Insurers Misprioritize AI Over Design, Hindering Adoption
Insurance Thought Leadership published a July 9, 2026 op-ed arguing insurers have a design problem, not simply an AI problem, when deployments fail to fit real workflows. The author, Cake & Arrow CEO Josh Levine, says insurers often buy models and automation before understanding agents, customers, employees, and brokers who will use them. A related ITL item citing Camunda says only 11% of agentic AI projects reached production last year. For practitioners, the lesson is operational: model choice matters less than user research, trust signals, exception handling, and workflow redesign that make AI outputs usable in claims, underwriting, and service.
The useful practitioner takeaway is that insurance AI failures often look like product and operating-model failures before they look like model failures. If a copilot or agent lands in a broken claims, sales, or servicing workflow, better model accuracy may not translate into adoption, trust, or measurable value.
What happened
Insurance Thought Leadership published a July 9, 2026 op-ed by Josh Levine, founder and CEO of design agency Cake & Arrow. Levine argues that insurers risk wasting AI investments by prioritizing models, platforms, copilots, and automation before understanding the workflows and needs of the people expected to use them. A related Insurance Thought Leadership item cites Camunda's State of Agentic Orchestration and Automation 2026 research, reporting that only 11% of agentic AI projects reached production last year.
For practitioners
The operational lesson is to design AI work around users, exception paths, controls, and trust signals from the start. Claims adjusters, underwriters, brokers, and service teams need interfaces that explain uncertainty, preserve handoff points, and fit the cadence of their existing work. A technically impressive model can still fail if it creates extra review burden or hides why it produced a recommendation.
Industry context
Insurance has high regulatory, fiduciary, and customer-trust constraints, so AI adoption depends on governance and human-centered process design. The same pattern applies beyond insurance: pilots move to production when owners can measure sustained use, decision quality, cycle-time improvement, and safe escalation, not simply when a model demo looks persuasive.
A practical implementation sequence starts with the decision and user, not the model. Teams can map the current workflow, identify high-friction handoffs, define which judgment remains human, and test whether an AI-supported path reduces time without increasing corrections or escalations. That makes adoption evidence observable instead of relying on a demonstration or procurement milestone.
Insurance teams should also segment results by workflow and user group. A tool that helps experienced adjusters summarize files may perform differently for new staff, brokers, or service teams. Review rates, override reasons, complaint signals, and exception outcomes can reveal whether low usage reflects training, interface design, weak output, or a process that should not be automated.
What to watch
Watch whether insurers pair AI-agent pilots with user research, process redesign, and measurable trust metrics. The stronger signal will be production adoption in claims, underwriting, and servicing workflows, not announcements that a carrier has purchased a new model or automation platform.
Key Points
- 1The article frames insurance AI adoption as a workflow-design problem, not simply a model-selection problem.
- 2The cited production gap reinforces that trust, controls, and orchestration determine whether pilots become deployed systems.
- 3Practitioners should measure sustained use, exception handling, and decision quality alongside model accuracy and automation speed.
Scoring Rationale
The item is a practitioner-relevant insurance AI adoption essay with useful operating lessons, but it is not a new model, regulation, funding round, or platform release. The score is solid rather than high because the evidence is mostly thought-leadership and adoption research context.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems
