At some point, you need an AI orchestration platform

I am going a bit down memory lane today. Before founding Peak3 in 2018, I was one of the early hires at ZhongAn, China’s largest digital-only P&C insurer. For those who don't know the story: ZhongAn scaled organically to roughly US$5bn in P&C gross written premium in a little over a decade, built cloud-native from day one, on an open, API-first architecture. 

ZhongAn was a pioneer in embedded insurance more than a decade ago, and one of the first insurers globally to operationalise social commerce and short-video platforms as a primary distribution channel. ZhongAn scaled to tens of billions of micro policies each year, reaching half a billion customers. 

Today, I’d argue ZhongAn is the insurer that has scaled AI most systematically and most successfully in the industry. Every few months, I catch up with former colleagues and I'm still surprised by the pace. 

Entire workflows have been moved to autonomous AI agents. A/B tests against control groups of human staff show on-par or better performance on customer satisfaction, conversion and risk management. It's not just about efficiency. It's about outcomes. 

What impresses me most is the systematic breadth. By my last count, there are more than 200 distinct AI agents in production, with an idea-to-production rate somewhere around 80%. These AI agents span the full value chain from marketing to underwriting to claims. These agents made more than two billion LLM calls and consumed more than three trillion tokens last year. 

How? Early and deliberately, ZhongAn invested in a centralised AI orchestration platform on which all agent applications are built and their lifecycles managed. In practical terms, that means: 

• A no- and low-code agent framework so business teams can build and deploy simpler agents without waiting on engineering. 

• A pro-code framework for AI engineers to develop complex, multi-step agents with deep system integrations. 

• A managed RAG layer with curated, domain-specific knowledge bases: policy clauses, underwriting rules, claims precedents, regulatory circulars. 

• Short- and long-term agent memory to sustain multi-turn interactions and personalisation across sessions. 

• An LLM gateway providing unified access, cost metering, traffic control, fallbacks and audit across multiple foundation models. 

• Guardrails to enforce behavioural boundaries and compliance filters on inputs and outputs. 

• Observability, evaluation and CI/CD pipelines tailored to agent assets (prompts, knowledge bases, models) rather than just code. 

Strip any one of these capabilities out and two things happen. First, you slow your pipeline: every new agent becomes a bespoke engineering project. Second, you lose control of cost, compliance, consistency and security. Agent sprawl arrives before scale does. 

At Peak3, I have been building with my teams on a similar platform that insurers can use to centralise their agentic AI operations. It incorporates the capabilities mentioned above and we pre-build our AI agents on the same framework. 

Most insurers I talk to are still at a very early stage. They are POCing individual agents, and a centralized platform may feel like overhead. But I think it isn't. It's the difference between a handful of agents in production and two hundred or more. 

If you're past the “can we POC an agent?” question and heading into “how do we scale efficiently, compliantly and securely?”, the platform conversation is overdue.