Credit: InfiniteFlow/Adobe Stock
Artificial Intelligence will transform the insurance industry soon. Yet, so far, most enterprise AI initiatives have not delivered value. Why is that, and how are successful AI projects different?
A recent report from Fortune and MIT on generative AI deployments found that only about 5% of AI pilots achieved rapid revenue gains, while roughly 95% stalled or produced little measurable business impact. Meanwhile, an S&P Global survey in 2025 showed that 42% of companies abandoned most of their AI initiatives and scrapped nearly half of all proofs of concept before reaching production.
While the hype reflects a true AI revolution, these disappointing results highlight an expensive failure risk, such as wasted capital and unmet expectations in underwriting and contracts. To capitalize on what AI brings, companies must understand how to mitigate that risk.
What’s actually failing—and why it matters to insurance
AI is not failing because models and agents don’t work. It’s failing because organizations cannot embed new technologies into everyday operations. Over 80% of AI projects fail because of cost overruns, data privacy concerns, and security risks, largely due to underprepared or resistant workers. Experts attribute about 70% of failures to people- and process-related issues, not technical ones.
For insurers, that translates directly into project risk, regulatory risk, and even professional liability exposure when firms overpromise AI capabilities to clients or regulators. When a vendor, carrier, or broker touts a revolutionary AI solution, it should be viewed with healthy skepticism.
Patterns behind failed AI pilots
Enterprise AI failures follow a familiar pattern insurers know from other large programs. First comes hopeful pilots: projects live in safe sandboxes where technology appears impressive but fails in the real world due to compliance issues or lack of user training.
Then, engineering teams fall in love with the latest models, optimizing new features while overlooking data quality, workflow redesign, and governance (issues critical in financial and regulated industries).
Next comes organizational friction. Product, infrastructure, data, and compliance teams operate separately with different priorities, so nobody owns the AI process end to end. For insurers, that can mean a claims triage model that “works” in a lab but is unusable for adjusters; an underwriting assistant that can’t pass model risk management review; or a generative AI output that conflicts with outside counsel’s expectations.
Lessons from the 5% that succeed
The AI projects that succeed follow a playbook highly relevant to insurance leaders. Successful teams start with clearly quantified business metrics, like call handling time or quote turnaround, before discussing models. In other sectors, companies like Lumen Technologies and Air India began with specific operational bottlenecks, then designed AI solutions to address them, achieving tens of millions in annual savings.
Insurance organizations can mirror this by targeting a defined reduction in claim-handling minutes or a measurable improvement in quote-to-bind conversion rather than imagining a broad, company-wide AI transformation. Industry surveys also cite data quality, readiness, and training as key points of failure. So it is not surprising that successful programs devote 50–70% of budget to data cleaning, preparation and governance.
Human–AI collaboration and the missing role: AI translators
The most durable AI deployments are built for human–AI collaboration, not full automation. In other industries, AI improves fraud detection and sales productivity by combining AI suggestions with human judgment. In insurance, adjusters, underwriters, and claims counsel should retain final authority over coverage and settlement decisions.
Here a newer role also becomes vital: the AI translator (often evolving from the analytic translator). This professional speaks the language of both AI and business, bridging goals, data constraints, and regulatory requirements. AI translators help define workflows, explain risk, and ensure all stakeholders agree on the balance of AI and human control.
What insurance professionals should do now
Begin with a narrow focus. When a carrier markets AI-enabled underwriting or service, ask for the specific business problem it solves, how it’s measured, and what guardrails exist around bias, privacy, and error handling. The explanation should be easy to understand.
Moving forward, start every AI proposal with a quantifiable operational or financial outcome. Budget more for data and integration than model experimentation. Design workflows where AI assists, rather than replaces, expert judgment. Most importantly, invest in or cultivate AI translators—people who can connect actuarial, legal, regulatory, and technical perspectives so AI reduces, rather than introduces, risk.
Finally, lawyers and claims professionals should anticipate AI will likely solve problems while also creating new ones. The same governance discipline that separates the 5% of successful deployments from the 95% that fail (i.e., clear ownership, monitoring, and human oversight) will also separate organizations that use AI to reduce litigation from those that inadvertently create it.
Dr. Wendy Lynch is a researcher, author, and long-time “analytic translator” who helps organizations turn data and AI into decisions that actually stick. For more than three decades she has worked with Fortune 100 companies to bridge the gap between business leaders and technical teams, improving how questions are framed, results are communicated, and value is realized. Find her training at Analytic-Translator.com.
Opinions shared in this piece are the author's own.
See also:
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.