The Claims Spectrum: Why Not All Claims Are Equal
Authors: Chris Brown - The Build Paradox , Mike Daly - Insurtech World
Published: 17/02/2026
My first question is always: "Which line of business?" Why?
Claims aren't one thing. The spectrum from gadget replacement to cyber incident involves fundamentally different levels of complexity, judgment, data availability, and litigation exposure. An AI strategy that makes sense for one end of that spectrum could be catastrophic at the other.
Getting this distinction right is foundational. Get it wrong, and you're either under-investing in automation where it's safe and valuable, or over-investing where it's dangerous. And there's a related trap worth flagging early: the assumption that AI is always the answer. Sometimes a well-designed deterministic workflow performs better, faster, and with far less governance overhead. The goal isn't maximum AI. It's the right tool for each problem Explained over the following sections:
- How claims get routed
- The spectrum of complexity
- How complexity impacts your AI strategies
- The portfolio strategy view
- The organisational realityThe uncomfortable implication
How Claims Get Routed
We need to know how claims routing works today and how customers and claims professionals will demand change in the future. This is the foundation on which any AI approach must be built.
Most modern automated claims journeys are orchestrated workflows: branching question sequences that classify the claim, validate basic coverage, and route to appropriate handlers or service providers. The system classifies the claim type early (e.g., a single vehicle low-speed bump in a car park, versus a multi-vehicle pile-up) and routes it accordingly.
Vulnerability indicators trigger immediate redirection to the call centre. Third-party involvement triggers capture workflows because insurers want to engage the third party quickly to manage costs. Ancillary claims, such as glass replacement, are entirely offloaded to specialist service providers. None of this requires AI. It's deterministic routing based on clear criteria.
The claim data flows into the core claims systems where it follows standard workflows with defined validation points, known inputs and outputs, and clear next steps. This is the environment in which any AI system must integrate, not replace across the whole spectrum of claims with differing claims process routing.
The Spectrum of Complexity
Mapping the claims spectrum, is foundational to any sensible AI strategy across key categories: -
- High volume, low complexity claims
- Medium Complexity claims
- High Complexity Claims
- Extreme complexity and emerging lines
High-volume, low-complexity lines:
Gadget insurance involves receipt, proof of ownership, and replacement value. Clear data, binary decisions, minimal judgment required. Litigation exposure is negligible.
Routine pet claims require a veterinary invoice, policy limits, and standard treatments. Slightly more variability but still largely formulaic.
Windscreen replacement requires a photo of damage, an approved supplier network, and fixed pricing. Almost entirely automatable and typically offloaded to third-party specialists anyway.
Basic travel involves flight delay documentation and standard compensation tiers. Data-driven, minimal judgment.
These claims share characteristics: clear data inputs, limited judgment requirements, established pricing, and minimal litigation exposure. The decisions are largely binary or formulaic.
Medium-complexity lines:
Standard motor without injury involves damage assessment, liability determination, and repair network management. More judgment is required. Was the vehicle roadworthy? Is the damage consistent with the reported incident? What's the fair repair cost?
Routine home claims, such as water escape or storm damage, require cause assessment, scope of loss determination, and contractor management. This needs interpretation. Gradual damage versus sudden event? Maintenance failure versus insured peril? A simple commercial property is like a home, but with business interruption considerations adding complexity.
These claims exhibit greater variability, require interpretation, and have meaningful consequences if decisions are incorrect. Customer complaints and FOS referrals become real possibilities. Third-party motor claims alone account for 70% of all motor insurance claims costs [1].
High-complexity lines:
Motor with injury isn't one category. Soft-tissue claims follow different protocols than serious injury claims, which differ from catastrophic or life-changing injury claims involving multi-year rehabilitation and care costs. The judgment required is substantial. The stakes are someone's quality of life. Litigation is common. The 2021 Whiplash reforms appear to have lowered minor injury claim costs, but rising care and legal costs continue to drive up serious injury claim costs [2].
Large commercial claims involve business interruption calculations, aggregation questions, and coverage disputes over policy interpretation. Significant sums and sophisticated policyholders who will challenge decisions. Professional indemnity requires liability assessment, notification requirements, and coverage trigger disputes. Legally complex and highly contestable.
These claims involve irreducible complexity. The judgment required has been developed over decades by specialists. Getting decisions wrong means significant litigation, regulatory censure, and reputational damage.
Extreme complexity and emerging lines:
This is where things get genuinely difficult. And where "extreme complexity" encompasses very different operational challenges that shouldn't be conflated.
