Author: Timo Loescher, Insurance Industry Lead, Quantexa
Even when technologies are new, you hear some stories again and again. The details differ, but the shape of them is the same.
Say an insurer launches an underwriting copilot. The pilot is well scoped, focused on one product line where the data in the workbench is clean and robust. The copilot helps underwriters summarise submissions, review appetite fit, and even draft pricing narratives. At this size, it works.
The C-suite is excited about the results, and leadership soon announces a broader rollout. But when it comes time for a wider launch, outside the parameters of the initial pilot, everything goes pear-shaped.
People task it with more than had been tested, and at a wider scale. The copilot starts returning confusing responses to queries. It pulls outdated exposure details, fails to connect multi-line submissions, and ignores loss histories the underwriter would expect to see. It might suggest a form tailored for a different jurisdiction, mistake a broker for the insured, or create a host of other problems.
As frustration builds, trust fades, and eventually most underwriters avoid the system altogether. The AI team is left wondering where they went wrong.
Sound familiar?
Many insurers have deployed lightweight copilots, things like meeting summarisers, call note generators, or email draft assistants, and done so with success. But when they try to scale GenAI for higher-value tasks, things break.
Why GenAI struggles in insurance
Part of the reason is that large language models (LLMs) are trained on public data. Because of this, they lack context for any specific business use, so the LLM does not know your policies, claims, or customers. To solve this, AI firms created a technique called Retrieval-Augmented Generation (RAG), which grounds an LLM's responses in an organisation's actual data. Instead of relying solely on its trained knowledge, the model retrieves relevant internal documents to inform its answers. Ideally, this produces more accurate and useful outputs.
But as the old computing maxim goes: garbage in, garbage out. RAG is only as good as the datasets it draws from. They have to be accurate, distinct enough to query successfully, rich in context (within the organisation and industry, but also connected to the real world), and connected in traceable ways.
Too often, despite an organisation's dedicated data management resources, those datasets fall short on one or more of these criteria. So when copilots are given complex tasks, they fall apart. They can summarise a meeting, but it is dangerous to trust their advice on how to react to a problem. Descriptions are easy. Informed prescriptions are hard.
Copilots' challenges at scale
When insurers attempt to operationalise GenAI copilots, these four areas are the biggest sticking points.
- Unresolved entities
These are entities that show up in the dataset under slightly different names. For example, the "AB Energy" named in a submission from the broker is not recognised as the same as "Alan-Behr Energy Limited" in the underwriting workbench or "AB Energy Ltd." in the claims history. Often copilots cannot stitch the story together.
- Over-resolved entities
This is the opposite problem. When five prospect references live in five systems, each with a similar common name, copilots either guess at which version to use or overconfidently collapse unique records into one picture. Either way, it can lead to a hallucinatory response.
- Missing relationships
Sometimes the copilot cannot see an individual as a representative of a company and treats them as separate entities. It might see "Henry Loomis" as a personal lines policyholder, and not in his role as the CEO of a business operating in a high-appetite market. Or it might miss that a top-producing agent recently changed firms, or a company's involvement as a third-party claimant under another line. Relationships are fluid, and even when they are updated in one part of the dataset, that update might not carry over throughout the whole.
- Scattered, stale data
Outdated relationships are just one example of this. Even when copilots retrieve the "right" data (or, more precisely, the data that was requested), it is often outdated or incomplete. This is because it comes from one system, not the full view across the organisation's data estate, which leaves the LLM to guess, average, or hedge, all of which can lead to a bad result.
While agentic frameworks have shown promise, AI agents are ultimately still reliant on LLMs. And it is unwise to expect an LLM to resolve similarly named records into a single entity or choose which version of conflicting data to trust. It might miss connections that are there, or see connections when there are none. Letting an LLM substitute for human reasoning is beyond its skill set. They are not built for that.
That is where Contextual RAG comes in.
What is Contextual RAG?
Graph RAG brought a useful advance to retrieval-augmented generation by layering in graph-based techniques to map connections between entities. But it has real limitations in practice. The relationships it works with are often manually defined, which means they only go as far as the solution architect imagined they needed to. And the underlying data quality is taken largely on faith; as with any AI tool, what you get out depends entirely on what you put in (garbage in, garbage out, remember?).
Quantexa's Contextual RAG was built with those limitations in mind, and with the specific messiness of insurance data as its starting point. Insurance records are complex, inconsistent, and full of relationships that are easy to miss if you are not looking for them in the right way.
The approach pulls together a few capabilities that work better in combination than they do individually. Entity resolution sits at the base, automatically identifying who and what is actually being referenced in the data, then cleaning and consolidating those references across the full data estate. From there, relationships between people, places, and locations are derived from the trusted operational records that insurers already rely on, things like claims and policies, rather than being manually mapped or assumed.
The system is also built to understand that different parts of the insurance business have different priorities. What matters in underwriting is not the same as what matters in claims, and the graph analytics reflect that. Finally, when a query comes in, entity context is used to sharpen the prompt itself, steering the model toward sources that are both relevant and reliable.
The result is a copilot that does not just locate information. It understands what that information means in context and responds accordingly.
Putting it into action
Here is how it works. Quantexa's Decision Intelligence Platform provides a contextual fabric: a connected, contextual, and AI-ready knowledge layer that sits between your data lake and operational applications. This fabric provides a reliable, enterprise-wide view that serves as the single source of truth.
Q Assist, Quantexa's agentic copilot, then works alongside your teams, allowing users to talk to their data. Quantexa AI ensures that any AI, from copilots to autonomous agents, operates with verified, contextual information. And Quantexa's Agent Gateway enhances any third-party agents and copilots you build, bolstering security and explainability across distributed agent ecosystems.
This results in better grounding, fewer hallucinations, deeper insights, and, ultimately, higher adoption, because users can actually trust what they see.
Expanding copilot capabilities
With Contextual RAG, insurers can finally scale copilots beyond summarisation and stenography to actually move the business forward.
On the distribution side, it gives you a clearer picture of which partners send good risks, which prospects to target next, and where cross-sell or upsell opportunities exist on any given deal. For underwriting, it helps prioritise key submissions and surface hidden risks or missed appetite matches, speeding up the quoting cycle and improving both win rates and deal profitability.
Better context also means better experiences. The full-context servicing and optimised digital experiences that Contextual RAG enables make every customer interaction feel more personal and individualised, which drives stronger retention. And in claims, broader confidence in complexity and severity, connected risk patterns, and customer intelligence can accelerate reserving, settlement, and payment decisions.
From copilots to connected intelligence
These are the kinds of opportunities that become possible when GenAI is grounded in accurate, contextual data. The value goes beyond individual copilots and comes instead from connected intelligence that can scale across the business.Without the right data foundation, even the most sophisticated GenAI tools are constrained by the limits of what they can see and understand. With it, the technology can reason across relationships, surface patterns that would otherwise go unnoticed, and support decisions with a level of context that generic solutions cannot match. That requires a data foundation that understands not just insurance, but the specifics of how your organisation operates.
