Security and compliance that decides, records, and proves - on every action.

AI governance for finance is deciding, at runtime, whether an action - taken by a person or an AI agent - should be allowed to proceed against financial systems, then proving that decision afterward. In Fintra this is one engine, end to end: AegisGraph enriches every action against a reasoning graph at ingest and computes its blast radius; the action is mapped to specific controls across 76 frameworks; SentriAI scores it with a deterministic Action Trust Score and issues a verdict - allow, step-up, hold, or block; Trust CORTEX ranks the decision for a human; and the whole thing seals into a hash-chained evidence ledger. Consequential actions are governed, draft-first, and human-confirmed: an AI agent can propose, but a named person approves before anything irreversible posts.

That is the question SentriAI answers on every consequential action. Reading a report is not the same as posting a journal entry or moving money. AegisGraph reasons about the action in context - the graph, the toxic-combination paths, the blast radius - and AgentFence intercepts the agent’s tool call at the MCP boundary, checks it against allowlist and argument policies, and routes financially material actions through human-in-the-loop approval. Low-risk reads flow freely so agents stay useful; irreversible or money-moving actions wait for a named human, and both the proposal and the approval are recorded in the hash-chained ledger.

The Action Trust Score is a deterministic, explainable number SentriAI computes for every consequential action from weighted signals - identity confidence, account and vendor trust, compliance state, and Fintra’s financial context. “Deterministic” means the same inputs always produce the same score, and every score ships with a factor breakdown, so a reviewer sees why an action was trusted or held. It is the runtime companion to the composite Enterprise Trust Score: one is per-action, the other is the organization-wide roll-up.

Every governed action - the prompt, the tool call, the verdict, the approval - is written to an append-only ledger where each entry embeds a cryptographic hash of the previous entry. Altering or deleting any record breaks the chain, and verify_chain recomputes it on demand to prove integrity. Each entry is also mapped to the specific controls it satisfies across frameworks, so operations become audit evidence automatically rather than being reconstructed after the fact.

AegisGraph is the reasoning engine at the front of the flow. Its signature move is graph-at-ingest: before anything reasons about an event, it enriches it against the identity ↔ asset ↔ agent graph, runs deterministic detection rules, and computes a risk score and blast radius. It surfaces toxic-combination attack paths - dangerous chains of individually-benign conditions - that single-signal tools miss. Today it reasons over a graph seeded with design-partner data; it decides, records, and can gate, but it does not silently auto-execute anything.

A SIEM or dashboard tells you “that happened.” This engine decides whether it should happen - before it does - scores it, enforces or holds the verdict, and proves it. Because it sees actions through Fintra, which owns the money, it grounds identity, fraud, and compliance signals in real financial events (pending approvals, payment and payroll changes, new-vendor and bank-change events) that a standalone monitoring tool never sees. Trust CORTEX outputs a decision, never a passive chart.