Skip to main content

detected late or never - what insurers miss in their own dataAsk anyone working in an insurance company whether they have enough data, and the answer is almost always yes. Policies, claims, premiums, brokers, providers, payments. Years of it, stored across systems built for different purposes at different times.

Ask whether they can get a reliable answer to a simple business question by the end of the day, and the conversation changes.

That gap, between having information and being able to understand it, is one of the most expensive problems in the industry. And it rarely shows up as a line item.

The numbers arrive late, and they don’t always agree

Production, loss ratios and portfolio figures are the pulse of an insurance business. Yet in many organizations they still arrive through reports built by hand, pulled from several sources and stitched together in spreadsheets.

By the time a report reaches the people making decisions, the situation it describes has already moved. And it’s common for two teams to show up to the same meeting with two different versions of the same number. The discussion shifts from what to do to which figure is right.

The result is a slow, fragmented view of the business.

Late answers, missed patterns

The questions that shape pricing and underwriting, like how claims experience relates to profitability by region, channel or broker, require crossing the most data from the most places. So they get answered after the fact.

The same applies to fraud. Inflated claims and suspicious payments hide in the relationships between data points, and when systems are reviewed one report at a time, those patterns are caught late, or never.

In both cases, the cost isn’t the data the company lacks. It’s the time it takes to see what’s already there. By the time the answer arrives, the decision it should have informed has already been made.

The root cause: an analyst’s loop

Behind all three symptoms is the same workflow, repeated thousands of times a year.

An analyst receives a question. They figure out where the relevant data lives: which system holds the policy, the client, the claim, the provider. They write the query, export the file, pull it into a spreadsheet and build the report. Then the question shifts slightly, and the whole process starts again.

Two things stand out in this loop. First, every analysis begins from zero. What was learned answering yesterday’s question rarely makes the next one faster. Second, the focus is always on what was asked. Nobody has the time to look at what else the data might be saying.

Many insurers have already tried to close this gap with a chatbot. It helps, up to a point. Ask for the status of a specific policy and the answer comes back in seconds, because it’s a single lookup. Ask why renewals dropped in one region this quarter and the chatbot has nowhere to go. The answer isn’t in one place. It’s spread across the operational database, a spreadsheet maintained by hand and a BI tool that covers only part of the picture.

A lookup and a question are not the same thing. Most tools are built for the first. The business runs on the second.

Meet our Agentic Data Intelligence Solution

Nybble has been working with the insurance industry for more than 20 years, accompanying its digital transformation for two decades. Through all that time, this loop has been one of the most persistent patterns we’ve seen. The Agentic Data Intelligence Solution we developed was built to break it.

It works through a single chat interface, the assistant a company already runs or one built as part of the delivery. Direct lookups keep working as they do today. What changes is what happens with open-ended questions: instead of hitting a wall, it opens an investigation. It explores the relevant sources, connects what it finds and reports back in the same conversation.

Along with the expected result, it delivers its own findings drawn from the data. The patterns nobody had time to look for become part of the answer.

The analyst stays at the center. What changes is that they now work with an AI copilot, reviewing results, asking follow-ups and making the final call.

How it works

  1. Ask. An open-ended question goes into the chat, for example, “Why did declined renewals spike last quarter?”
  2. Provision. A sandbox is set up from the company’s schema and guardrails, with no data or secrets built in.
  3. Investigate. it connects to more than 10 databases, a centralized data lake through AWS Athena and external systems via API, running read-only queries with sensitive fields redacted. Depending on the question, it processes data at volume and applies statistical or machine learning analysis.
  4. Deliver. The analysis comes back with findings, sourced figures and a notebook, saved and versioned.
  5. Return. Follow-ups happen in the same chat. The analysis can be reopened weeks later or reused the next time a similar question comes in.

Built to be trusted with sensitive data

In insurance, speed means nothing without control. An agent that can query production data has to earn that trust in its architecture, not in a policy document.

  • Read-only by design. Every query is restricted to reading at the tool level. It can look, never write.
  • Sensitive data stays out. Social Security numbers, dates of birth and other sensitive fields are redacted before anything reaches the model.
  • Disposable sessions. Each investigation runs in its own sandbox, with credentials provided only at runtime.
  • Nothing gets lost. Every analysis is saved as a record that can be reopened, even after the sandbox closes.

An analyst that gets better with every question

Every analysis is kept as a recipe: the steps, the queries, the assumptions that are held. When a similar question comes in, it starts from what already worked instead of starting from zero, and improves a little each time.

Underneath, every query is grounded in an ontology generated from the company’s own schema. Subject matter experts can review it as an interactive graph and adjust it against real data, so it reflects what the business needs, not just what the tables contain. Each session feeds back into it, so it keeps evolving with the questions people ask.

And it doesn’t stop at databases. Documents, forms and correspondence are read through DocuGPT, Nybble’s document intelligence framework, and folded into the same investigation. Partner systems and third-party services connect the same way, as one more source the agent can query.

Your data. Your agent.

This isn’t a product on a shelf. It’s delivered into the company’s own systems, and the capability stays with the company.

Each conversation and each analysis session shows its own token usage and cost, so teams see what an investigation costs instead of waiting for a monthly invoice. The model behind the agent can be changed by project or by session, so cost and capability match the question. And its answers are evaluated over time, not just on the day it went live.

It’s already live in production at a life insurance carrier, plugging into its servicing assistant’s existing tools without changing them.

From answering questions to understanding the business

The value isn’t only in getting answers faster. It’s in changing what kind of questions a team can afford to ask. When a cross-cutting analysis takes minutes instead of weeks, people stop asking only what’s urgent and start asking what’s important.

For insurers, that means seeing loss and profitability trends while there’s still time to act, spotting suspicious patterns before payments go out, and working from a single version of the truth.

Insurance has never lacked data. What it needs is a better way to turn that data into answers.

If your team has plenty of data and very few answers, let’s talk. We can show you how this works with the questions your business is asking today. Reach us at info@nybblegroup.com or visit nybblegroup.com.