AI will Replace or Elevate You

"AI will not replace humans — but humans with AI will replace humans without AI."

That's the common story told across most industries, including the insurance sector. I agree with this statement but with a big caveat.

In operations roles, the number of humans without AI to be replaced is significantly larger than the humans with AI replacing them. Also, the real question isn't whether AI replaces insurance underwriters, claims handlers or customer service staff. It is what the job looks like once AI agents are inside it. And, importantly, how many human staff are still required. 

My working view, from conversations with insurance operations leaders over the last eighteen months: the underwriter of the future is an agent orchestrator. So is the senior claims adjuster. So is the complex-case customer service specialist. They don't disappear, but their focus shifts. 

Instead of working one case at a time, an experienced underwriter reviews the decisions of ten or twenty AI agents running in parallel: approving the clean cases, correcting the borderline ones, rejecting the misguided ones, and feeding the corrections back into the agents' evaluation loop. The human role gets denser, not thinner. It also gets harder. It needs judgement and pattern recognition. And it requires the discipline to push back on a confident-looking recommendation that is quietly wrong (something everyone having dealt with AI chatbots over the last years will already know).

I see two pathways: 

1. Human-in-the-loop is the target state for complex issues, not a transition phase. For the next 5 to 10 years at least, high-stakes decisions in insurance — complex underwriting and claims, sensitive claim denial and coverage dispute — will sit with a human. That human has to be able to see the AI agents' reasoning, intervene at different points in the workflow, and override cleanly, without breaking the process. Most core platforms handle this clumsily today because they were designed for human-first workflows, not human-in-the-loop workflows.  

2. Autonomous handling is the target state for most issues; human-in-the-loop is the just the transition phase. Standard underwriting can be done autonomously by AI agents, same as the processing of most claims. Customer service requests will be almost exclusively handled by AI agents — fully autonomously. For some of these, humans may still need to be required to click the final button due to regulatory requirements (e.g., under the AI Act in the EU) or until our confidence in the AI is high enough. I can already see the autonomous agent potential at my previous employer ZhongAn. There, the vast majority of customer service requests, claims and underwriting decisions are done fully autonomously, combining powerful rule sets that are enhanced with AI agents. 

I mention “human-in-the-loop” (HITL) above. HITL is the most commonly used term. But in fact, the target state should be “human-on-the-loop” (HOTL). HITL refers to operating models where humans are directly embedded into the AI agent's execution workflow as a mandatory synchronous checkpoint. When an agent encounters predefined critical decision points or when uncertainty exceeds a threshold, it immediately pauses execution and waits for explicit human input (approval, modification, or rejection) before proceeding. HITL is a “blocking” collaboration model. 

HOTL, on the other hand, is a "supervisory" collaboration model. Humans shift from being "executors" within the process to higher-dimensional "supervisors." AI Agents are granted full end-to-end execution authority and can autonomously complete tasks within human-defined rules and boundaries. Humans only intervene when the system triggers anomaly alerts, risk exceeds thresholds, or strategic adjustments are required. 

In any case, talent implications follow. We will see a shift from junior to senior roles. The impact on talent strategies and career paths is still unclear. Also, the senior skill profile is changing. These seniors will need to be fluent in pattern recognition on agent output and prompt crafting for edge cases, amongst others. This can become a multi-year professional evolution. 

Irrespectively, explainability and traceability are operating requirements, not a compliance checkbox. An underwriter reviewing twenty agent decisions an hour needs to know, in seconds, why each agent concluded what it did. What's needed is per-decision provenance: which rules were executed, which data was retrieved, which prompts ran, how each confidence score was calculated, etc. If the core or AI platform can't surface this, the human orchestrators are flying blind. 

That's why we emphasise explainability and traceability in the AI assessor agents that Peak3 has been building. Our clients of these pre-built multi-agent systems can realise seamless HITL and HOTL processes with clear evidence chains. Each rule triggered, the rationale for each decision and each data source accessed are clearly made visible — as are confidence scores. Also, human experts can provide immediate feedback and override AI decisions.

-Bill Song, Peak3 Co-Founder and CEO