Agentic Workflow

Automating a single task is straightforward. Automating a sequence of tasks that requires judgement at each step is where most process automation projects stall. An agentic workflow uses AI to orchestrate multi-step processes, deciding what to do next based on what it has learned so far. For financial services, this means replacing manual hand-offs between teams with intelligent pipelines that route, escalate, and resolve.

What is an agentic workflow?

An agentic workflow is a structured sequence of tasks where an AI agent (or multiple agents) handles the orchestration: deciding which step to execute next, routing work to the right system or person, and adapting the process based on intermediate results. Unlike a fixed automation script that follows a predetermined path, an agentic workflow can branch, loop, and escalate based on what it encounters.

Consider a new customer onboarding process. A fixed workflow follows the same steps for every customer. An agentic workflow assesses the initial application, determines the risk tier, routes high-risk customers to enhanced due diligence, checks sanctions lists, verifies documents, and escalates exceptions to a human reviewer. The agent decides the path based on data, not a predetermined rule tree.

The distinction from traditional robotic process automation (RPA) is important. RPA automates repetitive, rule-based tasks. Agentic workflows handle tasks that require interpretation, judgement, and adaptation. RPA clicks buttons in a predictable sequence. An agentic workflow reads a document, interprets its content, decides what to do with it, and routes it accordingly.

The landscape

The major AI and automation platforms are converging on agentic workflow capabilities. Microsoft's Copilot Studio, Salesforce's Agentforce, and open-source frameworks like LangGraph provide the building blocks. The tooling is maturing, but production deployments in regulated financial services remain early-stage. Most organisations are running pilots, not enterprise-wide implementations. The EU AI Act's requirements for human oversight in high-risk applications apply to any agentic workflow that touches credit, insurance, or fraud decisions.

The PRA's operational resilience framework applies directly to agentic workflows that support important business services. If an agentic workflow handles claims processing or customer onboarding, it is part of the service delivery chain and must meet the same resilience standards as any other critical system. Fallback procedures for when the agent fails or produces uncertain results are a regulatory requirement, not a nice-to-have.

The challenge of multi-agent systems is emerging. Complex workflows may involve multiple agents collaborating: one that handles document extraction, one that performs risk assessment, one that drafts communications. Coordinating multiple agents introduces its own failure modes. An error in one agent's output propagates through the chain. Current tooling handles single-agent workflows well. Multi-agent coordination is less mature, with debugging and observability tooling still catching up.

How AI changes this

End-to-end claims handling is the most developed use case in insurance. An agentic workflow receives a claim notification, extracts information from the claim form and supporting documents, checks the policy terms, assesses whether the claim falls within coverage, estimates the reserve, and routes to the appropriate handler or, for simple claims, settles automatically. Insurers deploying this pattern report straight-through processing rates of 40 to 60 per cent for motor windscreen claims, up from near zero with purely manual processes.

Customer complaint resolution benefits from agentic workflows that combine investigation and response. The workflow retrieves the customer's history, identifies the root cause of the complaint, checks for similar past complaints and their resolutions, drafts a response, and routes it for approval. What previously took an analyst three to five days of elapsed time can complete in hours.

Regulatory change implementation is an emerging application. When new regulation is published, an agentic workflow can assess which internal policies are affected, identify gaps between current practices and new requirements, draft updated policy language, and create implementation tasks for relevant teams. The output is a structured change programme, not a single analysis. The value compounds across the 600-plus regulatory changes that major UK financial institutions track annually.

What to know before you start

Map the process before automating it. An agentic workflow that automates a broken process produces broken results faster. Document the current process, identify the decision points, understand where human judgement adds value, and define where the agent can act autonomously and where it must escalate. This process mapping exercise often reveals inefficiencies that can be fixed before AI enters the picture.

Design for graceful degradation. The agent will encounter situations it cannot handle: ambiguous documents, contradictory information, edge cases outside its training. Define clear escalation paths for each failure mode. A workflow that freezes when the agent is uncertain is worse than a manual process. A workflow that escalates to a human with full context is better than both.

Measure end-to-end cycle time, not just automation rate. An agentic workflow that processes 60 per cent of cases automatically but creates a bottleneck for the remaining 40 per cent may not improve overall throughput. The underwriting team that handles exceptions needs capacity planning, not just a dashboard showing how many cases the agent resolved.

Start with a single, well-bounded workflow where the steps are known, the data sources are accessible, and the success criteria are measurable. Customer onboarding for a specific product, first notification of loss for a specific insurance line, or vendor risk assessment for a specific category. Build, test with the operational team, measure against the manual baseline, and expand only after the first workflow is stable.