AI in insurance operations after the hype: what has to be true first
Most insurers have heard what AI could do. Fewer have asked what has to be true of their data and processes before it does it.
The second question decides whether an AI pilot becomes part of operations or stays a demonstration. This post looks at where AI helps in claims, underwriting support and servicing, and what we would check first.
Where AI helps in operations
In our experience, the useful applications of AI share a shape. They take work that is repetitive, document-heavy and rule-bound, and leave a person to make the call that carries judgement.
Claims: reading a first notice of loss, triaging by type and urgency, and flagging exceptions for a handler.
Underwriting support: extracting data from submissions and documents, so the underwriter starts from a structured file.
Servicing: classifying and routing requests, and drafting responses for a person to approve.
ACORD, the insurance data standards body, describes similar territory. In its Reference Architecture release of 30 July 2026 it says: "Agentic AI can assist with insurance processes such as claims triage, fraud flagging, and document intake, but accountability must remain clear across both digital and human participants." (ACORD, 30 July 2026)
The data has to be ready
AI works on the data it is given. Somesh Mukherjee, VP Solutions Architecture at ACORD Solutions Group, wrote in July 2025 that "AI models require structured, high-quality, and interoperable data to function effectively." (ACORD on LinkedIn)
It matches what we would check in practice:
Is the same thing recorded the same way across systems, such as a claim, a party or a coverage?
Is the data held in a form a system can read, or in documents and spreadsheets?
Can you trace where each field came from?
Who owns the quality of the data and fixes it when it is wrong?
The process has to be defined
An AI step needs a place to sit. Before adding one, we would write down three things: what the step receives, what it must produce, and what happens when it is unsure. Without those, nobody can say whether it is working. We would map the current process first, including the manual steps and workarounds that never made it into the process document.
Keep a person accountable
Decide in advance which outcomes need human review or approval, who gives it and how it is recorded. In a regulated process, in our view, that is what lets an AI step earn trust rather than ask for it.
Measure before you start
Pick one process, record its current speed and accuracy, and set the target before the pilot starts. Our work with large insurer on their technical pricing programme was a platform replacement and not an AI project. Its reported outcome was 1.5x improved speed and 15% improved pricing accuracy.
The takeaway
AI is an operations question before it is a technology question. Get the data readable, the process defined and the accountability named, and the choice of tool becomes much easier.
Which single process in your operations has data clean enough and steps clear enough to try an AI step on? If you want to talk it through, send us a message through our contact page.
Want to be first to hear these insights by email? Join the Genesis reader list on our Insights page.