Subsidence is slow, technical, and predictable in its unpredictability. It can be validated against geographical and geological data, including existing claims in the area, soil types (particularly clay content), weather patterns, and the property's age and construction. The British Geological Survey maintains national datasets mapping shrink-swell susceptibility [5]. Worst-case scenarios involve underpinning or rebuilding. Most subsidence claims take between one and two years to resolve,” and the process includes a monitoring stage [3]. The challenge is technical complexity and long-tail exposure, not speed. Average payouts currently stand at around £17,264, but underlying cases run substantially higher [4].
Cyber is the opposite in almost every dimension. The April 2025 attack on Marks & Spencer illustrates why cyber claims can't be handled like traditional property losses. The attackers had infiltrated M&S systems, impacting them over the Easter weekend. Contactless payments failing, click-and-collect down, and online shopping suspended for 46 days [6]. This wasn't just an M&S problem. The attack hit their supply chain, franchisees, and service providers. The Co-op was struck by the same threat actors using similar tactics. The Cyber Monitoring Centre classified it as a "Category 2 systemic event" with a combined financial impact of £270-440 million across both retailers [7].
The regulatory and criminal complexity compounds things. The National Crime Agency investigated. M&S's chairman testified before Parliament, calling for mandatory disclosure of material cyberattacks [8]. Social engineering through a third-party IT contractor provided the entry point. Ransomware deployment, data exfiltration, and double extortion tactics. These aren't scenarios with decades of precedent. M&S expects the attack to cost over £300 million in lost operating profit, offset partly by its £100 million cyber insurance coverage [9]. The attack erased £750 million from their market value within days [10].
Subsidence and cyber are both "complex", but that's where the similarity ends. They require different expertise, timescales, data sources, and decision frameworks. Your AI strategy needs to reflect that.
Beyond these, specialist lines like terrorism, marine, and catastrophe are outside the AI conversation for now. They're handled by experts dealing with genuinely novel situations where precedent is scarce. If there's a role for AI in these lines, it's more likely in proactive risk mitigation than reactive claims processing: modelling exposures, identifying accumulation risks, supporting underwriting decisions. Not automating claims.
How Complexity impacts your AI Strategy
The strategic question isn't "should we do agentic AI?"
It's "where on this spectrum do our claims sit, and what does that mean for our approach?"
High-volume, low complexity: Automate aggressively. For gadget, routine pet, windscreen, and basic travel claims, near-full automation is feasible. Customers benefit from speed. A gadget claim resolved in hours rather than days is a better customer outcome. Costs reduce materially when you're not paying handlers to process formulaic decisions. The governance requirements here are manageable. You need quality assurance through sampling. You need monitoring for drift. But you're not making decisions that will end up in court.
Medium-complexity: Augment, don't automate. For standard motor, routine home, and simple commercial claims, AI should enhance the customer experience and the work of handlers rather than replace them. AI handles intake, extracts data from documents, identifies relevant policy terms, and suggests an initial assessment. The handler makes the decision based on what the AI surfaces, not on what the AI concludes. This is augmentation: the AI does the heavy lifting on data processing and presents the right information. Deterministic rules in the workflow can get you close to Straight Through Processing for low-risk, low-value claims. For the rest, human judgment determines the outcome. The handler remains accountable. The audit trail shows a human decision informed by AI analysis.
High complexity: Assist, don't augment. For injury claims, subsidence, large commercial, and professional indemnity, AI can add value in bounded ways. Document gathering and organisation. Policy wording retrieval and comparison. Similar claims identification for benchmarking. Schedule and workflow coordination. Draft correspondence for handler review. But decisions remain human. The AI is a research assistant, not a decision-maker. The complexity and stakes are too high for anything else.
Extreme complexity: Wait and learn. For cyber, terrorism, marine, and catastrophe, focus on data capture and organisation. Build the foundation for future capability. Use AI for analysis and pattern identification.
The Operational Reality: Running Multiple Strategies
Here's what the vendor demos don't address. If you take this framework seriously, you're not implementing one AI system. You're running multiple AI strategies simultaneously across different parts of your portfolio.
That means different systems or configurations for each tier. Full automation for gadget claims, augmentation workflows for motor, assistance-only tools for complex lines. Each has distinct governance requirements, monitoring approaches, and human oversight models.
Complexity isn't the only axis. Claims reach insurers through fundamentally different channels, and each channel shapes what AI can do and who controls it.
Broker-intermediated claims mean the insurer isn't the first point of contact. The broker fields the initial notification, captures data in their own systems, and passes it through. Data quality, completeness, and format depend on the broker's processes, not yours. Your AI system is working with whatever arrives, and that varies enormously across a broker panel.
Third-party administrators manage the entire claims process on the insurer's behalf. The TPA operates the workflow, employs the handlers, and makes day-to-day decisions within agreed authority levels. If you're deploying AI into a TPA-managed line, you're not just asking "what should the AI do?" You're asking who owns the AI governance: the insurer whose risk it is, or the TPA whose process it is? That's a contractual and regulatory question, not just a technical one.
