Direct Pojišťovna automates claims, cutting processing time by 87% with Azure AI

Insurance runs on documents, and rarely tidy ones. For Direct Pojišťovna’s (Direct) claims team, a single windshield claim could contain multiple forms, photos, handwritten reports, invoices, PDFs, emails, or legal documents. Each in a different format and quality. Sometimes incomplete. Sometimes ambiguous. Every case had to be read, interpreted, and reconciled by hand.

"The people in our claims-handling department were genuinely overwhelmed," recalls Jakub Lada, AI Product Owner. "The whole process was very manual, so our people spent a lot of time doing repetitive work.”

Direct also had a second, quieter constraint: the clock. Claims could only move forward when the office was open. "Our opening hours directly limited how many claims could actually be processed," says Lada. Direct had a backlog that paused every evening and weekend, and even relatively simple claims could take far too long. For more complex cases, the amount of documentation and the need for human review meant that processing would sometime take days to complete. This increasingly became out of step with what customers expected.

Overcoming the manual strain of unstructured data

But making a change wasn’t going to be easy. On a technical level, that variability of input posed a significant challenge. “Claims data does not arrive in one clean, predictable format,” notes Pavel Lukeš, Head of the Information Technology Department. “The hard part was not only extracting data, but also validating it, connecting it to the right claim context, and making the result reliable enough for operational use.”

The solution? A modular, agent-based system with people in the loop. “Human-in-the-loop was an essential component,” says Lukeš. “It would allow us to introduce AI into the process safely, giving us better visibility into where the model was reliable and where additional rules or process changes might be needed.” Working in partnership with BigHub, Direct turned to Microsoft Azure AI Foundry and Azure Document Intelligence, with an initial focus on windshield insurance. These high-volume claims were going to be an ideal proving ground to test the capabilities of their solution. 

“Azure Document Intelligence was a good fit because the core problem was document understanding and structured data extraction from semi-structured and unstructured inputs," says Lukeš. "Azure AI Foundry helped us approach the pilot in a more systematic way. We could experiment with models, evaluate outputs, and build the solution in an enterprise-ready environment. The Azure ecosystem also made sense from the perspective of security, scalability, and integration with the rest of our architecture.”

Piloting a new way of working

Two early results showed promise. The first came from the document-reading layer.

“When we tested text analysis through Optical Character Recognition (OCR) with Azure Document Intelligence, we saw 100% precision in understanding the text," says Lada. "That was the very first thing that really resonated." The second was just as important: the system's automated assessment of a claim matched the judgment a human handler would have reached. Both were strong signals that the pilot could move forward with confidence.

And once the system started processing live claims, the difference was immediate. Claims processing reduced from 15 minutes of manual work to 2 minutes (an 87% cut), while 60% of processing was able to be fully automated by AI. Routine cases were easily flagged and prepared for quick resolution, while clear summaries were provided to claims handlers, reducing dependence on expert knowledge. Customers with uncomplicated claims received payouts within minutes, and communication became faster and simpler, both of which improved customer satisfaction scores.

“This was innovative technology,” notes Lada. “It was transformative for the company and freed our staff away from simpler cases to working on more interesting, complex ones.” 

Faster, around-the-clock, and trusted

What’s more, processing was no longer limited by working hours. The system could prepare or process claims continuously, so work could move forward during evenings, nights, or weekends. Direct’s operational rhythm slowly shifted. Instead of waiting for the next working day, the processes would be partially completed or ready for human review in the background. And that automation slotted into existing operations without disruption.

“After implementing automation across those claims, there was no negative impact and no difficulty adapting it into the processes the company already had,” says Lada. “That was the real success.”

For Lukeš, the proudest milestone was the moment the system crossed from assisting to delivering. “We saw the system was working so reliably that we decided to include the actual claims payout directly in the AI-supported process,” he says. "That was an important shift. The solution was no longer just helping with one isolated step but became part of an end-to-end process. From the customer’s perspective, that’s where the change becomes very tangible: the claim is handled in a much faster, smoother way, with a noticeably different level of service."

The pilot also changed how people inside Direct viewed AI. Once staff saw it working in a real, customer-facing process, hesitation gave way to demand. “It became clear that people wanted to be involved in these AI and automation initiatives,” says Lada. "They could see the innovation wouldn't wait for them, so they felt a genuine need to be part of the projects.”

Expanding beyond windshield claims

Buoyed by the pilot’s success, Direct is now extending beyond windshield claims to other claim types. Bringing into play a crucial aspect of the initial implementation. Rather than build a one-off automation bolted to windshield claims, Direct designed for reuse from the outset. The architecture was assembled from interchangeable components that integrated back into the core claims process.

"We tried to avoid designing the solution as a one-purpose automation," says Lukeš. "The idea was to create a pattern that can be adapted to other insurance lines by changing the data model, rules, and document types, rather than rebuilding the whole solution from scratch."

In time, Direct plans to expand the solution across car insurance, property, liability, pet insurance, and further lines. The target is ambitious and specific. “In two to three years, we want 70% of all claims received to be fully automated,” says Lada. “The remaining 30% would be the most complex cases, and that’s our people thrive. We want to move them toward more direct, customer-facing roles, solving the problems that genuinely need human judgment.”

“In an environment as un-dynamic as insurance,” he continues, “our AI-readiness means we can prototype fast and test new ideas quickly. We can dramatically cut the time it takes to prove an idea is good. And with this pilot, we’ve done just that.”

Discover more about Direct Pojišťovna.