The question was posed from the outset by the moderator, , partner advisory at PwC Luxembourg: One year on from the previous Insurance Day, when it was announced that AI would become a "decisive competitive enabler" in three years' time, is this prediction now materialising in insurance?

For , chairman of the Foyer group and vice-chairman of the Aca, there can be no doubt: AI is "a revolutionary and disruptive technology" for the sector. Even if he addresses insurers by paraphrasing Forrest Gump: "Run, Forrest, run", he immediately warns against the "wet dreams" surrounding AI, sometimes perceived as a magic solution to all problems. The urgency is there, but it does not obviate the need for lucid work on the limits, risks and real use cases.

This urgency is shared by , chief customer officer at LuxProvide. According to him, "it was already the right time a year ago, and it's even more the right time today." Models have matured, infrastructures have matured and companies have started to experiment. The risk now, he warns, is of "missing the train": in an environment where the learning curve is very steep, any delay today will be hard to make up.

At Liberty Mutual Insurance Europe, the tone is even clearer. For , general manager Europe, AI is no longer just a competitive advantage but "a question of survival" for business models. In the near future, he says, AI will be as commonplace as the internet or electricity: all players will be using it. Those who fail to adapt to it will simply be putting their survival at risk.

Investor Orna Carni, managing partner of the FinTLV Ventures fund, is nevertheless calling for us to run in the right direction. Referring to the Cheshire Cat, she reminds us that "if you don't know where you want to go, it doesn't matter which way you go". Translated to the level of companies, this remark is less anecdotal than it might seem: her conviction is that generative AI needs to be part of a clear strategy, with explicit priorities, rather than sprinkling tools over existing processes.

On the ground, the first use cases are still largely focused on productivity and operational efficiency, rather than product transformation. Marc Lauer estimates that, in the long term, "75% of non-life insurance claims should be able to be managed without human intervention", with AI making it possible to automate a large part of the processing. But he acknowledges that, for the time being, companies are mainly inserting AI in bits and pieces into legacy processes, without rethinking them in depth. The risk, he points out, is then "taking the old process, adding a little AI in places, without redesigning the end-to-end chain".

Deployment in local mode

Filipe Pais, for his part, mentions projects carried out with Luxembourg financial institutions, in both banking and insurance. Most of these involve 'small areas' of activity: assistance for software development teams, automatic classification and faster response to IT support tickets, and deployment of internal chatbots for specific processes. The aim is to generate rapid productivity gains, rather than launch disruptive models. Data sovereignty constraints are also pushing many players to deploy these solutions on local infrastructures, which LuxProvide is supporting.

At Liberty Mutual, investments are becoming more structural. Christian Rola describes the creation of an internal "AI factory", supported by the group's engineering teams, and the development of a "submission intelligence" tool for policyholders. In concrete terms, exchanges with brokers and files sent in PDF or other formats are ingested by the platform, which builds a risk profile in a matter of seconds and notifies intermediaries of any missing data. Where an underwriter could spend days on a complex file, AI drastically compresses timescales and frees up time for higher value-added tasks, particularly relations with brokers.

Warning signs, preventive insurance, automation

Generative AI is also invading customer relations and prevention. Marc Lauer stresses the potential of predictive models to identify dissatisfied customers before they cancel their policies. Policyholders do not leave their insurer "on a whim one morning": they have accumulated frustrations. By finely exploiting the available signals, a model could identify some of these situations, enable proactive intervention by the insurer and, if only half of these customers were ultimately retained, generate a substantial portfolio gain.

Orna Carni extends this reasoning to the transformation of the business: AI and the proliferation of data - in-car telephony, connected objects, sensors - are opening the way to more preventive insurance, capable of providing early warning, helping to reduce exposure to risk, while remaining present to compensate when the claim occurs. She points out that on the employee side, this tooling is also seen as an opportunity: by offloading repetitive tasks, AI enables teams to concentrate on complex claims or underwriting.

Innovation, however, is not just about productivity. Christian Rola anticipates a future where "the insurer's AI agent will talk to the broker's AI agent" to handle simple risks and standardised flows from end to end. In this scenario, the AI becomes an additional member of the team, responsible for routine tasks, while the human employees focus on negotiation, expertise and advice.

