From Earthquake Models to AI Start-ups: Sam Zheng on the Bet No Actuary "Runs the Numbers" On

Most actuaries are trained to run the numbers before they act. Sam Zheng's biggest career decisions came from knowing when not to.

Sam started out modelling earthquakes and pricing motor insurance at Suncorp and Zurich, before a data science role at Quantium put him on a different path entirely. He's since co-founded three AI companies: Hyper Anna, acquired by Alteryx, Curious Thing, a voice AI platform backed by Westpac's Reinventure. He now leads the third, 477 Studio , with long-time business partner Dr Han Xu. This year, that path earned him the Actuaries Institute's 2026 Actuary of the Year award.

We asked him what that journey actually looked like from the inside: what he had to unlearn, what went wrong and what he'd tell an actuary weighing up the same move.

What was the opportunity that made you change from a stable actuarial career to build a start-up? 

I first started thinking about doing a startup when I was in a data analytics consulting role at Quantium. There wasn't one moment — it built up over those two years. By then my work had already shifted away from traditional actuarial pricing. I watched the industry slowly transform from statistical modelling into machine learning and data science. I also picked up a lot of engineering skills along the way. I met a colleague who had a similar idea, so we decided to give it a shot. That's how I quit the job and started my first startup. 

You taught yourself to engineer. What was hardest about that, and what does a normal week at 477 Studio actually look like now? 

My first programming language was R, a statistical language rather than an engineering language. The hard part of teaching myself was the unknown unknowns — type systems, environment management, things you don't know exist until you hit them. I picked up most engineering skills by building, solving real-life problems, and asking good engineers a lot of questions. I learned Scala, Python, JavaScript and a bit of TypeScript along the way. I still don't see myself as a professional engineer, but I can work on most engineering tasks proficiently. And it's much easier now with AI. 

My current company, 477 Studio, is the AI-native service business we're trying. We do three things here: build managed-agents for clients, run AI consulting for tech and non-tech teams and embed with client teams as Forward Deployed Engineers (FDEs) for AI development tasks. Unlike the previous two companies, the work now is different every week — some weeks we build the whole time, others are large-team discovery sessions with clients. 

What did you need to unlearn from your actuarial training when you moved to a startup? 

Actuarial training gave me a lot, especially a quantitative thinking framework. The habit that helped most is being the person who always asks for reconciliation. When two things can't independently confirm each other, you don't move on until you can explain the difference — and in a startup that's usually where the real answer is. 

If you run the calculation properly, you never start — the chance of the business surviving a few years is small, the pressure is personal and on expected return alone it rarely makes financial sense. You have to be able to make the bet anyway, and be honest that it is a bet. 

When did your actuarial background clearly help your decision making at Hyper Anna or Curious Thing?

A small but interesting example: we had a self-serve product, most of our users in the US, and we had grown to about 10k users in six months. Growth was good, but churn was really high. 

The team's read was that it was an onboarding problem — first-week churn was very high. I wasn't convinced — we had a small amount of feedback saying the opposite that our onboarding was easy. Small sample, but it didn't feel right. 

Before we spent time rebuilding the onboarding process, I spent few hours on the churn triangle, by cohort and by all possible dimensions from our marketing and usage data. The first-week churn wasn't spread evenly. Almost all of it sat in the cohorts from one channel we had switched on and off — it brought people who came to look, not to use the product. 

On the surface it was an onboarding problem. Underneath it was a channel problem. Nothing about that analysis was complex — it's the triangle every actuary has built a hundred times, pointed at a different question. But it stopped us solving for the wrong thing. 

Not everything in a start-up goes to plan. What has gone wrong along the way, and what did you take from it? 

As a founder, and now also an angel investor, I would say most things in a startup don't go according to plan. Almost everything that can go wrong will go wrong, unfortunately.  

The way it actually works is that you have a long-term goal, you make a plan to get there, and then you change that plan constantly as new information arrives, or as the business reality turns out to be different from what you assumed. Usually what stays fixed is the goal — a North Star with a metric attached. Then you focus on the short-term problems in front of you (the goal can change too — that’s what we call a "pivot"). The last thing you want is a master plan you are determined to follow step by step. 

What do actuaries consistently get wrong about themselves? 

Most actuaries wouldn't identify themselves as builders. We didn't go through structured technology and engineering training, and as a profession built on precision, we default to not wanting to build things where our knowledge isn't perfect.

So two things I'd recommend. Start building with AI — it can be work-related, or even a personal project. And I would encourage all actuaries to learn how to use Claude Code or Codex. You will be surprised how much you can achieve with them. 

Your next step

Read more about the Actuary of the Year award , or explore the Institute's AI and data science resources to take Sam's advice further.