Embedded insurance, where cover is distributed by a major brand at point of sale, introduces yet another dynamic. The customer's relationship is with the brand, not the insurer. Claims expectations are set by the brand experience. A consumer who bought device cover at checkout from a retailer expects the same seamless experience for the claim. That's an argument for aggressive automation, but the insurer may have limited control over the customer journey if the brand insists on owning it.
Each of these channels interacts with the complexity spectrum differently. A high-volume, low-complexity line distributed through embedded insurance is a strong candidate for end-to-end automation. The same line intermediated through brokers with inconsistent data capture is a messier proposition. Your AI strategy needs to account for both dimensions.
It means routing decisions at intake, before you know the true complexity. A claim that appears simple at FNOL may reveal complexity as it progresses. Injury severity might not be apparent initially. Fraud indicators might emerge. The AI system that handles initial intake must route correctly and provide graceful escalation when complexity becomes apparent.
It means escalation paths when claims change category. What happens when a "simple motor" claim develops injury complications? When does a "routine home" claim reveal subsidence? The system needs to hand off cleanly, with full context, to different workflows.
It means skills and training for handlers working across modes. Your handlers need to understand what the AI is doing, where they need to apply judgment, and when to override or escalate. Different claims require different interactions with AI tools.
This is a significant step from today's choreographed workflows. In a deterministic system, data validation (both input and policy validation) occurs at defined points, with known inputs and outputs that determine subsequent steps. In an agentic system, you're asking AI to determine workflow, handle user interaction, make decisions about data capture and validation, and orchestrate across multiple subsystems.
Even in an agentic architecture, claims will likely begin in the same way: the nature of the incident, date, time, location, and the parties involved. This basic information is needed for policy coverage validation. There is no point in filing a full claim if it can be rejected quickly. That's better for the customer and insurer alike. But the orchestration logic (the criteria for dropping claims to handlers, triggering automation, or routing to specialists) becomes critical. Testing that orchestration requires a different skill set than testing traditional workflows.
Building and testing these systems requires a mix of AI expertise, prompt engineering, development for integration into core and third-party systems, and both manual and automated testing. The crucial requirement is for teams with a balanced mix of skills and experience across all these domains.
The Portfolio View of Strategy
Most insurers don't operate at a single point on this spectrum. They have a portfolio of lines with different characteristics.This means your AI strategy isn't one strategy. It's multiple strategies matched to different parts of your portfolio.
The framework: map your claims portfolio to this spectrum; identify where you have volume (the automation opportunity) and where you have complexity (the augmentation opportunity); match your AI approach to each segment; and build governance appropriate to each segment's risk profile. And remember that each segment may be distributed through different channels, direct, broker, TPA, or embedded; each adding its own integration, governance, and data quality considerations to the mix.
A 2% error rate on gadget claim triage is very different from a 2% error rate on injury claim coverage determination. Your governance, monitoring, and human oversight requirements should reflect that difference.
The Organisational Reality
Large insurers rarely have a single claims system. Divisions often operate on different platforms, frequently the result of acquisitions in which an insurer inherits systems and processes built for different purposes. Integration is expensive and risky, so parallel operations persist for years, sometimes permanently.
Any AI strategy must integrate with existing systems. That means starting with one line, one system, one well-defined problem. Prove value and compliance there first so you are netter equipped to tackle the next compelling problem.
The Uncomfortable Implication for full automation
This framework has an uncomfortable implication for the "full automation" narrative. For a significant portion of most insurers' portfolios, human judgment isn't a transition phase to be automated away. It's a permanent requirement.
AI will transform what handlers do. It will make them more productive. It will let them focus on judgment rather than administration. But for complex claims, it won't replace them.
Insurers who recognise this will develop sustainable AI strategies. The ones chasing vendor promises of full automation across their entire portfolio will learn expensive lessons about what happens when you automate judgment that shouldn't be automated.
To discuss this in greater detail please feel free to contact me
Next in series: Agentic AI, when nobody can define it, how do you govern it?
References
[2]: Financial Conduct Authority, "Motor Insurance Claims Analysis - Multi-firm Review" (July 2025).
[3]: Aviva, "Subsidence: What You Need to Know" (December 2024).
[5]: British Geological Survey GeoSure datasets, "BGS GeoSure.
[7]: Cyber Monitoring Centre, "Statement on Ransomware Incidents in the Retail Sector - June 2025".
[9]: Insurance Journal, "M&S First-Half Profit Hammered by Impact of Cyber Hack" (November 2025). & Reuters, “M&S forecasts rebound after cyber hack halves first half profit“.
[10]: Sangfor, "Marks & Spencer Cyberattack: A Wake-Up Call for Supply Chain Cybersecurity" (May 2025).
