In 1966, the philosopher Michael Polanyi wrote that we know more than we can tell. It is the best explanation I have found for why end-to-end automation in insurance is so rare. Every agent can read an ACORD form now. Few can handle the first exception that depends on knowledge nobody wrote down. That exception goes back to the manual queue.

Two things decide whether an agent gets past that exception: a learning method in which experts correct the agent on real cases and their reasoning is captured, and governance built on the operational controls carriers already run. Both are hard to see when assessing options. Nearly every proposal will also offer a way to get started, from existing procedures or a pre-built agent, and a way to keep the agent current as procedures change. Those matter, but they are table stakes. Research on how experts hold knowledge and how people apply it to govern processes explains why it’s the knowledge transfer and oversight that are the most important.

The agent has to learn from your experts

MIT economist David Autor showed that Polanyi's observation is the reason automation stalls. Computers take over what can be written as rules and struggle with the unspoken tacit knowledge that that never makes the requirements document.

Large language models got partway around the problem by learning from enormous volumes of published examples. That is why reading an application or summarizing an attending physician statement is now effectively solved. A model learns only what it has seen, though, and the knowledge that matters most in insurance operations was never published. Sociologist Harry Collins calls it collective tacit knowledge, which is held in a group's shared practice. It covers how a 1035 exchange is handled when surrender proceeds arrive short, which trust language Legal accepts, and why one type of beneficiary change gets a second review. A general-purpose agent that runs too far from the experts stops at the first decision that depends on it.

Interviews and documentation will not surface this knowledge, because experts cannot write down what they do not consciously know. Insurance already knows how it transfers. A new examiner works real files beside a senior colleague, who corrects decisions and explains why. Agents can learn the same way. The expert overturns a decision on a real case and explains the reason, and the agent incorporates that reasoning into the procedure it follows. The procedure stays readable to the people accountable for it.

Oversight depends on staying close to the work

Engineering psychologist Lisanne Bainbridge described the ironies of automation in 1983. When people stop doing the work and start watching it, they lose the practice that made them able to judge it.

Insurance has managed this for decades. Managers who approve exceptions often no longer work those transactions closely enough to assess them, so carriers do not rely on approval alone. Authority limits define what each person may decide on their own. QA reviewers rework a sample of complete files, which keeps their judgment current. Referral rules send the hardest cases to the people who still handle them every day. Agents create the same dynamic, and the same controls apply with the addition of a detailed audit of every micro decision and why. Sampled review of whole cases keeps experts working on real files and shows them how agents fail so they can better improve them in the future.

The question to ask

Ask someone how their agent learns what your experts never documented, and how your existing controls will govern it. The answers tell you whether you will get end to end or another queue.