The Underwriter's Nose: Kenneth Stanley and the Search for Value

In our first two articles, we looked at two major hurdles facing the insurance industry. First, in The Real Physics of Automation, we saw how digital initiatives get bogged down in manual data checking, stealing up to 40% of an underwriter's day. Second, in Moravec's Mirror, we looked at the mathematics. We saw how autonomous AI models inevitably drift and make sequential errors when they are cut loose in complex areas.

The industry's favourite fix for this drift is the "Human-in-the-loop" model. But in reality, this is just a fancy rebranding of the old administrative waste.

Instead of freeing underwriters from tedious data entry, we have simply shifted them into a new kind of digital prison. Underwriters are now tasked with chasing AI hallucinations, checking black-box errors, and playing the role of highly paid compliance janitors. We have traded a legacy data bottleneck for a digital audit bottleneck, continuing to waste elite human minds on low-level chores.

To understand how to automate safely without turning our best thinkers into glorified error-checkers, we have to look at a concept from evolutionary biology. Professor Kenneth Stanley calls it the Principle of objective-less discovery.

The Seduction of the Goal

In his book Why Greatness Cannot Be Planned, Professor Kenneth Stanley argues that having a rigid, measurable goal actually stops us from discovering new things.

Think about how human music evolves. If we had managed musical history using a strict computer algorithm tasked with maximising "rhythmic efficiency," we would never have invented jazz. Jazz was not a pre-planned target. It was the unpredictable byproduct of a chain of interesting "stepping stones"—spirituals, ragtime, and the blues. Each step was chosen not because it looked like the final destination, but simply because it felt interesting to the musicians at the time.

This is the big secret of discovery. The stepping stones that lead to something wonderful and complex do not look like the final product. The vacuum tube was invented to boost radio signals, not to build computers. But without the vacuum tube, we would not have computers. The moment you set a rigid, metrics-driven goal, you shut down the very stepping stones you need to get there.

As Stanley pointed out in his interview on Machine Learning Street Talk (MLST), human curiosity and a sense of "interestingness" are much more reliable guides through complexity than any automated checklist.

The Underwriter's "Nose"

Now, let's bring this principle back to the specialty insurance market at Lloyd's of London.

In the real world, an underwriter's job is not a clean maths problem. It is a messy mix of digital archaeology and old-school diplomacy. Day-to-day, they must navigate decades of clunky legacy IT systems, piece together fragmented historical data, and maintain deep human relationships.

An underwriter looking for a highly profitable, unusual risk does not find it by ticking boxes on a checklist. They navigate a sea of messy, unstructured information using what the market calls a "nose for value." This is an intuitive, highly refined ability to spot when a risk is priced wrong. It is the exact human equivalent of Stanley's search for "interestingness."

Why is this human "nose" so important? Because specialty insurance is constantly shifting. Geopolitical events, volatile economic conditions, and changing weather patterns rewrite the rules of risk every single day. A static algorithm cannot adapt to these shifts in real time without constant, incredibly expensive retraining.

Even more importantly, specialty risk is solved through human-centred transactions. It relies on trust, negotiation, and the subtle reading of a partner's risk appetite. You cannot automate a relationship.

Because large language models operate purely on statistical correlation and—as we established in Moravec's Mirror—completely lack an internal "world model" to simulate reality, they lack this intuitive human compass. Letting them run on autopilot in bespoke markets is a direct path to regulatory ruin. When we try to patch this by forcing underwriters to sit downstream of the AI—manually checking every sentence for errors—we do not solve the efficiency gap. We just standardise it.

The Organic Alternative

This brings us back to the conversational ecosystem of Lloyd's. Specialty markets work precisely because they leave room for human-to-human negotiation to handle complex, shifting realities.

If we want to automate safely, we must abandon both the dream of autonomous AI agents and the nightmare of the human-as-error-checker. We have to move toward a framework where the human remains the master architect upfront. We need a system where the expert's intuitive intent is captured and mapped into clear, version-controlled logic, allowing the AI to do what it actually does best: translate that clear human intent into flawless execution.