💎 AIB - Executive Series - William Steenbergen on how insurers can harness AI across the end-to-end value chain (Federato internal use only)

Hey, Fabio here!

In today’s Executive Series, I am joined by William Steenbergen, CTO at Federato.

Across the market, insurers are discovering that the real challenge with AI is the architecture underneath it.

The bigger opportunity is no longer automating isolated tasks.
It is redesigning end-to-end operations so the business can use AI across the full workflow, with the right context, the right tools, and the right portfolio intelligence to improve decision-making. 

So we focused on a practical question:

What does it take to build a truly AI-native insurance operation, rather than simply layer AI on top of fragmented systems and workflows? 

William explained why old core architecture still blocks value, why context and tools matter more than the model itself, why real-time portfolio steering is critical if insurers want AI to improve underwriting decisions, and what will separate carriers that genuinely harness AI from those that keep layering it on top of foundations that were never designed for it. 

Let’s dive in.

The real value from AI doesn't come from replacing your existing process. It comes from rethinking the process itself. That means treating an AI agentically, not as a chatbot that answers questions, but as something that executes decisions and actions.

An AI agent needs three things. 

First, an LLM. Everyone has access to models like Gemini, Opus, or GPT, and that's not where carriers differentiate. What matters more is context and tools.

Context is the data an agent needs to make decisions: underwriter interaction history, policy financials, exposure data, quote documentation, and forms.

Tools are what let an agent act, not just answer. Running a rater, generating a document from a template, and accessing the web.

The problem with old core systems is that many of them were built 25-30 years ago and expanded through years of acquisitions. The result is not just legacy technology, but fragmented schemas and inconsistent data structures. Insurers already struggle to pull this data out manually. For an AI agent, it's nearly impossible. 

The same applies to tools. They often don't exist in a form an agent can use reliably at scale, slow SOAP APIs, weak validation, and no clear error handling. We've had agents interacting with old core APIs and literally brought the systems down. You can have the best LLM in the world. Without the right context and tools, the agent will fail.

The difference is whether you are thinking about the full process or just a series of isolated tasks.

I am seeing more people take a best-of-breed approach. They may buy one AI-focused tool for ingestion, another tool for content management, another tool for rating, and so on. Each tool may be AI-native in its own area. But connecting all of that and putting agents on top is still very hard. Often, it fails.

Insurers may get 5% efficiency gains across isolated subtasks, but never reach the much larger transformation possible when the full underwriting process is redesigned holistically. That is not just a problem with old core systems. It can also happen when insurers assemble a group of point solutions. The tools may be modern individually, but the operating model is still fragmented.

A truly AI-native insurance operation is not just about automating tasks. It is about giving agents the context and tools to operate across the full workflow.

If you make underwriting much more efficient but do not improve decision-making, you are just making wrong decisions faster. That is worse.

For example, if an AI agent helps underwriters start from a quote before they have even reviewed the submission, that creates speed. But if you are not also thinking about portfolio strategy and better decision-making, that speed can create a deeper problem.

The context of the portfolio needs to live in the same place as the underwriters and the AI agents. The agent needs access to underwriting goals, underwriting strategy, and portfolio goals when making decisions.

It could also mean creating dynamic portfolio goals, such as limiting a certain type of exposure within the portfolio. The point is that those goals need to be available as real-time context.

Today, guideline changes often arrive as a circular, a SharePoint upload, or an email telling underwriters to look somewhere new. Many executives know some underwriters will still be using the old guidance. If the agent works from a single updated source of truth, that inconsistency becomes easier to manage.

Once you do that, you can also start tracking something very important: when and where underwriters deviate from what the AI agent recommended. If that happens consistently, there are two possibilities.

One possibility is that the humans are right and the underwriting strategy needs to change. In that case, you update the guidelines and give the agent better context.

The other possibility is that the AI agent is following the strategy correctly, but the underwriters are not. That creates a management issue around coaching, performance, and consistency.

That visibility does not really exist today. Many executives do not know whether underwriters are consistently following the guidelines or portfolio strategy. Real-time portfolio steering gives them a way to see that.

