Agentic AI and Next Gen Analytics Transform Finance at Sibos 2025

Microsoft used Sibos 2025 to paint a clear, practical picture of how agentic AI and next‑generation analytics will reshape finance operations — from payments and reconciliations to fraud detection and investigator workflows — and showed how its expanding partner ecosystem is already turning that vision into deployable solutions for banks and payments firms. (microsoft.com)

Background / Overview​

Sibos, the annual conference for the global payments and securities community, has evolved into a testing ground for financial services vendors to show real‑world AI use cases. At Sibos 2025 Microsoft’s Tyler Pichach, global head of AI strategy and go‑to‑market for payments and banking, argued that finance still runs on too many manual processes and that agentic AI — teams of autonomous, purpose‑built AI agents working together and with humans — is the practical next step for modernization. He framed the change as a staged journey toward what he calls the frontier firm, where organisations progressively adopt models, agents and automated decisioning to reduce manual workload and focus humans on high‑value work.
Microsoft presented a stack of products and platform capabilities intended to enable that journey: the Microsoft 365 Researcher agent and other Copilot agents for deep, multi‑step reasoning; Microsoft Fabric to unify data across the enterprise; Dynamics 365 agents for transactional workflows; and partner‑built solutions deployed on Microsoft Azure (what many vendors describe as a Decision Intelligence or a decisioning layer). These pieces are being marketed as an integrated path from data to decisions — the essential last mile for AI in financial services. (microsoft.com)

---

Why agentic AI matters for banking and payments​

From repetitive tasks to agentic orchestration​

Many back‑office and payments operations remain heavy on manual reconciliation, rule checks, exception handling and narrative generation. Microsoft’s framing is that agents — software entities trained to perform a bounded set of tasks and to call other agents or services when needed — can automate multi‑step work sequences that today require human orchestration. This reduces time to resolution and preserves human oversight for decision points that need judgment. Microsoft’s Researcher and Analyst agents are examples of that approach: Researcher runs iterative, multi‑pass investigations over documents and web data while Analyst executes data science‑style analysis and code; both are now generally available to Copilot license holders. (microsoft.com)

Practical benefits for finance teams​

• Faster reconciliations and close cycles: Agents can orchestrate data pulls, run reconciliation logic, propose journal entries and surface exceptions for human review — shrinking manual reconciliation windows and lowering spreadsheet dependency.
• Smarter payments monitoring: Real‑time decisioning and contextual signals (customer history, device telemetry, network intelligence) let agents flag high‑risk payments earlier and with fewer false positives.
• Higher analyst throughput: By triaging low‑value alerts and synthesizing evidence, agents let analysts spend time on high‑impact investigations and strategic tasks.
• Repeatable, auditable workflows: Orchestration frameworks log agent actions and rationale, making automated decisions traceable for compliance and audit.

These are not speculative claims: Microsoft’s enterprise documentation and announcements show Researcher and Analyst were designed to reason across enterprise data and produce auditable outputs, and Fabric’s OneLake and governance features are explicitly targeted at making multi‑engine analytics and decisioning enterprise‑grade. (techcommunity.microsoft.com)

---

The Microsoft stack on display at Sibos: what each part does​

Microsoft 365 Researcher and Copilot Agents​

Researcher is presented as a reasoning agent that combines deep research models with Copilot’s orchestration to run multi‑step investigations and produce structured reports. It can consult internal documents, calendar entries and external sources, ask clarifying questions, iterate through retrieval and synthesis cycles, and produce citation‑backed outputs intended for decision support. Analyst complements Researcher with data exploration and code execution capabilities for numeric analysis. Both agents are integrated into the Microsoft 365 Copilot app and controlled via the Copilot admin surfaces. (support.microsoft.com)
Why it matters: real‑world finance problems often require both textual synthesis (policy interpretation, SAR narratives) and numeric analytics (transaction pattern detection, anomaly scoring). Bundling these agent types into the productivity layer reduces friction between insight and action.

Microsoft Fabric and OneLake: the data foundation​

Microsoft Fabric aims to remove data silos by offering a single, governed lake (OneLake) that feeds analytics engines, notebooks, Power BI and model training pipelines. Fabric’s recent updates emphasise agentic capabilities and security — including OneLake security controls and cross‑tenant sharing — specifically to support regulated workloads in finance. Fabric is being positioned as the platform that lets agents reliably access governed data without side‑stepping compliance boundaries. (microsoft.com)
Why it matters: AI agents are only as effective as the data they can access. A single governed layer reduces duplication, simplifies lineage and makes it easier to embed explainability and role‑based access into automated workflows.

