 million of that on their own. The topic is large, but it ranks ninth, because its feasibility depends on portfolio quality rather than on technology.

How to size a flow before committing budget

If only 22 % of insurers in production can say what their AI returns, it is not because the gain is absent. It is because nobody measured the starting point, and it cannot be reconstructed afterwards. Five variables are enough, and one week of sampling gives them all.

Three corrections make this calculation honest, and they are almost always the ones missing. The first is about the automation rate. Only count the items that pass quality control without rework. On a heterogeneous document flow, the observed starting point runs from 40 to 60 %. On a standardised flow with clean reference data, from 60 to 80 %. Above 85 %, ask for proof.

The second correction is the one finance departments wait for. Freed capacity is not cash. Part of it is redeployed onto other files, which is real value but invisible in the income statement. Only the other part becomes an avoided cost. In the example above, out of 212,160 euros of addressed capacity, roughly 64,000 euros is genuinely avoided cost.

The third correction adds the value beyond time. Processing time, error rate, call answer rate, non-compliance penalties avoided, premiums not lost. On contractual flows and on fraud, that share frequently exceeds the productivity gain. It is also the most fragile, because it only materialises if somebody picks up the phone.

What five brokerage assignments taught us

Tandem publishes the sector, the volumes, the scope sold and the sequence followed. Not our clients' figures, which they have not authorised us to circulate, and which would be useless to you outside their volume context. The assignments below are therefore anonymised.

A wholesale health and protection broker, 51,000 cancellations a year

Processing was entirely manual. At twelve minutes per request, that amounted to the equivalent of six full-time roles spent copying information from a PDF into a policy administration tool. The structuring constraint came from health data, which required hosting on the client's own infrastructure and ruled out any solution calling a third-party interface outside the European Union.

The transferable lesson is counter-intuitive. The difficulty was not reading scanned documents, which current models handle well. It was matching them against the policy reference base, on a portfolio containing namesakes and duplicates. A match rate of 85 % leaves one file in seven for manual rework, which is enough to cancel out the gain.

A property and personal lines broker, two assignments in two months

The sequence started with a short audit of customer support, two and a half days. Call volumes, reasons, handling times, answer rate. The build came two months later, in five days. A first-version voice agent qualifies inbound calls, answers at level 1 and takes a structured message.

A specialist broker in international student insurance

Three and a half days of scope, live within a quarter, on a voice agent plugged into a highly seasonal and highly repetitive flow. Questions about certificates, cover and pre-authorisation, concentrated over two months of the year. This is the typical case where human staffing is structurally impossible.

The lesson holds for the whole sector. A seasonal flow is an excellent first ground, because the cost of the manual solution is already known and quantified. Temporary staff and call overflow show up on an invoice, which makes the comparison immediate and the calculation hard to dispute.

Two further assignments cover flows structurally identical to insurance ones. The first is an audit of third-party payment and customer service, at a network of more than twenty health centres. The second is a voice agent paired with a written customer service, at a banking group subsidiary. The document corpus was built there first.

The voice agent, tenth in the ranking and first in requests

Three of our five insurance assignments involve a voice agent, while the grid ranks it tenth. The contradiction is only apparent, and it is worth explaining. This project scores 4 out of 5 on regulatory risk. The reason is article 50 of the EU AI Act. Since 2 August 2026 it requires telling a person they are addressing an AI system and not an advisor.

That obligation blocks nothing, but it is designed into the opening greeting from the start. Added after go-live, it forces a rewrite of the whole script. The project therefore stays in wave 3 until a higher-value, lower-risk flow has shipped.

The mechanics come down to four components. A telecom operator carries the line, a transcription turns voice into text, a model generates the answer, speech synthesis makes it audible. Each component is configured separately, which keeps cost and voice under control.

The trade-off that decides a voice agent's success is not about technology, it is about scope. I have found that an agent loaded with too many scenarios gets lost among them. Three or four well-written resolution paths, plus a human transfer for everything else, beat any attempt to cover every case. In my own tests, handled calls last under a minute on average.

Outbound calls are the forgotten side of the topic, and the one that speaks loudest to insurance. An annual renewal to confirm, an unpaid premium to chase, a missing document to request. Or a prospect who just filled in a form and needs a call back within the minute. Nobody has time to do that flow by hand, so it is net gain rather than displaced time.

