The first AI agents your organization deployed probably delivered. A chatbot that summarized case files. An assistant that surfaced customer data. An agent that triaged inbound claims. Real value, genuine wins, and a growing confidence that scaling would be straightforward.
Scaling is a different problem. A 2026 survey of 1,150 senior IT and business decision-makers found that while 71% of organizations are deploying AI agents, 80% of those deployments remain assistants that summarize and answer questions rather than execute mission-critical work. Only 11% of agentic use cases have reached production. Gartner projects more than 40% of agentic AI projects will fail by 2027. The technology works. The operating models built to absorb it don't exist yet.
Banking Faces a Speed Gap. Insurance Faces a Distribution Earthquake.
Lily Wang, CIO for Wholesale Onboarding & Group Controls at Barclays, named the pressure at CamundaCon Amsterdam. "The key objectives has been similar in that for many years. But how we deliver that now increasingly has to be much more fast-paced because the environment is changing."
"The key objectives has been similar in that for many years. But how we deliver that now increasingly has to be much more fast-paced because the environment is changing."
- Lily Wang, CIO for Wholesale Onboarding & Group Controls, Barclays
Regulatory velocity is outpacing the speed at which most banks can re-engineer their response. T+1 settlement, evolving AML/KYC requirements, increasingly sophisticated financial crime: each shift demands process adaptation, and the current cycle for adapting a core process runs months to years. McKinsey estimates 50–60% of banking operations work is tied to activities that agentic systems can handle. The banks that redesign those processes first frees capacity that compounds exponentially. For example, analysts redirected from manual evidence gathering to complex risk decisions, operations teams focused on exceptions rather than routine handoffs. Every quarter that re-engineering is delayed, that compounding advantage accrues to a competitor.
Insurance faces a different kind of disruption. AI agents are becoming the buyers. Within 18–24 months, consumer-facing agents will conduct automated policy reviews across dozens of carriers, and distribution advantage will shift to whoever is most legible to AI decisioning upstream of purchase. Carriers whose products, pricing, and underwriting logic can't be expressed in machine-reasonable terms will be invisible to this new channel. 86% of insurance organizations plan to increase AI spending in 2026, with agentic AI topping the list. The investment intent and technology is present. The operating model to absorb it isn't.
The Agentic Value Trap
Early agent deployments deliver quick wins. Confidence grows. More agents get deployed without coordination infrastructure. They begin conflicting with each other and fragmenting across the landscape. Teams are building the same AI agents on repeat. The efficiency gains that motivated the original deployments get consumed by the breakdowns caused by agent sprawl. In their O'Reilly book on agentic orchestration, co-authors Bernd Ruecker and Amy Johnston call this the "Agentic Value Trap": agent deployment without process governance generates coordination debt that compounds with every new agent.
Half of organizations say their AI agents operate in silos, disconnected from end-to-end processes. Researchers at the University of Vienna confirmed the structural gap in March 2026: today's agent frameworks are insufficient alone.
The market is converging on what fills that gap. Forrester describes a shift from task-level automation to process orchestration at enterprise scale. Gartner predicts 60% of organizations will adopt independent AI agent orchestration platforms by 2030. The common requirement: governed processes where agents operate with bounded autonomy, full audit trails, and human checkpoints where stakes demand them.
In regulated industries, this is table stakes. An agent that can't be audited can't be used.
What Re-Engineering Looks Like at Barclays and HDI
At CamundaCon Amsterdam 2026, Barclays presented how they are re-engineering client due diligence end-to-end: the financial crime prevention workflow spanning prospecting through ongoing monitoring and offboarding. Gautam Verma, head of financial crime core platforms and client due diligence technology, told Computer Weekly the process "can get very involved and can take months."
Barclays brought AI agents for data collection, intelligence, and policy guidance directly into their core processes with an open agentic orchestration layer. The targets: faster time to revenue, reduced client friction, and full regulatory compliance. Verma's insight at CamundaCon speaks to why this re-engineering is end-to-end: optimizing a single stage of the process in isolation doesn't improve time to revenue. The gains come from redesigning the full lifecycle.
At HDI, part of the Talanx group (one of the largest insurance groups worldwide), the mandate came from the top. Eric Hellmich, Head of Process, described it plainly: "Six months ago, our CEO just asked me, 'If you start rethinking operations, how would you do it? What will change?' And that's what we did."
"Six months ago, our CEO just asked me, 'If you start rethinking operations, how would you do it? What will change?' And that's what we did."
- Eric Hellmich, Head of Process at HDI
HDI runs more than 40,000 process variants across its German operations alone. Classical automation takes six months per process and stops being cost-effective at that scale. HDI's approach: use AI agents to parse unstructured inbound communication, generate process models from documented operational knowledge, and compress the data-gathering cycle. Their expected results include customer response times dropping from two to three weeks to near real-time.
Closing the Gap Between Technology and Organizational Velocity
Re-engineering for AI is the right approach. The math just doesn't work at human speed. A typical enterprise runs hundreds of business critical processes. Re-engineering one takes roughly 12 months. That gap between business need and execution will only continue to grow. The winners will be the ones who can move fastest – while keeping the rhythm of business.
ProcessOS, announced at CamundaCon 2026, is Camunda's answer to that gap that applies AI to the process lifecycle itself. It’s a multi-agent swarm that is split into four distinct phases of the lifecycle from Discovery to Deployment.
The discovery phase ingests existing documentation, operating procedures, business rules, and operational logs to map how a process actually runs in production. You control how to feed the data; whether automatically through Github or MCP connected to your data, or using a secure filesystem.
Design combines that context with defined business outcomes you set and rethinks the shape of the entire workflow from the outcome backward. You control the levers based on your goals whether it’s efficiency and reducing cycle time, improving quality, lowering costs. You control what matters most, The output is a visual model in BPMN that functions as both the stakeholder review artifact and the actual execution contract. What's designed is what runs, and what runs is what an auditor can trace at every step.
Build and Deploy generates all production artifacts (executable models, decision tables, integration code) and deploys it aligned to your existing CI/CD pipelines and policies. Optimize monitors performance data from running instances, identifies bottlenecks, and proposes changes with a value hypothesis for a process owner to approve. The re-engineering stays current after deployment.
Camunda ran ProcessOS on its own Quote-to-Cash process. Four weeks from discovery to deployed reimagined process, freeing an estimated 6,000 person-hours annually. For a bank or carrier running hundreds of core processes, this changes the math from years of sequential re-engineering to a continuous improvement cycle.
Lily Wang put it directly at CamundaCon: "The next thing we are needing to embrace is actually now how do we re-engineer."
See how ProcessOS works
ProcessOS is currently in closed beta with a planned Early Access release in October with Camunda 8.10.
