Demo – Ava

Ava Workbench — Intake & Decomposition

This looks like an ordinary AI chat, and that is deliberate. The governance is underneath. Every element you just used maps to one of the four requirements regulators place on AI in banking. In Canada these are OSFI Guideline E-23 (govern every model touching a credit file) and B-10 (govern the third parties banks rely on). Every banking jurisdiction is converging on the same four, in one form or another.

Model inventory

One known, registered model. The badge reads Ava · Claude on Bedrock, ca-central-1 — not whatever public tool a person happens to open. The model surface is fixed and documented by construction.

→ the model badge

Usage logs

Every turn carries a log strip: deal ID, session, model, timestamp, request ID, and provenance on agent notes. Who used which model, on which file, when — recorded automatically.

→ the log strip under each turn

Data governance

The boundary made into a control. Names and UIDs live together inside Ava; the de-sensitize toggle governs what leaves. With it on, only UIDs cross to the model. Client identity never enters a public model.

→ the de-sensitize toggle

Human review

It is a chat. Nothing moves on its own. The human reads every output, points out errors, and accepts the record before it goes anywhere. Review is the act of using it — structural, not a checkbox.

→ the whole interface

What this UI does

Ava leans into one strength: organizing a pile of intake documents into a faithful, structured record, efficiently, inside governance. Its capabilities are modest next to a frontier model, and they are meant to be. What it does do is extract and organize the source documents into a governed, de-sensitized format, ready for further analysis.

Why it matters

Because the output carries UIDs and no client identity, it can be taken straight into an enterprise-grade frontier model for the heavy analysis. Ava organizes intake inside governance; the frontier model does the underwriting, on a record that is safe to send.