Last week I sat across from the global AI lead of a specialty carrier. His team did what most carriers haven't: asked the whole company, globally, how AI could help. They got back more than two hundred ideas.
Two hundred ideas. And no decision.
The bottleneck is no longer building. It is choosing. Every carrier I meet has run the pilots and proved the concept. Nobody has answered which of the two hundred things to do first, who decides, and what "done" looks like. The intake form was the easy part. The prioritization is the work, and it is unstaffed.
Across my conversations with chief AI officers, I keep hearing the same three problems — data, boundaries, and scale — in different words.
First, data. His enterprise data strategy has stalled for two years — not for lack of technology, but for lack of continuity: leadership changes, changes of opinion, IT living separate from the business. Two years is not a delay. It is a decision nobody has made — and every idea on the list inherits it.
Second, boundaries. What should AI do, and where should it stop? How does security get involved, and how much is enough? How do you move a heavily governed SDLC fast enough to matter? The boundary question is a negotiation re-run on every project, and nobody has written down who wins it.
Third, scaling. Growing the team, having operations ready for the day a tool reshapes forty percent of someone's job, and keeping builders on work that matters. His executives are fully bought in and funding the work — and they want it yesterday. His job is to stay realistic about what the team can deliver without breaking the promise that bought the funding.
What it means for carriers
First, the idea pile is the new pilot pile. Two years ago it was a hundred proofs of concept and nothing in production — nearly half of carriers still report AI innovation below expectations. Now the disease has a new form: two hundred ideas in a spreadsheet and no mechanism to choose among them. The failure mode moved upstream, from the build to the decision.
Second, every one of those three challenges is a decision problem wearing a technology costume. A stalled data strategy is a leadership decision; a boundary dispute, a governance decision; a scaling question, an operating-model decision. Carriers keep buying technology to solve them, and the technology keeps arriving to find the decisions still unmade. It is the same pattern this series has tracked all year: the submission that waits because nobody made it readable; the renewal decided on half the picture. The work is never the model. The work is the decision the model is supposed to serve.
Third, the economics question is now the first question, not the last. His team is asking what every token leaves behind — what value compounds, and what the cost per decision actually is. That is the right question, and most AI programs still cannot answer it. A pilot that cannot state its cost per decision cannot be scaled, because scaling it just scales the unknown.
From the operator's chair
One: put every idea through the same three questions before it gets a team. What is the production problem — where does the work break today, with volume and exceptions? How is it governed — who owns it, how is it explained, where does human judgment sit? And what could you demonstrate in ninety days — one workflow, one measurable outcome? An idea that cannot answer all three waits. Two hundred become twenty. Twenty become three.
Two: name one owner for the data decision. Not a committee, not a program — a person, with a date. A two-year stall ends the day someone is accountable for it.
Three: underwrite the controls, not the model. Inventory every deployment, name an owner, set the guardrails, keep the human in the loop where the risk is high. The carrier that shows that discipline in its own operations can underwrite it in others — and answer the examiner, the auditor, and the board with the same record.
Four: price the token. Know your cost per decision — inference, review minutes, rework, exceptions, the price of being wrong. If you cannot state it, you are not ready to scale, whatever the pilot showed.
Five: prove one thing in ninety days. One high-volume workflow, one measurable outcome — a demonstration, not a roadmap. The team that can do that once can do it again. The team that cannot will still be choosing among two hundred ideas next year.
Two hundred ideas is not a pipeline. It is a symptom. The carriers that win the next three years will not have the most ideas. They will be the ones that learned to decide.
— Krishna