The insurance industry has always been data-intensive. Policies, claims, underwriting submissions, regulatory filings — the sector runs on documents. So when artificial intelligence promised to transform how organisations process information, insurers were among the first to pay attention. But three years into the AI wave, the gap between expectation and reality remains wide.
At Imagefast, we work with insurers across the London market and beyond. We've seen AI implementations that genuinely transformed operations, and we've seen expensive pilots that never made it past the proof of concept. The difference between the two almost always comes down to the same set of factors.
The Problem with Starting from Technology
Most AI conversations in insurance start the wrong way round. A vendor demonstrates an impressive capability — extracting data from unstructured documents, summarising policy wording, classifying claims by type. The technology is genuinely compelling. Leadership gets excited and asks the IT team to find a use case.
This is where things go wrong. Finding a use case for technology you already have is the opposite of solving a business problem. The result is typically a pilot that works in a controlled environment but fails when it meets the complexity of real-world insurance operations. Handwritten endorsements, multi-party correspondence, legacy document formats, and the sheer variety of submission types create challenges that a demo environment never anticipated.
The organisations getting real value from AI start with the process. They identify where human effort is being spent on repetitive, low-value work. They map the specific bottlenecks where manual document handling slows down underwriting decisions or claims processing. Then they look for AI capabilities that address those specific problems.
Where AI Is Actually Delivering Value
Across our insurance client base, we're seeing consistent returns in three areas.
Intelligent document classification
Insurance submissions routinely arrive as bundles of mixed documents — policy schedules, loss runs, financial statements, certificates of insurance, and broker correspondence all combined in a single PDF or email chain. The manual effort required to separate, identify, and route these documents to the right team is substantial. AI-powered classification is now handling this at scale, with accuracy rates that match experienced human operators. The key is training the model on the organisation's own document types rather than relying on generic classifiers.
Data extraction from submissions
Underwriters spend a significant portion of their day pulling data from submission documents into structured systems. AI extraction tools can now handle this for the majority of standard document types — pulling key fields like insured name, coverage limits, policy period, and loss history into the underwriting workbench automatically. The time saving is meaningful, but the bigger value is consistency. AI extracts every field every time, eliminating the variability that comes with manual data entry across a team of underwriters.
Claims triage and routing
First notification of loss often arrives in unstructured formats — emails, letters, phone transcripts. AI can now read these incoming claims, classify them by line of business and severity, and route them to the appropriate handler. For straightforward claims, this can reduce the time from notification to first contact from days to hours. For complex claims, it ensures the right specialist is engaged from the start rather than after multiple handoffs.
The Data Foundation Problem
The single biggest barrier to successful AI adoption in insurance isn't the AI itself — it's the state of the data that feeds it. Most insurers are working with document management systems that were implemented years ago, often with inconsistent metadata, mixed-quality scans, and content scattered across multiple repositories.
AI models need clean, well-structured input to deliver reliable output. If the underlying documents are poorly scanned, inconsistently filed, or stored in formats that resist extraction, no amount of AI sophistication will compensate. This is why we frequently find that the first step in an AI programme isn't deploying AI at all — it's getting the document management foundation in order.
This doesn't mean a multi-year transformation programme. It often means targeted improvements: upgrading scan quality for incoming documents, standardising metadata for key document types, consolidating content from disparate systems into a single searchable repository. These foundations make AI viable and make the organisation more efficient regardless.
Governance and the Human Layer
Insurance is a regulated industry. Decisions about coverage, claims, and pricing carry real consequences for policyholders and the broader market. AI can support these decisions, but it can't replace the judgement that experienced underwriters and claims professionals bring.
The most effective AI implementations in insurance treat the technology as an assistant, not a replacement. AI handles the routine extraction, classification, and summarisation. Humans handle the exceptions, the judgement calls, and the relationship management. The governance framework makes clear where AI operates autonomously and where human review is required.
This isn't just good practice — it's increasingly a regulatory expectation. The organisations that build governance into their AI programme from day one avoid the painful retrofitting that comes when regulators start asking questions about automated decision-making.
Getting Started Without Getting Stuck
The insurance organisations that are furthest ahead with AI share a common approach. They started small, with a single well-defined use case in a specific part of the business. They measured the impact rigorously — not just the technology performance, but the operational outcome. They iterated based on what they learned and expanded only when the first use case was delivering sustained value.
They also invested in their people. The underwriters, claims handlers, and operations staff who work with AI every day need to understand what it does, what it doesn't do, and how to handle the cases where it gets things wrong. Training isn't a one-off event — it's an ongoing part of the operating model.
AI in insurance is no longer experimental. The technology is proven, the use cases are clear, and the early adopters are building genuine competitive advantage. But the path from pilot to production is paved with process expertise, not just algorithms. The organisations that recognise this are the ones turning AI from a buzzword into a business capability.