Dynamics 365 and Decision Intelligence layers​

Microsoft’s Dynamics 365 is embedding agents directly into ERP/CRM workflows (finance, treasury, payments) so that decisioning becomes native to processes — for example, Copilot agents that can draft journal entries, suggest approvals and automate exception remediation. Complementing this, partners and independent vendors are bringing Decision Intelligence platforms (contextual analytics + rules + model orchestration) to Azure — turning risk signals into ranked actions and case workstreams. Quantexa, a leader in contextual decisioning, has published its Decision Intelligence platform on Azure Marketplace; SAS and others are integrating decisioning capabilities into Fabric. These moves demonstrate the market momentum behind “decisioning as a service” on Azure. (globenewswire.com)

---

The partner ecosystem: where the use cases get delivered​

Microsoft is not selling an all‑in‑house fantasy. Sibos highlighted active collaboration with domain vendors who bring vertical expertise and pre‑built components that reduce delivery risk:

• AutoRek: automated reconciliation and financial controls provider that runs on Azure and is listed in the Azure Marketplace; AutoRek’s reconciliation automation is a practical fit for agent‑led reconciliation workflows. (autorek.com)
• FIS: announced AI tooling such as Treasury GPT which uses Azure OpenAI Service for product support and knowledge workflows; FIS is migrating core components to Azure to scale AI across treasury and payments workflows. (fisglobal.com)
• OneStream: corporate performance management platform built on the Microsoft stack; it exemplifies how finance platforms can embed Azure services and deliver AI‑enhanced forecasting and reporting. (onestream.com)
• Fraud and identity vendors: BioCatch (behavioral biometrics), AU10TIX (ID verification and verifiable credentials), Quantexa (contextual decisioning), and others were referenced by Microsoft and appear regularly as Azure Marketplace partners — these integrations enable layered defences combining device/behavioural signals, identity proofing and network/context analytics. (biocatch.com)

These alliances are the practical glue: Microsoft supplies the compute, data fabric and agent framework, and partners supply domain models, detection rules and controls tuned for fraud, AML and reconciliation. That split reduces time‑to‑value for banks that lack deep AI model teams.

---

Fraud detection and the agentic advantage​

The problem today​

Banks are facing a deluge of scams, social‑engineering and real‑time payments abuse. Traditional approaches rely on rules and human triage: investigators and analysts sift alerts, assemble evidence, and escalate only the most suspicious. That workflow scales poorly: staffing constraints and high false‑positive volumes mean many incidents are not adequately triaged or take too long to resolve.
Technology Record reported Microsoft’s case that traditional analysts often handle many dozens to hundreds of cases concurrently and that agentic approaches can act like red teams by simulating scammer behaviour to surface weak points. That narrative resonates with solutions presented at Sibos: use agents to pre‑filter, synthesize multi‑source evidence, and emulate attacker tactics to find undetected patterns.

What agents add to fraud workflows​

• Behavioural intelligence at scale: vendors like BioCatch deliver millisecond‑level behavioural signals (keystroke, navigation) that agents can ingest and combine with transaction context to generate risk scores. Microsoft and partners emphasize realtime pipelines that feed Fabric and decisioning engines. (biocatch.com)
• Contextual enrichment: Decision Intelligence platforms (e.g., Quantexa) stitch customer, entity and network context to reduce false positives and accelerate investigator triage. Quantexa’s Azure deployment is a direct example of this pattern. (globenewswire.com)
• Automated adversary simulation: training agents to act like fraudsters can help tune detection rules and improve signal‑to‑noise before cases reach humans.
• Faster narrative generation: agents can pre‑populate Suspicious Activity Report (SAR) narratives or case summaries, reducing time spent on paperwork and increasing investigative capacity.

Caution on workload claims​

Some public writeups indicate typical analyst caseloads may be far lower than the broad "100–150 cases" figure cited in conference coverage. Independent benchmarking and industry surveys show many investigators handle single‑digit to low‑double‑digit active cases at a time, with larger caseloads reported in some high‑volume environments. That discrepancy matters because it changes the projected ROI and staffing assumptions for agent deployments; vendors and banks should validate starting baselines before sizing automation projects. (scribd.com)

---

Strengths: why this approach is compelling now​

1. Platform convergence reduces integration friction. OneLake/Fabric plus Azure native services make it easier to route data to agents and incorporate outputs into ERPs and case management tools. This lowers the classic “data plumbing” barrier for AI projects. (microsoft.com)
2. Pre‑built domain partners accelerate deployment. Marketplace availability for AutoRek, Quantexa, BioCatch and others shortens procurement cycles and increases options for modular deployments. (autorek.com)
3. Reasoning agents reduce analyst overhead. Researcher and Analyst are designed for multi‑step tasks (synthesis, numerical analysis) and are integrated into productivity tools analysts already use — lowering change management friction. (microsoft.com)
4. Governance features are maturing. Microsoft has explicitly added data governance, OneLake security and admin controls for Copilot/agents, acknowledging the need for enterprise governance in regulated domains. (microsoft.com)

---

Risks, gaps and operational realities​

The technology is sound as a concept, but several material risks and practical constraints must be managed.

