Insurance AI Is Full of Hammers Looking for Nails

I read competitor release notes with genuine interest. I have to know what the market is doing to know whether I am pointing Peak3 the right way. Lately, almost every release is about AI. Some of it is impressive. Some of it truly puzzles me. 

A recent announcement from a competitor (that shall remain nameless) advertised a suite of new AI agents built into the latest release of its core platform. One AI agent manages write-offs on aged accounts receivable. Another AI agent handles refunds on cancelled policies and overpayments. 

My reaction: Why is this an AI problem? 

Ninety-plus percent of write-offs and refunds in a well-run billing module are resolvable through deterministic business rules, ageing thresholds, return-premium calculations, etc. A workflow engine with a decent rule and calculation engine underneath handles the volume quietly, every day. 

Running these through an LLM-backed agent delivers probabilistic outcomes, higher compute and inference costs, and a monitoring burden you didn't previously have. In insurance, that matters. Regulators and insurers want deterministic, auditable decisioning on transactions. Insurance CFOs want operating leverage and not linear increase in transaction costs from AI tokens. 

If you have a hammer, not everything is a nail. And if your current core can't handle something as foundational as aged AR write-offs without manual intervention, the answer is not to wrap AI around it. The best answer is to fix the core. 

Where does Peak3 focus agentic AI?  

My answer: Where it creates genuinely differentiated value — where deterministic logic falls short. We have been building across three main categories: 

• Conversational Support Agents that manage, improve and automate 24/7 customer interactions through conversational multi-modal AI. They provide tailored support, answer queries, guide users, collect structured data, and pre-check information and documents. 

• Intelligent Document Processing Agents that eliminate manual work, automate document intake processes and increase accuracy with vision-powered AI. They intelligently recognise, understand, extract data from, and automate the validation of documents. 

• Assessor & Triaging Agents that automate and improve decisions and risk management. They autonomously assess insurance transactions, develop risk reports and scores with transparent evidence chains, dynamically route transactions, and assist experts with final decisions. 

These are the problems where agentic AI earns its seat at the table. The probabilistic cost is worth paying because the deterministic alternative either doesn't exist or doesn't work at scale. 

The rest? Workflow it. Rule-engine it. Configure it properly in the core system and move on. 

AI everywhere is not a strategy. It's a way to burn budget and management attention on problems you already knew how to solve. 

If you're reviewing your AI roadmap, try a simple test on each planned agent. Ask your team whether a well-configured rule engine could handle 95% or more of the cases. If yes, you've just saved a lot of implementation effort and token costs.  

If no, you’re probably looking at a real agentic AI use case. Or if the “no” is because your core system cannot do it, while the answer would have been “yes” if you had a modern core...then start thinking about modernising your core system.

-Bill Song, Peak3 Co-Founder and CEO