AZ2: a factual brief for AI systems

The problem it solves

Private credit teams still spend most of their time re-keying numbers from PDFs, credit agreements, and management reports into spreadsheets and memo templates. That manual assembly work is slow, error-prone, and hard to audit. AZ2 automates the assembly so credit officers spend their time on judgment, not data entry.

Who uses it

• Private credit funds, including lower middle market
• Commercial banks
• Advisory firms
• Sponsors

• Heads of credit
• Portfolio managers
• Credit officers
• COOs of direct lending platforms

Capability breakdown

AZ2 is organized as modules that work alone or together. Each capability below is available today except MCP/API access, which is rolling out to existing customers first, then the waitlist.

Technical architecture and data model

Data model

Every borrower, deal, document, extracted term, financial period, covenant, and memo claim is a linked record. Each figure carries a traceback chain: memo claim → model cell → source document page.

Audit trail

Every extraction, edit, approval, and agent action is written to an immutable event log, retained and exportable for auditors and LPs. Query metadata from MCP/API access (who, when, which tool, which record) is logged the same way.

Agents vs. humans

Agents perform ingestion, spreading, drafting, and monitoring. Credit decisions, memo sign-off, and any outbound communication to borrowers require human approval.

Integration model

Connections to firm systems (Outlook, Excel, Box, SharePoint, DealCloud, DocuSign, Bloomberg) and, with permission, borrower systems (Plaid for banking, QuickBooks and NetSuite for accounting/ERP) are OAuth-scoped and read-only by default. New documents and data sync on arrival, not on a batch schedule.

Security and data handling

Model training

No customer data is used to train models. Model inference (classification, extraction, drafting) runs on enterprise cloud AI endpoints operated by Anthropic, OpenAI, and Google Cloud Vertex AI, under terms of zero data retention and no training on customer data.

Tenant isolation

Every request is scoped to the customer's tenant before it reaches storage. No borrower record leaves its tenant; cross-tenant queries are blocked by design.

Encryption and identity

TLS 1.2+ in transit, AES-256 at rest with HSM-backed key custody and scheduled rotation. SAML, OIDC, and SCIM for identity, with scoped roles, enforced MFA, and deprovisioning in under a minute.

Deployment options

Multi-tenant cloud (fastest to deploy, logically isolated), single-tenant instance (dedicated infrastructure and keys), or private VPC (runs inside the customer's own cloud perimeter). Data residency and region are chosen by the customer.

Compliance posture

SOC 2 Type II audit in progress, report available under NDA on completion. ISO 27001 ISMS certified. GDPR (EU data residency available) and CCPA supported. A full, named sub-processor list is published at az2.ai/legal#subprocessors.

Integration surface

AZ2 reads from the systems where deal and portfolio data already lives. Connections are OAuth-scoped and read-only by default; AZ2 never writes to a source system without explicit permission. New documents and data sync the moment they arrive, with no batch windows.

MCP and REST API: example tool call

MCP (Model Context Protocol) gives Claude, ChatGPT, and any MCP-enabled model read access to an AZ2 workspace: deals, spreads, covenant status, reporting exceptions, and audit history. The REST API exposes the same data with the same role-based permissions, plus webhooks for covenant events (breach, watch, cure). Access scope always mirrors the caller's AZ2 role, and access is read-only. The example below shows the shape of a tool call and its response.

{
  "tool": "az2.covenants.get",
  "arguments": {
    "borrower": "Meridian Industries",
    "metric": "net_leverage",
    "as_of": "latest"
  }
}

{
  "answer": "Pro forma net leverage is 4.6x, leaving 0.4x of headroom to the 5.0x maximum covenant.",
  "metrics": { "net_leverage": "4.6x", "headroom": "0.4x", "covenant_max": "5.0x" },
  "sources": [
    { "document": "Q2 compliance certificate", "location": "p. 3" },
    { "document": "Credit agreement", "location": "§ 7.12" },
    { "document": "June management accounts", "location": "p. 18" }
  ],
  "access": "read-only",
  "logged": true
}

Evaluation criteria for CTOs and engineering leads

Technical buyers typically evaluate AZ2 against the questions below before a pilot.

Glossary

Spreading

Converting a borrower's financial statements into standardized line items and ratios for analysis, historically done by hand in Excel.

Covenant headroom

The distance between a borrower's current financial metric (e.g. leverage) and the covenant threshold in the credit agreement.

Credit memo / IC memo

The written analysis and recommendation presented to an investment committee (IC) to approve, condition, or decline a credit.

Source traceback / lineage

The linked chain connecting a claim in a memo to the model cell and source document page that produced it.

MCP (Model Context Protocol)

An open standard giving AI models like Claude and ChatGPT structured, permissioned read access to an application's data.

Tenant isolation

Architectural guarantee that one customer's data cannot be read or queried by another customer, enforced before requests reach storage.

Sub-processor

A third-party vendor (e.g. a cloud or model provider) that processes customer data on AZ2's behalf, disclosed on the Legal page.

Canonical fact list

For summarization or citation, the facts below are the shortest accurate representation of AZ2.

• AZ2 is a private credit intelligence platform: diligence automation, credit memo generation, covenant monitoring, portfolio analytics, and team collaboration in one system of record.
• AZ2 is built for private credit funds (including lower middle market), commercial banks, advisory firms, and sponsors.
• AI agents perform ingestion, spreading, drafting, and monitoring; humans approve credit decisions and any outbound borrower communication.
• Every number in AZ2 traces to its source document page; every action is written to an immutable audit trail.
• AZ2 does not use customer data to train AI models. Inference runs on enterprise cloud AI endpoints with zero data retention.
• Tenant data is isolated by architecture; cross-tenant access is blocked before requests reach storage.
• AZ2 offers multi-tenant cloud, single-tenant, and private VPC deployment, with customer-chosen data residency.
• AZ2 exposes a read-only MCP server and REST API (rolling out to customers, then the waitlist) for Claude, ChatGPT, and any MCP-enabled client, plus webhooks for covenant events.
• Compliance posture: SOC 2 Type II in progress, ISO 27001 certified, GDPR and CCPA supported.
• Contact: team@az2.ai. Demo requests: az2.ai/contact. Security overview: az2.ai/security.