Financial institutions continue to spend heavily on financial crime compliance, yet fragmented data and manual work still absorb substantial capacity. In a benchmark of leading banks, McKinsey found that KYC and anti-money laundering activities can account for 10% to 15% of full-time employees.
Client due diligence exposes the operating-model problem. Teams collect company records, identify beneficial owners, verify information across sources, screen for sanctions and adverse media, and assess risk. For complex corporate clients, the process can take weeks or months.
The decision itself is rarely the main source of delay. Manual evidence collection, repeated data entry, team handoffs, and fragmented systems create rework and hide the status of each case. In Fenergo’s survey, 70% said their firm had lost at least one client in the previous year because of slow or inefficient onboarding.
At CamundaCon, Gautam Verma, Head of Financial Crime Core Platforms and Client Due Diligence Technology at Barclays, explained how the bank is redesigning this work from end to end.
Barclays blended agentic work with control
Barclays began with an operating-model question: which parts of CDD benefit from AI agents, and which require predictable execution?
Predefined processes enforce required approvals, risk rules, controls, deadlines, and evidence. Agents handle variable, context-dependent work such as collecting information, comparing sources, identifying gaps, and verification while adhering to specific procedures and policies. People retain accountability for exceptions and high-consequence judgments.
This division gives the bank a complete record of exactly what happened down to the reason why. That traceability matters when a decision must withstand internal review or regulatory scrutiny.
The design also addresses a common weakness in enterprise AI programs. Automating one task may reduce local effort without improving the customer’s end-to-end experience. Faster onboarding depends on redesigning and measuring the full process.
Three specialist agents free analyst capacity
Barclays described three agent capabilities, each assigned a defined outcome.
- A data collection agent gathers information from third-party sources and recommends authoritative sources for the client and jurisdiction.
- A data intelligence agent checks the collected information against bank policy, flags gaps, and recommends which values analysts should review.
- A guidance agent converts policy into contextual prompts, helping analysts spend less time searching procedures and more time assessing risk.
This model keeps people focused on work that requires judgment while agents support evidence-intensive tasks that take up the majority of their time. The boundaries are explicit: agents contribute analysis and recommendations inside a governed process rather than operating separately
Agentic orchestration provides the control plane
Barclays is using Camunda to coordinate predefined processes, specialist agents, people, rules, and existing systems across the client due diligence process.
For IT leaders, this orchestration layer creates a control plane for the end-to-end process. It can make the sequence of work explicit, invoke agents within defined boundaries, route exceptions to people, record evidence, and provide visibility across systems. The bank can change the process without making the program dependent on replacing every underlying technology. Risk of disruption is greatly minimized while rate of change speeds up.
For business leaders, the same foundation supports measurable outcomes. Leaders can see where cases wait, which steps create rework, how often agents escalate, whether evidence is complete, and where policy requirements affect cycle time. And ultimately, how this all impacts churn and revenue recognition.
That visibility matters because isolated automation can shift a single bottleneck without improving the client outcome. End-to-end orchestration gives the bank a way to redesign and measure the full operating model.
Measuring the whole onboarding outcome
Agentic CDD programs need measures that connect technology performance to operating and business outcomes. A useful scorecard should include:
- End-to-end onboarding cycle time
- Analyst touch time per case
- Percentage of cases completed without avoidable rework
- Exception and human-escalation rates
- Evidence completeness and policy adherence
- Number of follow-up requests sent to the client
- Client abandonment during onboarding
- Time from application to revenue-generating activity
These measures help leaders distinguish faster task execution from genuine process improvement. They also create a basis for deciding where agents add value, where deterministic control remains essential, and where people should retain authority.
The operating model determines the value of technology
Barclays’ approach offers three lessons for financial institutions:
- Understand the end-to-end process and where agentic AI works best and what should stay predefined
- Assign agents clear roles to alleviate time-consuming activities and preserve human accountability for consequential decisions.
- Use an open agentic orchestration layer to preserve the technology that works and speed up your rate of change
The central decision is how work should move, where control should reside, and how the institution will prove what happened. AI models contribute capability. The operating model and orchestration layer determine whether that capability improves CDD and customer experiences at scale.
Watch Gautam Verma’s session to learn more.
