Frende Forsikring has launched an ambitious automation initiative aimed at transforming claims handling for travel and contents insurance. The project combines generative AI with robotic process automation (RPA) to create intelligent agents that streamline complex workflows and deliver faster, more accurate outcomes for customers.
The solution automates a wide range of tasks traditionally handled manually, including estimating claim amounts with deductions for age and deductibles, processing payments, generating and sending customer letters (such as confirmations, rejections, and informational updates), and gathering critical data like claim history, coverage details, and product information. For approved claims, the process is fully automated from start to finish, while rejections are validated through a “human-in-the-loop” approach to ensure fairness and compliance.
Currently, the system supports around 20 different outcomes for each insurance type, enabling nuanced decision-making and personalized communication. By leveraging generative AI, Frende ensures that customer interactions remain clear and empathetic, while RPA agents handle repetitive tasks with speed and precision. This combination reduces operational complexity and frees employees to focus on high-value activities.
The impact is already measurable: the automation corresponds to about a quarter of a full-time equivalent (FTE) annually, and the scope is expanding rapidly. As the project scales, Frende expects significant efficiency gains, improved customer satisfaction, and reduced turnaround times for claims. Beyond time savings, the initiative strengthens risk management and compliance by standardizing processes and minimizing human error.
This project reflects Frende’s commitment to being a highly data-driven organization. AI is embedded across the company’s operations, and this solution exemplifies how advanced technologies can deliver tangible benefits for both customers and employees.

Aura Guidance is an AI-powered onboarding companion developed by CCP Games as part of the EVE Evolved initiative. It is designed to help new players navigate EVE Online’s famously complex sandbox by providing contextual, data-informed answers directly within the game client.

EVE Online is widely recognized for its depth and freedom, but that depth can create significant onboarding friction for new players. Historically, rookies relied on volunteer player support in chat channels or external third-party resources to answer foundational questions. Over a nine-month period, players submitted more than 5.8 million messages in the in-game Rookie Help channel alone. This volume highlights both the strength of EVE’s community and the scale of the accessibility challenge.

Aura Guidance transforms this vast body of historical player support knowledge into a structured AI assistance system. Built using insights from millions of real player interactions, the tool surfaces relevant, context-aware answers during a new player’s early experience. Responses are informed by in-game context, such as location and ship type, increasing relevance and reducing confusion.

Importantly, Aura Guidance is not generative AI and does not create content, art, or gameplay systems. It is designed as a responsible, bounded assistance layer grounded in curated support data. When the system does not have sufficient confidence in a response, it redirects players to the Rookie Help channel, ensuring that community support remains central.

The feature launched as a controlled A/B test, allowing CCP to measure behavioral and retention impact between players with and without access to the system. This ensures the solution is evaluated based on measurable outcomes rather than perception.

Aura Guidance addresses a core business and design challenge: reducing early friction without simplifying the game’s depth. By making foundational knowledge more accessible, CCP strengthens player confidence, supports retention, and enhances the long-term sustainability of EVE Online’s complex ecosystem.

In 2025, Nextory launched Europe’s first dialogue-driven, agentic book recommendation system within a book streaming service. The “AI Librarian”, integrated directly into the Nextory app, transforms book discovery from static recommendation lists and keyword searches into interactive, conversational experiences. It combines a customer-facing agentic recommender with a production-grade synthetic metadata generation pipeline that continuously extracts and generates rich data from Nextory’s ever-growing catalogue.

The system is built on a multi-agent architecture paired with semantic retrieval, powered by continuously generated synthetic metadata derived from ebook and audiobook content. This enables the AI Librarian to understand books far beyond traditional metadata such as genre or author, capturing themes, tone, narrative structure, and audience suitability, delivering nuanced and highly personalized recommendations.

Motivation

Traditional recommender systems in streaming platforms rely primarily on collaborative filtering and historical behavior. While effective for popular titles, these approaches often limit discovery and struggle to capture nuanced reader intent, especially across large, multilingual product catalogues.

To address this challenge, Nextory developed a LLM-driven conversational recommendation system in-house that acts as an intelligent guide to the catalogue. Users can describe moods, themes, or reading goals in natural language and receive tailored, dialogue-based recommendations that exceed the capabilities of traditional recommendations and search.

