Six months ago, the CEO of a major European insurance carrier sat down with his operations leadership and asked them to rethink the whole thing from first principles.
"If you start rethinking operations, how would you do it? What will change?"
That carrier is HDI, part of the Talanx group, one of the largest insurance carriers globally. What they found says something uncomfortable about where the industry stands: the machinery built over two decades works beautifully on the processes it was designed for, and hits a hard wall on everything else.
The last 25% is where the economics struggle
HDI has automated process work for more than 20 years. Its high-volume operations run on Camunda with roughly 10,000 processes for the standardized, repeatable transactions classical automation was built to handle. By the carrier's own account, more than 75% of its business transactions are automated.
Then there is the rest.
The remaining work is the long tail: processes that run a few hundred times a year, each with its own variants. Yet these are the processes that consume employee time, increase costs, and frustrate customers. Automating the classical way, meaning analyze it, model it, build it, test it, deploy it, takes over six months even with an agile team. For a process with 200 to 500 annual instances, the math never pays back. HDI counts 13,598 multichannel process variants. Modeled deterministically, the carrier estimates that would require more than 7,000 individual orchestrated processes. "You can't handle it," as their operations leadership put it, "and it's too expensive."
This is not a staffing or tooling problem. It is structural, and it is not unique to HDI. McKinsey projects that by 2030, insurers will automate more than 90% of pricing and underwriting for simpler personal and small-commercial lines. The flip side is the work that resists standardization: complex, low-volume, judgment-heavy cases where carriers still handle many workflows manually despite sitting on enormous pools of unstructured data. The straight-through game is largely won. The value now sits in everything that was never worth automating one at a time.
The bottleneck is the conversation
A property ownership change shows why the long tail resists automation. On paper it has four clean, linear steps: the seller notifies the insurer, the buyer provides identification, the land registry confirms the transfer, the contract is reassigned. Until it hits the real world.
A homeowner passes away, and an agent contacts HDI on behalf of the widow to start a policyholder change. Mid-process, the daughter takes over: same process, different person, different authority. Months later the house is sold, so the contract transfers to the buyer, which opens an ownership change and a request for missing documents. Then the mother moves into a nursing home and passes away. The seller is no longer one person. It is a community of heirs, any of whom can call, write, or submit documents, and whose legal authority all has to be reestablished.
The case you started is no longer the case you are finishing.
A single ownership change runs six to nine months and generates an average of 6.5 internal business cases, with communication scattered across email, postal mail, phone, and online portals. The transaction was always simple. The expense lives in identifying who is allowed to act, chasing missing information, and holding one thread of context as the case mutates, which is where classical automation struggled.
What HDI built
HDI's answer was a framework for generating and running processes, instead of a faster way to build them one at a time.
It starts at intake. Most customers still arrive by email. A classification layer reads the request, identifies the intent, and routes it to the right business process regardless of channel. Today an email can wait two to three weeks for a first reply while it queues behind case handlers. The target HDI describes is a response in minutes.
For the long tail that never justified six months of bespoke work, a process-expert AI agent reads what the carrier already knows. HDI has more than 1,000 processes documented in its CRM, years of letters, replies, and work instructions the team calls its gold, and from that material the agent generates a working process description on demand. HDI has run it across 50 processes, and it is in internal production now.
A data-handler step checks that everything a process needs is present and requests anything missing on the spot. Structured, repeatable cases run deterministically. Variant- and context-heavy ones run through an agentic workflow, where you define the goal of the process and let the agent decide the steps.
Both run on one unified platform,rather than separate tracks.
Two controls keep it accountable: a scoring agent that lets a process run on its own only after it has run correctly 100 times under human review, and outbound templates the agent fills with specific variables so it cannot improvise at scale. HDI and its partner built this on Camunda in five months, designed to be reused across claims, policy changes, and service requests.
Questions to pressure-test your own operations
HDI's answer was an operating model, not a shopping list of tools. If your CEO asked the same question tomorrow, start here.
- Count your real long tail. The processes in your repository are the easy part. The ones that matter hide in inboxes, handler workarounds, and undocumented exceptions. HDI's number was 13,598. Yours is probably higher than you think.
- Measure your communication lag. Track the time between a customer's first contact and your first substantive reply. If it runs in weeks, the constraint is your communication architecture, not your headcount.
- Ask the reusability question. When you automate something, does it serve one domain or many? A portfolio of point solutions becomes a web of dependencies that grows harder to maintain and exponentially more expensive to run.
Every process running in your operation today was designed for a world without AI. ProcessOS, the intelligence layer of Camunda's agentic orchestration platform, is built to do exactly what HDI set out to do: discover how a process actually runs, redesign it for an AI-native world, and build it in weeks instead of quarters.
Your CEO's version of the question is coming.
Join the ProcessOS waitlist and have an answer ready.