On the start-up market, observes Orna Carni, most insurtechs have incorporated an AI brick into their value proposition, sometimes just for show. The role of an investor is to distinguish between technologies that are truly differentiating and those that are opportunistic, starting with a simple question: "Is there a real business problem behind them? In her view, AI only creates sustainable value when it tackles concrete frictions in the value chain, rather than remaining a marketing gimmick.

Data quality, the first imperative

The speakers converge on one point, however: without quality data, AI produces nothing useful. "Garbage in, garbage out", essentially sums up Marc Lauer, for whom the first building block is the organisation, inventory and "ontology" of internal data. He also believes that the Luxembourg regulatory framework is out of step with technological ambitions. He denounces the rules on professional secrecy "from the last century", designed for a world of "fortified castles" where data is locked behind walls, whereas insurers today want to process and enhance it securely.

Christian Rola points to the same paradox: the insurance industry has "a huge pile of data", probably its most valuable strategic asset, but projects get stuck when it comes to training models on sensitive data. He argues for regulatory changes that allow more supervised experimentation, with anonymisation or federation techniques, rather than a straitjacket that discourages innovation from the outset.

Filipe Pais points out, however, that technical solutions exist to reconcile data protection and model training, whether in the case of synthetic data or federated learning. In some cases, the data remains with the customer, with LuxProvide providing the computing capacity via dedicated connections; in others, it can be temporarily transferred to supercomputing infrastructures, while remaining within a framework of sovereignty. He also mentions the setting up in Luxembourg of the "AI Factory" aimed at pooling training, infrastructure, regulatory support and consultancy, to make access to AI easier for businesses.

Governance and explicability, levers for success

On the governance front, Liberty Mutual has set up a "responsible AI" committee, responsible for defining the principles governing the use of AI - respect for ethics, fairness, compliance - and for supervising projects, whether they involve the use of co-pilots or in-house developments. A supervisory body monitors progress, priorities and compliance with data protection, avoiding the sometimes tempting 'shortcuts' to speed up production. Christian Rola also stresses the need to train boards of directors, who bear the ultimate responsibility for the use of AI within the company.

Marc Lauer adopts a more critical tone with regard to European regulation, which he deems quick to "regulate what does not yet exist", in reference to the AI Act. He believes that before any written corpus, it is the values of companies that should play the role of first line of defence, in terms of protecting customers, employees and shareholders. He warns against the temptation to "over-frame" uses that are still in the exploratory phase too early.

The question of explicability, central to a sector founded on trust, is also being debated. How do you explain a pricing model or risk selection decision to a policyholder, a broker or a regulator? At Liberty Mutual, comparative tests are carried out in parallel between 'manual' teams and teams using AI tools, to check that the conclusions are comparable in terms of accuracy, while gaining in speed. The exercise remains tricky, admits Christian Rola, especially with those who do not spontaneously trust technology.

Talent, a model to be found

Marc Lauer points out that, in a way, "models are not new to insurance". Car and home insurance rates have always been based on statistics, first deterministic, then stochastic; AI is just a new generation of tools. As for the requirement for explanability, this is an old one: when presented with a proposed tariff involving a 50% increase, he would ask for explanations before approving it. In his view, the culture of questioning models is part of the DNA of the business and must remain so, even in an environment where AI is accelerating and automating.

And finally, there is the human dimension. Christian Rola describes a generational divide in the perception of AI: baby-boomers and part of generation X, often in management positions, remain the least convinced; subsequent generations are the biggest consumers of tools but have few levers to steer strategy; and generation Z sometimes fears for its job. The key, he believes, is to treat AI as a lever for quality of life at work, linked to the famous "work-life balance", rather than as a threat of substitution.

Marc Lauer identifies two challenges: attracting and retaining talent specialising in data or AI - which presupposes competitive working environments, tools and remuneration packages - and reassuring existing employees. In the short term, it is a question of not sending out anxiety-provoking signals, for example by promising to automate a given percentage of claims and announcing the loss of posts in proportion. In the longer term, he is betting that, as with the steam engine, the car or the internet, AI will improve living standards overall, provided that its benefits are shared across society.