Innovative solution and positive impact

A core objective was building a production-grade system for real-time user interactions across multiple markets and languages, requiring the project to address safety, reliability, and low-latency challenges that most agentic systems presented in 2025 did not encounter to the same degree, because they were deployed internally or designed to address non-real-time tasks. The solution to these challenges includes:

– Robust CI/CD pipeline for agentic systems
– Automated testing of LLM components, including evaluation using LLM-as-a-judge frameworks
– Continuous validation with both gold standard evaluation datasets and anonymized real user interactions and feedback
– Systematic monitoring and guardrails, ensuring safe, reliable, low-latency responses in a live consumer environment

In parallel, Nextory developed a continuous synthetic metadata pipeline that extracts a wide range of descriptors from book content. This data is used to power the AI librarian’s underlying semantic search. However, it also supports other use cases in Nextory, such as theme- and character-based collections for children’s books (see peer-reviewed publication referenced below).

The AI Librarian has been successfully deployed across 11 markets, demonstrating the scalability of conversational AI for multilingual content discovery. Across these markets, we observe high user engagement and read rates comparable to optimized search-based discovery. This is particularly notable because traditional search primarily serves cases where users already know exactly what they are looking for, while the AI Librarian supports more exploratory and nuanced requests that were previously difficult to serve, such as “fast-paced sci-fi horror with a strong female lead.”.

The project also laid the foundation for broad LLM-driven use cases, with a synthetic metadata pipeline serving as a strategic asset for discovery, personalized collections, and future innovations. The resulting pipeline is already feeding into other successfully deployed product discovery features, such as collections for popular characters and book themes for children. By combining conversational AI, large-scale semantic metadata, and robust system design, Nextory redefined book discovery and created a scalable framework for AI-driven innovation.

Ørsted and Amazon Web Services (AWS) teamed up to solve a practical engineering challenge in offshore wind: understanding the thermal conditions around undersea export cables that carry power from offshore wind farms to the onshore grid. These cables are critical, high-value assets. Their safe operating capacity depends on how efficiently heat from dissipates into the seabed. That makes accurate temperature prediction essential for reliability and long-term value.

Historically, thermal modelling relied on seabed property estimates (such as thermal resistivity) gathered early in a project’s design phase. These estimates are based on samples plus interpretation of ground modelling techniques. However, the models can never fully capture the reality, as complex effects from seasonal variations, complex material structures and disturbances from installation cannot be fully captured by modelling. The models deal with this uncertainty by making conservative assumptions, allowing a working model, but conservative, estimate to be made. That can lead to overly cautious loading limits and underused capacity.

To reduce this uncertainty, we worked with AWS to build a data-enhanced, physics-based modelling and optimisation framework. It combines real operational measurements with scalable cloud computing to improve how Ørsted estimates seabed thermal properties at a local level.

The solution brings together:

1. Operational datasets, including distributed cable temperature measurements, burial depth surveys, and environmental inputs.
2. A physics-based thermal model, TEEM (Thermal Electrical Equivalent Model), which simulates heat transfer by mapping thermal behaviour to electrical analogues.
3. Iterative optimisation methods (including gradient descent and Simultaneous Perturbation Stochastic Approximation, SPSA) that adjust uncertain parameters until model outputs align with measured temperatures.
4. Statistical techniques, such as bootstrap confidence analysis and temporal modelling, to validate results and distinguish genuine seabed variability from model artefacts.
5. Massively parallel processing on AWS infrastructure to run thousands of simulations efficiently, making high‑resolution calibration feasible in engineering timeframes.

With this approach, our engineering teams can move from broad, conservative seabed assumptions to more local, data-driven parameter estimates. That improves the accuracy of temperature predictions under different operating and environmental conditions. It also helps us understand where our assumptions can be updated with confidence, thus supporting safer operation, better utilisation, and more consistent performance across the cable network.

Cloud-based scaling on AWS made the work practical. Analyses that would previously have been difficult to run at the needed resolution could be completed quickly, supporting timely engineering decisions without disrupting operations.

The project also sets a clear direction for future development: a living, model-based capability “AI‑enabled cable digital twin” that can be recalibrated as new data arrives, strengthen anomaly detection, and inform next-generation export cable design and planning.

At its core, this project is an industrial AI system that learns from IoT‑style time‑series data to continually update a physics model of a critical asset. Overall, this initiative shows how data, AI and scalable computation can strengthen core engineering models and deliver tangible value in critical renewable energy infrastructure.

