Marcin Nowak, board member at Decerto, has 20+ years in insurance, focusing on automation, technology impact and software solutions.
getty
Since 2006, I have been building software for insurance carriers. In those 20 years, I have watched four transformational technologies arrive: service-oriented architecture, big data, the cloud and robotic process automation. Each promised to redraw the competitive map; each delivered value to carriers that knew what they were doing and failed in those that did not.
AI is the fifth wave, and the script is following the same pattern. The technology is genuinely powerful, and I use it daily. But the rule that separated winners from losers in every previous wave still applies, and the speed AI gives you has made it harder to follow.
The Standish Group’s original CHAOS report, which tracked IT outcomes and came out in 1994, showed that only 16% of projects came in on time, on budget and on scope. By the 2020 edition, the success rate had climbed to 31%. The technology kept changing, but the pattern of failure did not. What separated success from failure was almost never the tool, but whether someone had genuine clarity about the business problem and the team building the solution understood it at the same level. The tool was always a distant second.
Stack Overflow’s 2025 Developer Survey of 49,000 developers found that 84% are using or planning to use AI tools, up from 76% the year before, and 51% of professional developers now use them daily. Google and Microsoft report that AI writes around 30% of new code at their organizations, and Gartner forecasts that "75% of enterprise software engineers will use AI code assistants by 2028." The cost of validating an idea has fallen dramatically: a prototype that used to take six weeks and a senior engineer can now be built in two days by a business analyst.
The speed of building has not changed the rate at which the wrong things get built. McKinsey’s State of AI survey found that although 88% of organizations use AI in some capacity, only 39% report any EBIT (earnings before interest and taxes) impact at the enterprise level. Gartner separately predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of last year. In insurance, Grant Thornton’s 2026 AI Impact Survey found that 44% of insurance leaders said governance or compliance issues caused their AI project to fail or underperform.
The RAND Corporation’s August 2024 report, based on interviews with 65 data scientists and engineers, identified five root causes of AI project failure, all organizational rather than technical. The most decisive is the first: organizations begin by forcing a solution on the wrong problem. Instead of asking what business problem needs to be solved, they ask: “Where can we apply AI?” Rather than intentionally selecting a tool that will address an issue they have, they adopt a technology and then search for a way to use it.
The clearest example I have seen recently was a model with 94% predictive accuracy that never made it to production. The predictions arrived three days after the underwriter had already made the decision. The data science team did not realize that underwriters typically decide on the same day a submission lands. That is not a model failure but a process failure: a beautiful tool aimed at the wrong moment in the workflow.
The expertise gap behind such cases is not a shortage of machine-learning talent, but a shortage of people who understand both the technology and the business it serves. Insurance is not just rows and columns. It is a process with regulatory boundaries, a renewal rhythm, claim adjusters who pick up the phone instead of typing and endorsement types that arrive late on a Friday. None of that lives in the training data.
There is a distinction I rarely hear made in boardroom conversations about AI. In vibe coding, a person describes the outcome and the AI takes the lead on libraries, structure and technical decisions. It is excellent for testing whether an idea has merit, but rarely production-ready. In agentic coding, the team knows what it is building: the problem is defined, the architecture intentional, the constraints documented. The AI implements under supervision; the human leads.
The difference is not the tool. Cursor, Claude Code and Copilot are the same software in either mode. The difference is who knows what is supposed to be built. The trap of 2026 is treating a vibe-coded prototype as the output of agentic work and rushing it to production. Cognizant’s 2025 research found that 85% of senior executives at Global 2000 organizations are concerned their existing technology estate will impede AI integration. They are right to be worried.
Discipline starts with three questions:
1. What will stop happening when we deploy this? If nobody can answer, there is no business problem yet, only an idea.
2. Could a rules engine or simple automation handle this? If yes, that is the right tool. AI belongs where structured approaches fall short: unstructured documents, pattern recognition at scale, classification problems that defy explicit rules.
3. Who owns this in production? Without an architect accountable for integrating the result, the proof of concept becomes technical debt the moment it ships. Most organizations reverse this order, and today’s failure rates are the cost of it.
AI is the fifth wave of transformational technology in insurance, and it won’t be the last. The carriers that get real value will approach it the way they should have approached the previous four: by understanding the business problem before reaching for the tool, and by treating production as engineered rather than assumed.
The pressure to be seen doing something with AI is louder than for any prior technology, but the rule that decides which projects pay back is the same. It just matters more now.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。