We handed the controls of a virtual insurance company to student-built AI agents — then watched what happened over a simulated 20 years. The question underneath: what happens when AI stops being a tool and starts being a decision-maker?
We chose insurance deliberately. It's one of the most complex decision environments imaginable — every choice ripples across multiple stakeholders at once.
Premiums drive customer acquisition. Coverage affects profitability. Claims shape trust. Service quality drives retention. Risk determines survival.
An insurer must constantly balance growth, fairness, sustainability and risk. In other words — a miniature model of almost every modern organisation. If AI can navigate this, it can navigate logistics, healthcare, finance, public services.
Each agent had to make decisions across six competing priorities, every simulated year.
Win and keep policyholders in a shifting market.
Set premiums that attract without inviting ruin.
Pay fairly, fast — without bleeding the balance sheet.
Survive shocks the market throws without warning.
Keep trust high enough to retain for decades.
Still be standing in simulated year twenty.
Some teams built strict decision frameworks. Others gave the AI room to interpret. Some companies thrived. Others went bankrupt. All of them taught us something.
The AI wasn't just following instructions — it was developing strategies, trade-offs and behaviours nobody had explicitly programmed. Some agents chased aggressive growth into unsustainable risk; others turned so conservative they protected capital at the cost of any growth at all.
The most successful systems were rarely the ones with the most detailed instructions. They were the ones given clear objectives, sensible constraints, and room to adapt.
Agentic AI doesn't automate your existing process. It has the potential to redesign it entirely.
Many teams assumed better prompts would simply produce better outcomes. The reality was messier. Small changes in instruction sometimes produced dramatically different companies. Two models under near-identical objectives arrived at entirely different strategies.
This isn't an argument against trusting AI. It's a reason to design systems that can safely accommodate the unexpected while staying aligned with strategic goals.
As AI systems grow more autonomous, predictability gets harder — not easier.
The most important lesson had nothing to do with AI performance. Without guardrails, pure optimisation produced unintended consequences: an AI chasing profit became unfair; one chasing satisfaction became unsustainable; one chasing growth courted catastrophe.
The strongest teams understood that success wasn't about the smartest AI — it was about the most trustworthy system. Ethical guardrails aren't a compliance checkbox; they're a first-principles design decision, built in from the very beginning.
The winning teams didn't build the smartest AI. They built the most trusted one.
It was designed to help us ask better questions — the ones that will shape the future of every industry.
Insurance was the first model. Bring us a problem statement worth exploring, or co-host the next Hack to the Future with your organisation or campus.