Importantly, I do not think the agent should automatically learn from human deviations by default. If it simply learns from what humans do, it may also learn human bias. The carrier should decide whether humans need to learn from the agent or whether the guidelines need to change.

It comes back to context.

AI agents can interact well with data when that data is structured in a standardized, clear, and consistent way. It needs to be obvious what the data means. It needs to be consistent across systems and across the policy lifecycle. It also needs to be easy to link different pieces of information together.

For example, an agent may need to retrieve an expiring policy from three years ago for a specific new business or renewal case. That depends on data being standardized, clearly defined, and easy to link across the full policy lifecycle.

Metadata also becomes more important.

If you store a premium, the system should define how that premium is stored. Is it in dollars or pounds? Does it always have two decimal places? Is it expected to sit within a certain range?

The same applies to class codes. If only certain class codes are valid, that should be clearly defined. Even column names matter. I had a customer send files where the columns were called “field A” and “field B.” As a human, I would not know what those fields mean. An AI agent would not know either.

The foundation needs clear field definitions, valid values, proper formatting, hard validation, and standardization across the lifecycle. That is what allows data to become a useful context for AI agents.

The two big things are context and tools.

Context is probably the most important part, and tools are the other critical piece if the agent is going to execute tasks safely. Insurers need endpoints that allow agents to perform actions such as generating policies, generating documents, or running a rater. Those endpoints need to work in a standardized and safe way.

If an insurer does not have an endpoint that can run a rater in a stateless way, within roughly five seconds 95% of the time, and provide clear human-readable errors when something fails, then the architecture is not set up for success.

The tools also need permissioning. You do not want an agent to go off and do certain things without control.

A standardized interface is important. If an insurer has 20 different tools from a best-of-breed approach, that means the agent has 20 different interfaces to interact with. That becomes hard to maintain and hard to scale.

For agentic AI to work safely, the architecture needs structured context, reliable tools, clear permissions, strong validation, and standardized execution paths.

The difference will be insurers’ willingness to take architectural and vendor decisions seriously, rather than spreading risk across many point solutions. If insurers truly want AI agents to operate across underwriting and want underwriters to start with a quote, they may need to replace parts of their policy admin environment or other systems.

They may also need to take more risk by working with fewer vendors rather than spreading everything across many point solutions. In my view, carriers that think across the full policy lifecycle will be in a very different position from those that continue with a best-of-breed model.

We have seen longer implementation timelines and worse user experience when implementations require integration with many different tools. When we have to integrate with 10 different tools, implementation timelines become longer, even if the customer is using less of the Federato product. Users also report a worse experience when there are many integrations compared with when they do everything in Federato.

The issue is not just implementation time. When insurers keep building integrations, they often create a larger data mapping and data warehouse project. Those projects are hard to do well, tend to take a long time, and often fail.

That does not mean insurers need to go all in at once. They can start with one line of business. But they do need to lean into the full policy lifecycle rather than treating AI as another layer on top of a fragmented architecture.

The takeaway from William is hard to miss: the real AI divide in insurance will be defined by which carriers can give AI the right context, the right tools, and the right architecture to operate across the end-to-end value chain.

AI is not just about access to an LLM or automating isolated tasks. It is about redesigning the process itself, so AI can act with real-time portfolio context, consistent data, and reliable execution paths. Without that, insurers may get incremental efficiency gains, but they will struggle to deliver the larger transformation possible.

That is also why Federato’s focus on real-time portfolio steering matters.

If AI only makes underwriting faster, without improving decision-making, insurers risk making wrong decisions faster.

The broader point is clear: move beyond fragmented point solutions, think across the full value chain, and build the context, controls, and workflows needed for AI to operate safely and at scale.

See you on Friday!

Fabio Caravita
Founder, AI Insurer Brief
[email protected]

Reach insurance execs and decision makers weekly.
72% Exec / C-level / Decision-makers.
Limited number of partner spots open each quarter.
→ Reply “Partner” for details and we’ll get in touch.

For Insurers, Brokers, MGA, consultancies and tech providers, exploring AI or with a live AI mandate but limited clarity on governance, processes, operating model design and execution.
→ Reply “Advisory” for details and the team will be in touch.