Foreign objects—such as rock bolts, drill rods, plastic poly-pipe, tree roots, vent bags, giant boulders, or other debris—pose serious risks in conveyor systems, particularly in mining and processing operations.
These items can:

Tear or rip conveyor belts
Damage or destroy rollers
Block or obstruct chutes and screens
Drive up maintenance expenses
Trigger expensive production halts from unscheduled downtime
Lead to serious safety hazards, including rocks or materials falling (“raining down”) from elevated belts

Rapid identification and removal of these foreign objects—regardless of their size or type—prevents major damage, minimizes costly interruptions, and enhances overall operational safety.

To address these risks, Boliden—a leading Swedish mining company—has developed an innovative AI-powered computer vision system that detects foreign objects on conveyor belts in real time.

The in-house-developed solution uses advanced cameras and AI algorithms to continuously monitor the belt, instantly identifying anomalies such as oversized boulders, tools, rock bolts, or other debris that could cause damage.
Upon detection, the system automatically triggers an emergency stop of the conveyor, preventing tears, roller failures, chute blockages, and potential safety incidents.

What sets this apart as a world-first in the mining industry is its direct integration: the AI doesn’t just alert operators—it issues commands that immediately act on the control system to halt operations, representing a pioneering step in autonomous, AI-driven process intervention for enhanced safety and reliability.
This breakthrough minimizes unscheduled downtime, reduces maintenance costs, and significantly improves overall operational safety in harsh mining environments.

Platform is leveraging best of breed LLMs from Anthropic and fit-for-purpose software products to deliver scalable solutions for all parts of Nordea.

Most organisations talk about AI readiness. Saab has built AI into a fighter aircraft.
Gripen is not a platform with AI bolted on as a future aspiration. It is an operational combat aircraft with AI running at the core of its mission systems today. That distinction matters enormously. Moving AI from controlled research environments into a safety-critical, real-time airborne system, one where the consequences of failure are absolute, requires a fundamentally different standard of rigour, architecture, and design philosophy. Saab has met that standard, and in doing so has created one of the most advanced and credible AI implementations in any safety-critical system anywhere in the world.
At the heart of the Gripen AI framework is a sensor fusion and decision-support pipeline of exceptional sophistication. The system continuously ingests data from radar, electronic warfare suites, electro-optical and infrared sensors, datalinks, and the wider battlespace network, synthesising this into a single, highly curated tactical picture in real time. Machine-learning-based classification, probabilistic reasoning, and rule-based tactical logic operate in combination to identify and prioritise threats, propose manoeuvre options, recommend weapon employment, and dynamically re-plan as situations evolve. All of this happens within the stringent timing and determinism constraints of air combat, on certified airborne computing hardware. This is not a proof of concept. It is a working system delivering measurable results.
What makes Saab’s approach genuinely distinctive is the deliberate choice to combine data-driven AI with explainable, deterministic logic and human-centred design. The Gripen AI does not attempt to replace the pilot. It acts as a cognitive partner: absorbing information overload, surfacing the most relevant insight, proposing courses of action with clear rationale, and keeping the human in authoritative control of every consequential decision. This is responsible AI in the most demanding sense of the term, not as a compliance posture, but as an operational necessity. A pilot who does not understand or trust what the AI is telling them will not act on it. Saab’s human-machine interaction design is built around that reality.
The technical architecture reflects the same clarity of thinking. A modular design separates perception, reasoning, and interaction layers, allowing different AI techniques, including machine learning, heuristic methods, and optimisation algorithms, to be composed, tested, and verified. This hybrid model is not a compromise; it is the architecture that makes real-world deployment possible. Upgrades and new AI capabilities can be introduced in service without re-architecting the underlying system, giving Gripen a genuine capacity to evolve throughout its operational lifecycle.
The impact is already visible. In evaluation and exercise scenarios, pilots report significantly reduced cognitive load and improved situational awareness in dense, multi-domain environments. Mission effectiveness improves through faster, higher-quality decisions, more efficient use of weapons and sensors, and better coordination with other assets. Crucially, the same AI building blocks that operate in the live aircraft are reused in ground-based mission support systems and high-fidelity simulation environments. This creates a continuous learning loop across training, development, and live test operations, compressing the cycle between operational insight and capability improvement.
By taking AI all the way from research into a certified, operational fighter aircraft, Saab has done something that very few organisations in the world have achieved. And because the underlying framework is modular, scalable, and platform-agnostic, Gripen represents more than a combat aircraft. It is a reference implementation for how AI can be responsibly, rigorously, and progressively embedded into safety-critical systems across air, land, and sea domains. That is the standard Saab is setting, and it is one that the broader data and AI community should know about.