Banking, insurance, and capital markets run on actions that are regulated, irreversible, and time-critical. The Akka Agentic AI Platform lets AI agents execute those actions — payments, credit decisions, claims, regulatory filings — with audit trails, authorization controls, and human oversight embedded within the platform.
Financial institutions are deploying agents that initiate payments, submit trading orders, approve credit, file regulatory reports, and settle claims — at machine speed, without a human reviewing every action. That shift breaks control frameworks designed for human decision-making.
Every step survives failure, restart, and redeploy. A payment or decision runs exactly as intended, once — even when infrastructure fails mid-flight.
State changes are recorded as an append-only, replayable log — the immutable, time-stamped record auditors ask for, embedded within the platform.
Deterministic gates and human-in-the-loop escalation evaluate each proposed action against policy before it is applied — not after.
Data residency and active-active operation across regions for workloads carrying regulatory and availability obligations.
SAFR (Safeguards for Agentic Finance at Runtime) is the emerging industry reference model for governing agents at the moment they act — developed with the Monetary Authority of Singapore's Project MindForge and BuildFin.ai, and aligned to NIST's AI Risk Management Framework. It defines a governance checkpoint between every agent decision and its execution.
The Akka Agentic AI Platform satisfies SAFR's four runtime components with capabilities embedded within the platform. Agent identity, authorization controls, the pre-execution decision gate, and the audit trail are all provided by the platform itself.
| SAFR component | What it requires | Embedded in the Akka Agentic AI Platform |
|---|---|---|
| Governance Envelope | A faithful record of what the agent intends and how it got there — the action trace, authenticated to origin. | A sealed, signed evidence bundle — the action and its decision trace transparently captured (including for third-party agents), bound by a reproducibility hash and authenticated to origin. |
| Agent Identity | Every action bound to a registered, verified agent before evaluation proceeds. | SPIFFE identity + ACLs + JWT — a stable, verifiable identity for every agent. |
| Controls Repository | A configurable rulebook of policy, regulation, and capability-based mandates — bounded, machine-readable authority. | The regulatory corpus and its controls — policy matrices automatically enforced against each proposed action. |
| Disposition Engine | A deterministic, pre-execution gate resolving each action to Deny / Escalate / Auto-Execute / Observe — re-run every step. | Enforcement of evaluations, sanitizers, guardrails, HITL escalations, HOTL escalations (halt switches), and testing gates. |
| Audit Log | An immutable, tamper-evident record, reconstructable independent of the agent. | Petabyte-scale logging — append-only and tamper-evident, with 10-year retention and legal hold. |
FINOS AIGF is the financial-services industry's open standard for governing agentic AI, developed under the Fintech Open Source Foundation. It defines 23 agentic-AI risks and 23 mitigations — preventive and detective — and is the basis for the FINOS Common Cloud Controls adopted across banking, insurance, and capital markets.
Akka implements the FINOS mitigations as platform capabilities. The same runtime that enforces SAFR carries the preventive and detective controls FINOS specifies.
| FINOS AIGF mitigation | What it requires | Embedded in the Akka Agentic AI Platform |
|---|---|---|
| Agent authority, least privilegeAIR-PREV | Each agent may act only within explicitly granted authority. | SPIFFE identity + capability-scoped mandates — an agent cannot widen its own authority. |
| Tool-chain validation & sanitizationAIR-PREV | Tools and their inputs and outputs validated and sanitized before use. | Sanitizers + tool-call guardrails run in the decision path before any tool executes. |
| Multi-agent isolation & MCP securityAIR-PREV | Agents isolated from one another; the MCP tool supply chain governed. | Each agent an isolated component; MCP endpoints governed and access-controlled. |
| Model & data firewallingAIR-PREV | Inbound and outbound filtering of model traffic and knowledge-base data. | Runtime guardrails filter prohibited content on the way in and out. |
| Agent decision audit & explainabilityAIR-DET | Every agent decision auditable and explainable after the fact. | Tamper-evident, hash-chained log + sealed, signed evidence bundles. |
| Automated evaluation & human feedbackAIR-DET | Automated (LLM-as-a-judge) evaluation with a human feedback loop. | Evaluations (LLMs-as-judges) + HITL / HOTL escalation. |
SAFR and the regulatory corpus are enforced by a two-layer control model embedded in the platform — a real-time check on every agent action, and periodic verification over time.
Runtime filters catch prohibited or unsafe content before an agent can act on it.
Personal and confidential data is automatically removed from what agents read and write.
Every agent action is checked inline as it happens — pass, block, or flag — so a non-compliant action never executes.
High-stakes actions escalate to a person for approval, and halt switches let a human stop an agent instantly.
Scenario, replay, stress, and adversarial red-team tests gate every change before production.
Every decision is written to a tamper-evident, hash-chained log, retained up to 10 years with legal hold.
Akka maintains a live corpus of the financial-services regulations that govern agentic AI — mapped to enforceable controls and monitored as they change. Governance runs against this corpus at the agent boundary, so obligations are enforced, not just reported.
A selection of the 49 financial-services regulations in the corpus, mapped across the EU, UK, US (federal and state), Singapore, Australia, Canada, Hong Kong, and Brazil. EU AI Act Art. 72 alone mandates 10-year retention with tamper-evident records.
Initiate, sequence, and settle payments with per-action authority limits and full reconstructable traces.
Surface, prioritize, and act on suspicious activity — with escalation gates before any customer-affecting action.
Real-time decisioning under model-risk governance. See Capital One below — two-day batch to 200 ms.
Assemble, validate, and file against the mapped control set, with the audit trail retained by construction.
Straight-through claims with human-in-the-loop escalation on high-materiality or anomalous cases.
Bounded agent authority over quotes, endorsements, and renewals — every action policy-checked pre-execution.
Detect and contain in real time; reversible actions auto-execute, irreversible ones escalate.
Personalized, compliant interactions with conduct and data-protection controls enforced inline.
Agentic workflows across the trade lifecycle with deterministic containment and tamper-evident records.
Auto Navigator pre-qualifies buyers across ~4 million cars from 12,000+ dealers with no impact to their credit score — 486 applications per minute at 180–200 ms under load.
Read the story →Manulife selected Akka to build a high volume of trusted AI-powered applications on a secure and scalable foundation — the insurance proof point that anchors this vertical.
Read the story →