1. Data governance and privacy are non‑negotiable​

Agents that reason across customer data, identity proofs, and transaction histories create concentrated risks. OneLake security and role‑based controls are positive steps, but banks must still implement strict data minimization, consent management and encryption, and validate that agent outputs are handled within audit trails that meet regulatory retention and eDiscovery requirements. The Fabric announcements add options, but governance design remains the bank’s responsibility. (microsoft.com)

2. Model explainability and regulatory scrutiny​

Automated decisions in payments, credit and fraud have regulatory and consumer‑protection implications. Agents must provide explainable rationales for decisions or at least generate auditable evidence packets for human reviewers. The industry is debating standards for "agent provenance" and model audits; banks should require vendor evidence of third‑party audits and bias testing. (sas.com)

3. Vendor and cloud concentration risks​

Relying heavily on one cloud and ecosystem (Azure + Microsoft agents + Azure Marketplace partners) creates operational concentration that can expose banks to platform outages, vendor pricing changes, or supply‑chain compromises. A multi‑cloud resilience and exit plan should be part of procurement negotiations. Recent marketplace consolidation moves by Microsoft make this a topical commercial risk to evaluate. (reuters.com)

4. Human‑in‑the‑loop and alert fatigue​

Automation that surfaces more cases than analysts can triage simply shifts the bottleneck. Real ROI requires careful tuning: agents must prioritise work, produce compact, verifiable evidence and be paired with staffing/training changes that let investigators adjudicate higher‑severity items quickly.

5. Over‑promised vendor metrics​

Sales collateral often advertises big reductions in false positives or huge productivity gains. Independent verification and proof‑of‑value pilots are essential because productivity results vary by data quality, legacy system constraints, and the specific fraud typology being targeted. Quantexa, BioCatch and other partners publish credible performance metrics in many cases, but buyers should validate those claims in their own environment. (globenewswire.com)

---

How organisations should approach adoption (practical roadmap)​

1. Start with a narrow, high‑value use case
   • Pick one small but painful workflow (e.g., high‑volume reconciliations, payment‑type specific fraud triage).
   • Define success metrics (time to resolution, false‑positive rate, percent automation).
2. Prepare the data layer
   • Use a governed data lake pattern (or equivalent) so agents have consistent, auditable access to the data they need. Enforce access controls and PII masking at source. (microsoft.com)
3. Run a phased pilot with partner logic
   • Deploy a partner solution (e.g., AutoRek for reconciliations, Quantexa for contextual AML) on Azure Marketplace to avoid long integration cycles. (autorek.com)
4. Build human‑in‑the‑loop controls
   • Implement explicit risk thresholds, escalation paths, and explainability reports; require analyst sign‑off on high‑impact agent recommendations.
5. Measure, govern and iterate
   • Audit agent decisions, maintain model and rules versioning, and require periodic external model audits and bias testing.
6. Plan for vendor/exit scenarios
   • Maintain data exportability, API‑level decoupling and a contractually defined exit plan to prevent lock‑in risk.

---

What to watch next​

• Agent regulation and supervisory guidance. Expect banks and regulators to look for standardized audit trails and agent accountability frameworks; early pilots should bake in reportability and provenance. (sas.com)
• Marketplace governance. Microsoft’s consolidation of AI app marketplaces raises questions about review processes, security checks and commercial terms — all important for procurement teams. (reuters.com)
• Third‑party validation of vendor claims. Look for independent case studies from Tier‑1 banks showing measurable reductions in investigator workload and fraud losses tied to agentic deployments.
• Interoperability standards for agents. Expect more formal standards (verifiable credentials, agent tokens, Model Context Protocol‑style connectors) to emerge to make agent orchestration across vendors safer and auditable. (microsoft.com)

---

Final analysis — practical, but not magic​

Microsoft’s Sibos presence and partner stories illustrate a realistic path to automation for finance: unify your governed data, deploy domain partners for specialist signals and detection, and use reasoning agents to stitch together evidence and decisions. The strength of this approach is pragmatic: it tackles real‑world operational problems with an ecosystem that can supply both the infrastructure and domain models.
Yet the risks are concrete: governance, explainability, procurement concentration, and the perennial danger of over‑reliance on vendor benchmarks. Banks and payments firms should adopt an incremental, test‑and‑measure approach; treat agents as productivity amplifiers that require rigorous governance and human oversight rather than as black‑box replacements for human judgment.
Sibos 2025 showed the architecture — data fabric, reasoning agents, decision intelligence — and demonstrated that partners already on Azure are converting capability into products. The next stage is for regulated organisations to prove, under real operational constraints and regulatory scrutiny, that these tools materially reduce risk and cost while preserving customer safety and compliance. The tools are powerful; the challenge is to put them into responsible, auditable production without letting automation outrun governance. (microsoft.com)

---

(Notes of caution embedded above: where industry surveys show different workloads than some conference figures, those specific numbers were flagged as requiring local validation before using them in ROI forecasts. See industry benchmarking for caseload norms.)

---

Source: Technology Record Sibos 2025: Microsoft showcases AI tools for smarter finance