Beyond POCs: How forward deployed engineering is key to building AI for complex claims

There’s no shortage of ambition around AI in the P&C (re)insurance market. Almost every carrier is exploring opportunities. Some have multiple proofs of concept underway, some are investing heavily in internal AI development programmes, while others are still hesitating over issues around governance, security and operational questions.

A 2026 survey found that just 7% of respondents said they have AI fully embedded across claims workflows, while 38% said they were exploring (pilots and early trials not yet tied to core workflows) and 23% said they were using AI in a limited range of live-use cases (but not consistently end-to-end).

So, the appetite is certainly there. The challenge is how to move beyond AI experimentation to start generating meaningful operational value at scale. For many organisations, that transition is proving harder than expected. The issue is not the technology itself. AI is already capable of extraordinary things. The difficulty lies in making it work inside a live claims operation. Complex claims environments are highly specialised. Every carrier has its own processes, priorities, terminologies and ways of working. Different classes of business require different insights, and different teams need information presented in different ways. Given all this, it’s hardly surprising that claims teams often find themselves trapped in POC purgatory, running pilot after pilot without making that leap to production.

AI initiatives can often struggle to move beyond the proof-of-concept stage because they have started with the theoretical capabilities of the technology, rather than with the complex reality of the problems or opportunities they’re intended to address. An AI model may perform impressively in isolation, but less so in a production environment where they encounter unanticipated challenges around working with legacy systems, making sense of complex and often inconsistent claims data, supporting established workflows, satisfying governance requirements, and adequately reflecting the day-to-day realities of how claims professionals work in a complex claims environment.

Proofs of concept are often built around carefully selected datasets and narrowly defined use cases. They can demonstrate that an AI model is capable of performing a task, without testing whether it can cope with the volume, variability and operational complexity of a live claims environment, where data quality, business rules, governance and human workflows all interact.

Bringing AI development in-house can make sense for some organisations, but many carriers who take this route subsequently discover that building effective AI for complex claims requires not just technical expertise, but also specialist experience in applying AI within highly regulated, operationally complex environments. You shouldn’t have to become an AI company in order to benefit from AI.

When it comes to developing AI solutions that deliver real value, technology expertise and resource really do matter, but so too does understanding enough about a specific operational environment to build something that genuinely improves the way people work.

This is where Forward Deployed Engineering (FDE) comes in. Rather than developing products in isolation and presenting them to customers once complete, this approach sees engineers working directly alongside operational teams throughout the design process. The approach has been popularised by organisations such as Palantir, OpenAI and Anthropic, particularly where problems are complex, workflows are highly specialised, and success depends on deep operational understanding rather than technology alone.

The model is particularly applicable to complex claims. Every carrier has evolved its own operating model, governance, terminology, data landscape, and claims philosophy. Even organisations writing similar classes of business often reach decisions differently or organise information in different ways. Much of that knowledge exists in experienced practitioners rather than in documentation or process maps.

FDE is designed to surface that operational knowledge early. Engineers spend time understanding how work is actually performed, where information originates, how decisions are made and where unnecessary friction exists. Rather than attempting to solve every problem at once, development focuses on specific operational challenges that can be explored, tested and refined in collaboration with the people carrying out the work.

Where a proof of concept is primarily designed to demonstrate that a technology can work, FDE is designed to discover how it needs to evolve in order to work reliably within a specific operational environment. In that sense, the technology and the operational understanding mature together.

This changes the nature of the development process. Instead of lengthy specification documents followed by extended development cycles, feedback becomes continuous. Ideas can be evaluated quickly, discarded where necessary, or refined as operational understanding develops. The emphasis shifts from delivering predefined functionality towards solving clearly understood problems.

The objective extends beyond task automation. Across many claims operations, experienced professionals still spend a lot of time locating information, reviewing repetitive documentation, assembling context from multiple systems or manually identifying emerging patterns. These activities are clearly necessary, but they consume attention that could otherwise be devoted to investigation, judgement and decision-making. In that sense, one of AI’s most valuable contributions is protecting the attention of claims professionals by reducing unnecessary cognitive effort.

In practice, FDE typically begins with a period of close collaboration during which engineers work alongside claims teams to understand data sources, operational priorities and existing workflows. Early prototypes are developed and evaluated against live operational requirements, allowing the technology and the understanding of the problem to mature together.

Once there’s evidence that this approach is delivering measurable operational value, the focus shifts toward wider deployment and longer-term adoption. Rather than treating implementation as the end point for development, operational use becomes another source of learning, with further refinement driven by day-to-day experience.

This is a deliberate departure from traditional software procurement models. Instead of asking carriers to commit to a finished product based on assumptions made at the outset, carrier and technology provider work together to prove value in the carrier’s environment first. It’s a genuine co-investment model, with the provider providing significant engineering expertise upfront, and contributing the majority of the overall investment during the initial phase. This creates value for both organisations simultaneously. Carriers gain early access to AI capabilities and a direct influence on how they evolve. The provider gains a deeper understanding of the carrier’s operational challenges, allowing it to create stronger products for the wider market.

DOCOsoft has adopted this model as the basis for its AI development programme within the Lloyd’s and London Market. The intention is not in any way to replace carriers’ own expertise or encourage organisations to outsource AI strategy. It’s about combining specialist AI engineering with deep claims knowledge in a way that helps carriers move beyond experimentation to create practical, production-ready solutions firmly grounded in the realities of complex claims handling within a specific environment.