The models are ready. The infrastructure often isn't. And the gap between the two is where most AI investments go quiet.
Not long ago, I was in a discussion where everyone agreed the AI was working. The model accuracy was where we expected it to be. The pilot results looked promising. Yet nobody wanted to expand the program.
In insurance, AI rarely fails because the technology doesn't work. It fails because the organization around it isn't ready for what the technology requires.
The issue wasn't the algorithm. It was everything around it.
The Operational Intelligence Gap
The insurance industry has spent the last several years building AI competence at the model level. Carriers have invested in underwriting algorithms, fraud detection engines, claims triage tools, and customer-facing automation. Much of that work is technically sound.
But technical soundness isn't the same as operational readiness. And in insurance, where decisions carry regulatory weight, customer relationships, and actuarial accountability, the operational layer is where value is either realized or lost.
The failure mode I keep seeing isn't model accuracy. Over the years, I've started thinking about this as an operational intelligence gap: the difference between a model that performs well in testing and a business capability that can be trusted, governed, and sustained in production. That gap has four dimensions. In my experience, most organizations focus on only a few of them.
Places Where AI Loses Its Value
Part of that gap is data.
Insurance organizations rarely operate from a single source of truth. Policy systems, claims platforms, CRM environments, external feeds—each evolves at its own pace. AI doesn't break when those systems disagree. That's the problem. It continues to produce answers. Some of them are even convincing.
But data isn't where most projects get stuck. More often, the problem appears later, when AI has to fit into an existing workflow.
A claims adjuster receives a recommendation. An underwriter receives a score.
The model may be right. Yet if the recommendation arrives too late, or in a format nobody actually uses, the value evaporates surprisingly fast.
What This Looks Like in Practice
A few years ago, I led the technology architecture for a communications platform at a major insurance enterprise. The system eventually handled roughly 80 million customer communications annually across digital and traditional channels. The core AI components performed well in testing. The models did what models do.
What nearly derailed the program had nothing to do with the models. It was the integration layer between real-time AI routing decisions and a legacy policy system that updated on a 24-hour batch cycle. The AI was making decisions based on customer state data that was, by definition, always a day old.
Fixing that required infrastructure investment that wasn't in the original scope. What surprised me was how little of that conversation involved the AI itself:
Looking back, I don't remember many conversations about model accuracy. I remember conversations about budgets. Ownership. Which team would take responsibility when something went wrong.
When we addressed those layers, not just the algorithm, the results shifted materially. Complaint volume dropped by 90%. Delivery speed improved by 96%. The AI wasn't different. The operational architecture around it was.
Insurance Has Always Been a Trust Business
Insurance was a trust business long before it became a technology business. Customers trust insurers with their financial security. Regulators expect decisions they can explain and defend. Employees have a different concern altogether: they want technology to support their judgment, not quietly replace it.
Consider a claims environment. An AI model may identify potentially fraudulent claims with impressive accuracy. Yet if investigators cannot understand why a claim was flagged, or if the escalation process is unclear, the organization faces a difficult choice: trust a recommendation it cannot explain or ignore a recommendation it cannot defend. In insurance, that tension often matters more than the model's accuracy score.
Trust Changes Everything
Regulatory expectations are also changing. Across major markets, insurers are facing growing pressure to explain how automated decisions are made, monitored, and governed. Regulators are asking harder questions than they were two years ago. That shift is real, and it's not slowing down.
Some of the most mature programs I've seen weren't built by organizations with the most advanced models. They were built by organizations that invested early in governance.
When I look at AI programs that struggle, the same questions keep coming up.
The first question is, “Can we trace any AI-influenced decision back to the data that produced it, the model version that processed it, and the human who is accountable for it?” If the answer is no, the program has a governance gap that will surface at the worst possible moment.
Another question worth asking is, “Have the people closest to AI outputs been involved in designing how those outputs reach them?” Not briefed after the fact — involved in design. The distance between those two things is usually the distance between adoption and shelf life.
And the third one is, “Is the integration layer between AI systems and core operational data treated seriously, or as a technical detail to be handled later?” In every AI program I've seen fail quietly, the integration layer was a detail. In the ones that scaled, it was a strategic decision.
Beyond the Algorithm
I've stopped being surprised when a technically successful model struggles to create business value. I've stopped being surprised when a solid model fails to move the needle. By now I know where to look, and it's never the algorithm.
Most insurance organizations don’t need more AI pilots. They need the discipline to turn what they’ve already built into something that actually gets used by real teams, in real decisions, under real pressure.
Most of these organizations already have AI. What they're missing isn't access. It's the organizational work that makes AI usable, and that work doesn't show up in a vendor demo.