Steering runs on transcripts. Every call leaves a full text, a duration and a category, which makes it possible to alert teams on unexpected categories. The metric to follow is the share of requests closed without human intervention, corrected for the rework rate. An agent that handles 70 % of calls but where 30 % call back the next day really handles 49 %. For the complete method, see our AI voice agent guide.

What the AI Act and the ACPR change for an insurer

This is the part AI projects in insurance underestimate most, and the one that costs most when it arrives at the end. Brought in at scoping, compliance shapes the design. Brought in at the end, it is a veto after the budget has been committed.

The ACPR will be the market surveillance authority on high-risk financial use cases, from 2 December 2027. It is not waiting for the deadline to move. In July 2026 it published a discussion paper on algorithmic fairness in the financial sector. The text is open to public consultation until 30 September 2026, and guidance is expected by the end of the year.

One technical point in that paper deserves the attention of any team building a pricing model. Advanced systems are increasingly able to reconstruct sensitive personal characteristics, even when those variables have been removed from the modelling. Dropping the field is therefore no longer enough to demonstrate the absence of discriminatory treatment.

One last point often escapes business teams, because it targets the insurer as a risk carrier rather than as a user. AI-related risk is emerging from the grey area in professional indemnity policies. What was neither covered nor excluded is starting to be explicitly excluded. That is as much an underwriting topic as an operational one.

What AI cannot do in insurance

Knowing these limits avoids committing budget to a use case that cannot succeed. None of them is a temporary difficulty.

• Absorb exception cases. In a real flow, 15 to 25 % of files fall outside the frame. Illegible document, atypical situation, customer in dispute. Those cases carry most of the risk and will stay human.
• Repair inconsistent reference data. If the same product carries three labels and two cover tables depending on the source, no model will restore the truth. Tidying up is a prerequisite project, not an AI deliverable.
• Run a flow whose rules are not written. When the rule lives in the heads of two handlers, writing it down comes before coding.
• Replace the duty of advice. The distribution directive requires personalised, reasoned advice carried by an identified person. An assistant can prepare and document it, it cannot be its responsible author.
• Price on health data where that is forbidden. On a solidarity contract, health status cannot set the premium. A model that indirectly reconstructs that information through correlated variables exposes the organisation, even without intent.

Should you build your AI agent or buy it?

In 2023, everyone wanted to build. Since then off-the-shelf solutions have multiplied, technical debt has become frightening and the return on internal projects often disappoints.

I laid out this trade-off in a LinkedIn post on the build or buy choice for AI agents. The rule I take from it is simple. Buy when your processes look like everyone else's, build when the topic is strategic. Buy to move fast, build to differentiate.

Insurance has one feature that shifts the build or buy dial. As soon as health data enters the flow, the hosting constraint rules out part of the off-the-shelf market straight away. On the health brokerage assignment described above, the decision to build came from an obligation on the data, not from a technical preference.

On tooling, we build these workflows on n8n, which installs on the client's infrastructure when sovereignty requires it. See our n8n guide and our n8n agency offer.

How to choose an AI agency specialised in insurance?

The AI provider market is crowded, and most of them have never seen a cancellation letter or a delegation mandate. Five questions sort them quickly.

• Do they understand the insurance chain? A provider who confuses insurer, delegated administrator and broker will make routing errors.
• Can they deploy in a sovereign environment? Ask for a self-hosted workflow reference, not an intention.
• Where do they put the regulatory decision? If the answer is in the prompt, move on to the next one.
• What does the system do when it is unsure? An explicit handover to a human is a sign of seriousness, an always-confident answer is a risk.
• When does the first deliverable land? On our assignments, the first case reaches production between eight and twelve weeks after the audit starts. An audit that does not lead to a build within six weeks never will.

Tandem works across the whole chain, from claim declaration to broker reporting. The detail by role is on our AI in insurance and brokerage page. The scoping approach is described on our AI audit page.

Which project should insurance start with?

Start with the document corpus and the contractual flow. Those are the top two lines of the ranking, and that is no accident. The corpus conditions every later case. Cancellations carry the highest value in the table, with rules already written. The legal deadline makes automation almost mandatory beyond a certain volume.

Keep life and health pricing for later, and prepare it seriously. The 2 December 2027 deadline is coming, the ACPR consultation is open, and an undocumented pricing model will be a liability. Measure the volume, unit time and error rate of your three heaviest flows before any arbitration. One week of sampling is enough.

One last rule, valid in every case. If your system cannot say that it is unsure, it is not ready for insurance.