From Manual Reports to Agentic AI in Finance: Practical Lessons from Pavlé Sabic of Moody's
High-stakes decisions in finance outgrew traditional automation. Fragmented data, strict audit trails, and complex workflows make manual processes slow and risky. Agentic AI - systems that plan and execute multi-step tasks with oversight - is closing that gap by accelerating work while keeping experts in control.
Industry bodies such as the Financial Stability Board note a steady rise in AI adoption for internal operations and compliance across finance, driven by advances in data and compute. At the same time, enterprise leaders are prioritizing auditability and governance as adoption moves past surface-level tasks and into core workflows.
In a recent discussion, Pavlé Sabic, Senior Director of Generative AI Solutions and Strategy at Moody's, broke down how leading institutions are applying agentic AI today - and what it takes to deliver speed, compliance, and human accountability at once.
About Pavlé Sabic
Pavlé leads generative and agentic AI strategy at Moody's. A seasoned product and commercial leader, he previously drove double-digit growth at S&P Global across a $140M professional-services segment and led commercial strategy for its Insights division. He holds a Master's in Finance and Investment from the University of Edinburgh Business School.
Key Insight 1: Accelerate Decisions with Human-in-the-Loop Agents
Most agentic AI in finance still lives in low-risk workflows like customer queries and appointment routing. High-stakes processes - credit origination, risk reviews - remain expert-led, with agents handling the heavy lifting and analysts making the call.
According to Sabic, a human-in-the-loop setup has cut time to production by roughly 60% in some cases. The win isn't "autonomous risk assessment"; it's faster, cleaner information flow in a decision-ready format.
Consider the credit memo. It includes a sector overview, financial decomposition, strategy analysis, and implied probability of default - historically stitched together across tools and documents. Today, agents can draft each section to spec - format, charts, references - and route it to analysts for review and sign-off. As Sabic puts it: "Tell the agents the 10 sections you need, and they can go off and do that."
- Define "decision points" and approval gates. Agents prepare; humans approve.
- Break complex memos into modular tasks: data gathering, normalization, analysis, drafting, and evidence linking.
- Enforce sourcing rules: cite datasets, versions, and timestamps in every output.
- Track overrides and final decisions to improve prompts, tools, and controls.
Key Insight 2: Drive Compliance with Proprietary Data
Proprietary data is the backbone of reliable agents. Sabic emphasizes that the strongest results come from pairing advanced models with vetted in-house content: credit research, ratings analytics, sector taxonomies, and risk frameworks. That foundation enables agents to execute client-specific workflows and withstand audits.
Moody's launched a research assistant in December 2023 that blends proprietary research with generative capabilities for credit workflows. Sabic reports users process 60% more data and reduce tasks by about 30%, while handling larger content volumes without losing auditability.
The takeaway: public web data is noisy. Enterprise agents need controlled corpora and explicit rules on what they can use, how they cite it, and how outputs are logged. That's how you pass an audit - every time.
- Inventory internal data and map it to target workflows (e.g., origination, ongoing monitoring, model validation).
- Use retrieval over proprietary sources with strict citation and versioning.
- Gate external data behind trust policies; prefer licensed, traceable sources.
- Enable red-teaming and model cards for transparency on limits and risks.
Adoption Trends and Guardrails
Moody's 2025 report notes 53% of risk and compliance teams are using or trialing AI for fraud detection, KYC, and related functions. That traction comes with conditions: clear oversight, audit trails, and human governance at scale.
Research and commentary from organizations like the Financial Stability Board and publications such as Nature point to a wider shift: agents that plan and adapt multi-step processes are moving into data-heavy, regulated settings - provided supervision is built-in.
Implementation Reality: Orchestration and Workforce Readiness
Sabic draws a line between adoption and implementation. Natural language interfaces make advanced tools accessible, but that puts pressure on training and change management. Teams must learn how to prompt, review, and govern agentic workflows.
He also sees a split across industries. Non-regulated sectors can get by with off-the-shelf LLMs. Regulated institutions need centralized orchestration, model risk management, and a plan for model churn - every new model release can trigger re-validation and system updates.
"Enterprise-grade orchestration is key," says Sabic. "Scale requires platforms that coordinate agents across systems, centralized to avoid silos and ensure consistency."
- Centralize orchestration: one control plane for prompts, tools, connectors, policies, and logs.
- Isolate sensitive workloads; enforce data residency and access controls per region and entity.
- Instrument everything: prompt/response logs, tool calls, citations, approvals, and overrides.
- Plan for model switching costs: abstraction layers, evaluation pipelines, and rollback paths.
Metrics That Matter
- Cycle time: time-to-memo, time-to-approval, and time-to-funding.
- Work efficiency: tasks automated, analyst hours per case, cases per FTE.
- Quality and control: override rate, error rate, citation coverage, audit pass rate.
- Risk outcomes: PD/LDG drift, false positives/negatives in alerts, post-approval loss performance.
90-Day Starter Plan for Finance Leaders
- Weeks 1-2: Pick one high-friction workflow (e.g., credit memo drafting). Define scope, guardrails, and approval steps.
- Weeks 3-6: Stand up retrieval over proprietary content. Build an agentic flow with enforced citations and human sign-off.
- Weeks 7-10: Pilot with 10-20 analysts. Track cycle time, override rate, and user feedback. Iterate weekly.
- Weeks 11-12: Formalize policies, monitoring, and playbooks. Plan phased rollout and training across teams.
Practical Next Steps
- Choose "assist, not replace" use cases first. Win trust with measurable cycle-time cuts and clean audit trails.
- Invest in proprietary data readiness before model tuning. Data quality beats model novelty.
- Build a central orchestration layer to avoid agent silos and duplicated risk.
- Stand up continuous evaluation and change control for every model update.
Further Learning
- Explore curated AI tools relevant to finance professionals: AI Tools for Finance
- Browse role-focused AI training paths: Courses by Job
The bottom line: agentic AI won't replace expert judgment in regulated finance - it makes that judgment faster and more defensible. Start where oversight is clear, ground agents in your own data, and centralize orchestration. That's how you scale speed without losing control.
Your membership also unlocks:
