This article is by Jawwad Rasheed, Financial Services Advisory Lead, Camunda
For decades, the insurance sector has been layering technology on top of operating models that were never designed for it. Polished front ends and digital channels have done a good job of disguising the legacy technology and processes beneath, where disconnected policy admin systems, claims platforms and underwriting workbenches are still doing the work that matters. But now, the move to AI is exposing the ceiling of that approach.
The efficiency promise of AI agents in claims and underwriting has captured boardroom attention, but simply adding them into existing operating models risks compounding complexity rather than reducing it. As a result, recent research finds just 11% of agentic AI projects in insurance reached production over the past year.
The real reason pilots stall
Many AI pilots do not fall short because of the technology. Instead, insurance processes, even within digitally native InsurTechs, are built around human execution rather than the collaborative human-agent workflows that agentic AI requires. In addition, compliance is often an afterthought at the proof-of-concept stage, when it should be shaping design decisions. And without enterprise-wide buy-in, pilots stay as side-of-desk experiments that meet the needs of a single team or function, with no obvious path to the rest of the business.
Insurance also faces some sector specific challenges. For example, claims and underwriting workflows carry deep institutional knowledge built up over decades, and the regulatory bar for model risk management is rising. There is greater scrutiny on model explainability and on the full model development lifecycle, covering design, backtesting, regression testing, ongoing governance review and model drift assessment. Accordingly, insurers will need to reevaluate their risk and control frameworks.
These challenges all point in the same direction for insurers. When agents are dropped into a fragmented environment in isolation, they cannot coordinate across workflows, and governance suffers as a result. And without careful consideration given to overall alignment, the result is increased complexity and cost. What is missing is the orchestration layer that holds the operating model together, embedding coordination, governance and control into processes rather than adding it in afterwards.
The transformation dilemma
With balance sheets and customers to protect, insurers cannot redesign their operations overnight. In practice, most are working through some combination of three approaches.
The pragmatic option is to layer and improve, introducing an orchestration layer that ties existing SaaS investments together and exposes more value from the technology insurers already own. Alternatively, rip-and-replace can be expensive and high-risk at scale, and is generally reserved for specific functions or new product lines where the legacy debt has become costly. As a result, the ecosystem and partnership route is emerging as a popular path for incumbents looking to combine their brand strength with InsurTech agility.
Across all three approaches, accessing an orchestration control plane becomes key. This enables insurers to activate more value from existing SaaS deployments, support a controlled migration away from legacy infrastructure where required, and makes InsurTech partnerships operationally viable for areas like onboarding, rather than just commercially attractive.
What re-engineering looks like
Adding the orchestration control plane in is only part of the picture. Every insurance process in use today was designed for a world without AI, which makes process redesign a prerequisite for agentic deployment. But too many insurers are instead reaching for agents before they have established how their end-to-end processes should run, where bottlenecks sit, and when human judgment is crucial.
Governance must be embedded from the start. Process logic determines when an agent is engaged, and where human review is required, while decision logic defines what can be auto accepted, referred or escalated. With those decisions embedded into the process itself, agents work within enforceable guardrails rather than in the grey areas that causes many pilots to stall.
Allianz’s Project Nemo is a good example of this approach. Launched in Australia in 2025 to handle low-value food spoilage claims during weather events, Nemo uses seven specialised agents covering planning, coverage verification, weather validation, fraud screening, payout calculation and audit, with a human making the final payout decision. Allianz has reported an 80% reduction in processing and settlement time for eligible claims under $500, and has announced plans to extend the system in a controlled way into other low-complexity, high-frequency use cases over time.
Looking ahead
In 10 years’ time we may see many major insurers strip back to their core strengths of brand, customer relationship and underwriting liability, handing off areas like claims management and elements of servicing to specialised partners.
For that model to work, composability and orchestration are imperative. As models, vendors and regulations evolve, insurers will need the flexibility to swap out AI components and the governance structures to onboard partners at speed. Without an orchestration layer that sits above the participants and enforces consistent control, the model simply cannot be operated safely at scale.
That is why the winners over the next decade will not be the insurers that deploy the most agents. Success will belong to insurers capable of re-engineering their operating model around genuine collaboration between humans, agents and systems, with orchestration at the centre of how that collaboration is governed.
Be the first to comment