Skip to content

Vet Vendor SOC 2 Scope, Model Training Policies, and Subprocessor Data Retention

Skip generic 300-row firewall questionnaires and ask the 15 sharp questions that expose whether an AI vendor trains on your data or logs prompts for 30 days.

Copy the prompt and paste it into the AI assistant you already use.

Pre-screening checklists help business buyers eliminate high-risk tools early; formal vendor security approval, Data Processing Agreements (DPAs), and compliance attestations must be signed off by InfoSec and Legal.

Five red flags business buyers can catch BEFORE bothering InfoSec

  1. 01

    Distinguish 'Hosted on SOC 2 AWS' from the vendor's OWN SOC 2 Type II

    Early startups often slap a SOC 2 badge on pricing pages because AWS has SOC 2—ask explicitly for the vendor's own audited SOC 2 Type II report and bridge letter.

  2. 02

    Verify 'Zero Data Retention (ZDR)' at the underlying LLM subprocessor layer

    A vendor promising 'we don't train our models' means little if they call standard non-enterprise LLM APIs that retain prompts 30 days for abuse review—check subprocessor terms.

  3. 03

    Audit OAuth Scope Blast Radius (`Read All Mail` vs `Scoped Alias/Label`)

    Reject tools demanding tenant-wide super-admin read/delete permissions across the CEO's mailbox when scoped user-level or forwarded-thread permissions suffice.

  4. 04

    Confirm hard deletion SLAs for vector embeddings, caches, and backups

    Ask what happens within 30 days of contract termination or GDPR erasure: do derived vector search indexes and cached summaries purge alongside primary rows?

Common questions

What is the difference between SOC 2 Type I and SOC 2 Type II?

Type I only verifies that security controls existed on paper on a single snapshot date, whereas Type II proves an independent auditor tested that controls actually operated over 6–12 months.

Why should we check a vendor's `/subprocessors` page before signing?

An AI app is only as private as the third-party transcription, analytics, session-replay, and LLM hosts it pipes your text into—verify their subprocessor notification window.

Can we turn off human support-engineer access to our prompts inside AI tools?

Good enterprise vendors enforce explicit 'Customer-Initiated Support Access Only' with audit logs so vendor engineers cannot browse tenant workspace prompts without your ticket permission.

Learn more

Related guides and background articles for this task