Your AI rollout could cost you the experts it was meant to help

An adjuster opens a claim file that an AI tool has already summarized. They can’t see how it reached its recommendation, so they rebuild the file by hand. In Camunda’s new research, The AI Process Gap: Insurance, 43% of insurance employees who use AI or automation have had to manually override AI outputs because the underlying process wasn’t set up correctly.

Almost every insurer is trying AI, but mostly in isolated pilots that don’t change how work gets done. Allianz’s first agentic system uses seven AI agents on small food spoilage claims in Australia, with a claims professional deciding every payout. The processes these AI agents join were built for people, from claims handlers to agents and brokers, with a manual handoff at every step.

That makes now the time to change the operating model, before AI agents scale. Camunda calls this The Great Process Re-Engineering, and in its view every process built for people alone is legacy, even the ones built in the last few years.

Insurance employees are working around the AI meant to help them

Among employees not seeing significant time savings from AI, 42% selected distrust of the output, so they double-checked everything by hand.

Most weren’t asked how the work should change: 73% were not fully consulted on how AI would support their day-to-day work. AXA chief executive Thomas Buberl told investors that AXA’s insurance agents are asking for faster access to AI tools, and warned: “If you push AI on people, nothing will happen.”

When experienced adjusters leave, their judgment leaves with them

Four in ten surveyed employees would consider changing jobs over poorly implemented AI. Sedgwick expects nearly a quarter of claims adjusters to retire by the end of 2027, and carriers with high turnover among experienced adjusters have seen operational costs rise 12%.

What those adjusters know rarely lives in a system: which claims look clean but aren’t and need escalation and a deeper look. Orchestration keeps this knowledge in the business.

When claims and underwriting run as orchestrated processes, experts’ referral rules, approval thresholds and exception paths become part of the process model, and every decision and handoff is recorded. The next adjuster, and the next AI agent, follows the correct path and policies.

Because every step, action, and decision is tracked and machine-readable, you’re able to utilize this information to help expose this to further improve processes and AI agents helping to support the workforce.

Strong margins won’t last as long as re-engineering takes

US property and casualty insurers closed 2025 with a 92.9% combined ratio, which Verisk credited mostly to unusually low catastrophe losses.

That good fortune isn’t something you can price in long term. Global instability. Supply chain disruption. The growing risk of increasing climate volatility. All of these risks have the potential to compound and collide in a way that erases the buffer. AM Best expects the ratio to rise in 2026 as rates soften. Using the previous good years to help invest now will set up some carriers to weather the storms better than others.

Many industries are facing this same challenge of reimagining operating models quickly while ensuring that business keeps moving. Considering the radical shift required, and the sheer volume, the clock has already started. Insurance decision-makers surveyed expect to need more than four years on average to redesign their most important processes for AI. Camunda estimates one process has traditionally taken about 12 months, and a typical enterprise runs more than 500 core processes.

At that pace, four years covers four processes. Waiting for tighter margins means starting under pressure, so each process has to take far less than a year.

Phased migration lets you re-engineer without stopping the business

Among insurance decision-makers, 82% selected migration-driven re-engineering (moving off legacy platforms while delivering AI returns) as an investment priority. Design the whole process first, across the systems running it today. Then replace what isn’t serving the business one piece at a time, while claims and underwriting keep running.

Because the process logic sits above any core platform, replacing a system, revising specific rules, or updating a front-end UI changes only in the background.

Agentic AI needs a foundation built for insurance scale and complexity

An insurer rarely has one claims process. It has one for each product line, channel and jurisdiction, from FNOL to settlement and submission-to-bind, plus variations from every acquired book.

AI agents can only work safely across that on the right foundation: one orchestration layer coordinating AI agents, people and the core systems insurers already run, with each AI agent following the same referral rules and approval thresholds as the adjusters beside it, and every decision recorded for regulators.

ProcessOS, now in early access, applies AI to the redesign itself. It takes a process from discovery to a re-engineered version running in a test environment in 10 to 12 weeks, against about a year the traditional way. It ingests operational information you share, proposes a redesign, scores it against the metrics you set. Then it builds what the new process needs to run, with human review before anything reaches production.

Start where employees are drowning in menial data entry and re-keying: work that must be done correctly and often takes judgment and experience. Keep most processes predefined for cost, performance and consistency, and use agentic AI for the messy middle that calls for reasoning, which is also where experts’ knowledge is most at risk of leaving as the retirement cliff approaches.

Get the full findings.

The AI Process Gap: Insurance surveyed insurance decision-makers and employees across the US, UK, Germany and France, including where AI budgets go today and which processes insurers plan to redesign first.

Download the report