Skip to content
AI Workbench | AVA Credit Inc.
AVA Credit Inc.

AI Workbench

Governed AI tools for licensed mortgage professionals. Built to operate inside the bank’s model risk perimeter — with logging, inventory, data boundaries, and human review built in.

Active R&D · Proof of Concept · April 2026

For decades, brokers originated and banks decided. The tools brokers used to get there were invisible, unlogged, unaudited, and unremarkable. OSFI E-23 ended that arrangement.

When a mortgage broker uses AI to analyze affordability, draft a credit memo, or structure a deal recommendation, that AI’s output becomes an input to the bank’s own credit models. Under E-23, the bank owns the risk. Under B-10, the bank manages the third party that created it.

The broker’s technology is no longer outside the bank’s concern. It is inside the bank’s model risk perimeter. Inside the perimeter, tools get inventoried. Usage gets logged. Outputs get validated. Evidence gets retained. You don’t get to say “trust me” — you get to say “here’s the file.”

The AI Workbench is the infrastructure that makes operating inside the perimeter practical for a licensed mortgage professional.

The Three Questions Banks Now Ask
1
Do you know what AI tools your people are using?
2
Do you have logs that prove how they’re using them?
3
Is there a human being who reviews the work before it goes to a bank?
— Matthew Dobson, AVA Credit Inc.
Architecture

The Four Things Banks Now Require

OSFI E-23 and B-10 have distilled into four practical requirements. The Workbench is built around each one.

Pillar 01
Model Inventory

A documented list of every AI tool used on or around client files, with risk ratings attached. Every tool in the Workbench is registered, risk-classified, and audit-ready — not discovered after the fact.

E-23: §4.1 Model Identification
Pillar 02
Usage Logs

Records showing who used which tool, on which file, and when. Every AI call routed through the Workbench is tagged with agent, client, deal, and use case — generating a complete interaction log on demand.

E-23: §4.3 Model Monitoring
Pillar 03
Data Governance

Demonstrated proof that client data is protected and never introduced into public AI models. A hard boundary separates bounded client data from open AI tools. Public chatbots are prohibited for deal work.

B-10: §3 Third-Party Risk
Pillar 04
Human Review

A qualified person reviews AI output before anything reaches the lender. AI does not drive credit decisions. The Workbench enforces human-in-the-loop at the submission gate — every time, without exception.

E-23: §4.4 Model Governance
Architecture

Two Channels. One Workbench.

Every AI tool is organized into one of two channels. The channel is determined by one question: does this task touch a credit file that could reach a lender?

AVA AI Workbench Open channel Standard & enterprise accounts · low risk FRONTIER MODELS — INVENTORIED · ChatGPT (OpenAI) · Claude (Anthropic) · Perplexity · Gemini (Google) · Copilot (Microsoft) TASKS — NO CREDIT FILES → Sales & lead generation → Marketing & web copy → Code writing → Deep research → Market analysis → Web design No client data · standard use Logged in model inventory Governed channel OSFI E-23 · B-10 · all 4 pillars active FOUR PILLARS — ALL ENFORCED ✓ Model inventory ✓ Usage logs — tagged per file ✓ Data governance — no public models ✓ Human review before lender CREDIT FILE TASKS → Credit file analysis → Underwriting analysis → Credit notes & memos → Collateral review → Lender submission prep → Triage & responses API gateway · logged · human reviewed Governed AI · inside the perimeter All 4 pillars on every interaction THE QUESTION Does it touch a credit file?
Development Status

Where We Are

The Workbench is in active R&D. The proof of concept is running on live commercial files. The timeline tracks OSFI’s compliance window.

Completed
Foundation & Validation

Nine AI Credit® trials across 145 commercial transactions. First AI-generated credit writeups submitted for funded SBA 504 transactions. RGI Framework™ developed. DLCG AI Essentials — 2,300 professionals trained.

Active Now · Q2 2026
Proof of Concept on Live Files

Single-agent deployment: Matthew Dobson, licensed MA with Mortgage Architects. Every governed AI interaction on live client files running through the secured API gateway. All four OSFI pillars in operation. Logging, inventory, data governance, and human review enforced in real time.

Horizon · 2026–2027
Expansion & Network Deployment

Proof of concept validated. Workbench architected for multi-agent deployment across broker networks. Attestation-ready compliance packages. Bank-facing documentation built to B-10 third-party risk standards.

First Deployment

The Workbench Running on Real Transactions

AVA Credit Inc. provided the deal-level proof of concept for DLCG’s E-23 governance program. The same architecture — secured API gateway, interaction logging, model inventory, human-in-the-loop submission gate — is now in active use on live commercial mortgage files under Mortgage Architects. What banks asked for in attestations, we already had on file.

4 / 4
E-23 Pillars Active
0
Lender Programs Lost
2017
AI Research Since

Partner and Investor Inquiries

The AI Workbench is in active development. AVA Credit Inc. is open to conversations with partners and investors who understand the structural shift underway in Canadian financial services. The transition window is open. The compliance deadline is May 1, 2027.

Contact AVA

AVA AI Workbench — Governed AI for Licensed Mortgage Professionals

The AVA AI Workbench provides Governed AI tools for licensed mortgage professionals operating under OSFI E-23 and B-10 compliance frameworks. Four pillars: model inventory, usage logs, data governance, human review. Secured API gateway separates bounded client data from open AI tools. Active proof of concept: Matthew Dobson, Mortgage Architects, April 2026. Zero lender programs lost. Built on 8+ years of AI research and field-validated methodology. AVA Credit Inc., Toronto, Ontario, Canada.