Vendor and third-party risk
Every commercial LLM API Bizzi uses ships under an enterprise agreement guaranteeing the vendor does not retain your data and does not train their models on it. This is one of the ten public customer commitments published in the Executive Summary. It is enforced by signed contract, not a click-through checkbox.
Context
Section titled “Context”A B2B finance AI platform letting vendor LLMs retain or train on customer data has effectively re-exported the customer’s most sensitive records to an offshore third party. The single most consequential vendor lever you have is the Zero Data Retention clause. You refuse to ship a feature in production behind a vendor unable to sign it.
Vendor selection criteria
Section titled “Vendor selection criteria”Six criteria gate every commercial LLM vendor before production.
- Zero Data Retention in a signed enterprise agreement, not a terms checkbox.
- Region pinning capability. The vendor pins inference to a region you name.
- Standard DPA. Data Processing Addendum with a complete sub-processor list.
- SLA on latency and uptime. Must clear the B2B operating thresholds.
- Reasonable liability terms. No exclusions for vendor-caused security breaches.
- Audit rights. You request SOC 2 or ISO 27001 attestations on a defined cadence.
A vendor failing any of the six is not approved. Exceptions are rare, require CPTO and Legal sign-off, and live in the vendor risk register.
Open-source models. The license matrix
Section titled “Open-source models. The license matrix”When Bizzi deploys an open-source model, Legal reviews the license terms to ensure commercial use is valid. Common licenses and the Bizzi position on each.
| License | Commercial use | Must open derivative code? | Bizzi position |
|---|---|---|---|
| Apache 2.0 | Yes | No | Approved |
| MIT | Yes | No | Approved |
| BSD-3-Clause | Yes | No | Approved |
| Llama Community License | Conditional | No | Legal reviews per version |
| CC BY-NC | No (non-commercial) | No | Not approved for production |
| AGPL-3.0 | Yes | Must release modifications | Approved only without modification, or with explicit decision to open-source |
| SSPL | Restricted (server-side use triggers) | Network use triggers | Avoided in production |
| Custom research-only | No | Not applicable | Not approved |
Every fine-tuned and deployed OSS model has a license attestation document stored alongside the model registry entry.
The sub-processor list
Section titled “The sub-processor list”You maintain a sub-processor list, available to contracted customers, covering.
- Current LLM vendors and the regions they operate in.
- Cloud infrastructure providers.
- Observability vendors (the layer hosting your logs).
- Specialist vendors (cloud OCR, managed vector DB).
When you add a sub-processor, you notify customers at least 30 days before production cut-over and provide an objection window.
Quarterly review
Section titled “Quarterly review”Legal and the CPTO review every vendor and sub-processor each quarter.
- Have any terms changed?
- Was there a security incident in the quarter?
- Performance against SLA?
- Vendor financial stability?
Vendors failing review are escalated and replaced if needed. The AI Gateway with fallback routing is the technical control making a vendor swap operationally feasible.