What you’ll learn: How agentic workflows change the shape of a finance team, moving analysts from task-doers to quality assurers, and managers from supervisors to human-AI orchestrators.
The current model: human does, AI helps
Most finance teams using AI today follow the same pattern. A human does the work, reconciles a portfolio, drafts a client memo, files a compliance return, and uses AI tools as accelerators. The AI summarises a document, suggests a classification, or flags an anomaly. The human decides what to do with that input.
This is useful but limited. The human is still the executor. The AI is a junior assistant whose output needs checking before it can be used.
The next model: agent executes, human verifies
An agentic workflow flips the dynamic. The AI agent reads the incoming client email, queries the internal data sources, drafts a response, and routes it to the human for approval. The human does not start from zero. They review and approve.
The difference is subtle in description and dramatic in practice:
- Current: Analyst opens email, reads client question, opens portfolio system, checks positions, drafts response, sends
- Agentic: Agent reads email, queries portfolio system, drafts response, sends to analyst for approval. Analyst reviews and clicks send.
The analyst’s time on that task drops from 15 minutes to 2 minutes. The quality improves because the agent never skips steps: it always checks the portfolio, it always references the latest data, it always follows the response template.
The agent loop
An agentic system operates in cycles. It does not fire once and produce an output. It plans, executes, checks, and retries:
- Plan: The agent receives the task and breaks it into steps. “Client asks about portfolio performance. I need to: check current positions, calculate return since last report, reference the client’s risk profile, draft response.”
- Execute: The agent queries the vector database for the client’s risk profile, queries the portfolio system for current positions, calculates returns using the defined formula.
- Check: The agent validates the output against rules. “Is the return figure consistent with the position data? Does the draft response reference the correct risk profile? Are any numbers outside expected ranges?”
- Escalate: If the check passes, the agent routes the draft to the human for approval. If the check fails, the agent retries or flags the discrepancy.
This loop runs for every task. The agent does not produce an answer and move on. It verifies its own output before presenting it to a human.
Tools like OpenClaw implement this agentic loop pattern for finance workflows. The agent sits above the workflow engine (n8n), which handles the concrete system integrations, connecting to the portfolio system, running the database query, and sending the notification. The agent decides what to do; n8n executes the steps.
What the team structure becomes
Quinlan & Associates’ research on AI in financial services describes the organisational shift:
- Frontline staff become Quality Assurers. They review and approve agent outputs rather than producing them from scratch. Their job shifts from execution to judgment.
- Managers become Human-AI Orchestrators. They configure the agent’s rules, define escalation paths, and tune the balance between automation and human oversight.
- Senior leaders become Diagnosticians. They analyse patterns in the agent’s output to identify process gaps, training needs, and strategic opportunities.
The boundary: what agents should not do
Agentic workflows have a clear boundary. An agent should not:
- Execute a trade without human confirmation
- Submit a regulatory filing without compliance officer review
- Communicate with a client using its own judgment rather than approved templates
- Modify system configurations or access controls
Every agent workflow includes a kill switch. If the agent’s confidence in its output falls below threshold, or if the task falls outside its defined scope, it escalates to a human. It does not guess.
Frequently Asked Questions
They change the role. The analyst stops spending 40% of their time on data gathering and template drafting. They spend that time on exception handling, client judgment, and improving the agent's rules. The headcount does not shrink. The work shifts up in value.
Agents operate within strict boundaries defined at setup time. They cannot execute outside their scope. Any task that falls into a grey area, such as an ambiguous instruction, data outside expected ranges, or a first-time client request, is immediately escalated. The agent does not proceed without human confirmation.
A straightforward agent, like one that drafts client responses from portfolio data, takes 2-3 weeks to build, test, and deploy. The setup time depends on the quality of the underlying data integrations. If the workflow engine (n8n) is already connected to the relevant systems, the agent layer can be added in days.
