The framework
Apply this to every candidate vendor (or every analytical question of this type). Score each observation. Compare across instances. Document the result.
Download as PDFAI VENDOR DILIGENCE FRAMEWORK
The Six-Observation Method
─────────────────────────────────────────────────
Vendor: ___________________________________________
Tool / product: ___________________________________
Tier evaluated: ___________________________________
DPA executed?: [ ] Yes [ ] No [ ] Pending
Date of evaluation: _______________________________
Evaluator: ________________________________________
Each of the six observations is scored:
A = ACCEPTABLE (terms support Rule 1.6 / firm policy)
R = REQUIRES REDLINE (negotiable; track which clauses)
W = WALK-AWAY (do not contract on these terms)
═════════════════════════════════════════════════════
OBSERVATION 1 — TRAINING DATA RIGHTS
═════════════════════════════════════════════════════
The question: Can the vendor train its models on customer prompts and
outputs, or use them for "product improvement"?
What to look for:
• Default training-on-prompts language
• Opt-out vs opt-in
• Whether the enterprise tier has different terms
• Whether the DPA explicitly carves out customer data
Acceptable:
• Explicit, affirmative no-training-on-customer-data clause
• DPA confirms the clause applies to all customer data
• Customer can audit / verify
Walk-away:
• Default training-on-prompts with opt-out only
• Vendor reserves "improvement" rights without definition
• No DPA available
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
OBSERVATION 2 — RETENTION WINDOWS
═════════════════════════════════════════════════════
The question: How long does the vendor retain prompts, outputs,
conversations, metadata, and any derived data?
What to look for:
• Retention period (60 days? 1 year? indefinite?)
• Distinction between prompt history and metadata
• Soft-delete vs hard-delete
• Customer right to demand deletion
Acceptable:
• 30-90 day retention with customer-initiated deletion
• Hard delete confirmation
• Documented deletion mechanism
Walk-away:
• Indefinite retention
• "Soft delete only" / data persists in backups indefinitely
• Customer cannot trigger deletion
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
OBSERVATION 3 — SUB-PROCESSOR CHAIN
═════════════════════════════════════════════════════
The question: Who else, beyond the vendor, has access to customer data
in the course of providing the service?
What to look for:
• Cloud infrastructure (AWS, GCP, Azure)
• Underlying model providers (OpenAI, Anthropic, Google)
• Support and customer-success tooling
• Analytics providers
• Where each is located (jurisdiction matters for GDPR / CCPA)
Acceptable:
• Stable URL listing all sub-processors
• Customer notice obligation before sub-processor changes
• Customer right to object to a new sub-processor
Walk-away:
• Sub-processor list not disclosed
• Vendor reserves right to add sub-processors without notice
• Sub-processors include parties whose terms are weaker than the vendor's
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
OBSERVATION 4 — GOVERNMENTAL DISCLOSURE
═════════════════════════════════════════════════════
The question: Under what circumstances will the vendor disclose
customer data to law enforcement, regulators, or governments?
What to look for:
• Standard "we comply with legal process" language
• Customer-notice obligation before disclosure (where legal)
• Whether the vendor commits to challenge overly broad process
• Any specific carve-outs
Acceptable:
• Customer-notice obligation where legally permissible
• Commitment to narrow / challenge overly broad process
• Annual transparency report
Walk-away:
• Discretionary disclosure under broadly drawn clauses
• No customer notice
• No commitment to challenge
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
OBSERVATION 5 — ANONYMISATION CLAIMS
═════════════════════════════════════════════════════
The question: Does the vendor claim to anonymise customer data, and
does the anonymisation method actually work?
What to look for:
• Defined anonymisation method (k-anonymity, differential privacy)
• Whether anonymised data is treated as outside customer data
• Re-identification protections
Acceptable:
• Specific, documented anonymisation method
• Contractual commitment that anonymised data not be re-identified
• Anonymised data covered by the same protections as raw data
Walk-away:
• "Anonymised" means just stripping the customer name
• Anonymised data carved out of customer protections entirely
• Aggregation that reveals customer identity
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
OBSERVATION 6 — TIER DIFFERENTIATION
═════════════════════════════════════════════════════
The question: Are protections at the consumer / free tier the same as
at the enterprise tier? What activates the better protections?
What to look for:
• Whether the customer is on the right tier
• DPA execution as the activator
• Marketing claims vs operative contract terms
• What happens if a user accidentally drops to a lower tier
Acceptable:
• Enterprise tier is clearly demarcated
• DPA execution is documented
• All users in the customer's organisation are at the protected tier
• No silent tier-flipping
Walk-away:
• Marketing claims about enterprise terms not in the contract
• Free / consumer tier comingled with enterprise tier
• DPA terms apply only to "designated" users (rather than all)
Score: [ A / R / W ] Notes: ____________________
═════════════════════════════════════════════════════
SUMMARY
═════════════════════════════════════════════════════
Total observations: 6
Acceptable (A): ___
Requires redline (R): ___
Walk-away (W): ___
Overall recommendation:
[ ] PROCEED — all observations Acceptable
[ ] PROCEED WITH REDLINES — Requires Redline issues addressed
[ ] DO NOT PROCEED — one or more Walk-Away triggers
[ ] DEFER — additional information required
Specific redline asks (if any):
1. _______________________________________________
2. _______________________________________________
3. _______________________________________________
Walk-away triggers (if any):
1. _______________________________________________
Decision date: ____________________________________
Decision-maker: ___________________________________
This completed framework is filed with the firm's vendor-diligence
records as part of the procurement audit trail.How to use it
Inputs / fill-ins
The vendor’s privacy policy AND data-protection addendum (DPA) for the tier you’re evaluating. Marketing materials are not sufficient; the operative contract terms are.
For incumbent vendors, also gather: any prior DPAs, communications about sub-processor changes, and the firm’s prior procurement notes on this vendor.
What you get
Output
A completed 6-observation evaluation with explicit scores and notes. Becomes a firm-internal procurement record retained per the firm’s vendor-diligence policy.
Verification — what the lawyer must do
- Get the operative contract terms in writing. Marketing claims do not count. The DPA, the privacy policy, and the master service agreement are the operative documents.
- Compare across vendors. The framework is most useful when applied to 2-3 candidates and the scores compared.
- Document, file, and re-evaluate. Re-evaluate at contract renewal or when the vendor announces material changes.
⚠ Risks and failure modes
- Marketing-page risk: Vendors’ marketing pages overstate what their actual contracts deliver. Always evaluate against operative documents.
- Tier-flipping risk: A vendor may quietly migrate users between tiers. Confirm in writing that all firm users are on the protected tier.
- Drift risk: Vendor terms change. A vendor that scored Acceptable today may not score Acceptable next year. Re-evaluate at renewal.
Citations and further reading
- IXSOR: AI vendor diligence catalogue — the underlying analysis this framework operationalises.
- IXSOR: Legal Practice Management Software 2026 — worked examples applied to four PMS vendors.
- ABA Model Rule 1.6.
- ABA Formal Opinion 512.
- IXSOR Resources: AI Vendor Privacy Policy Analyzer — the prompt that runs Phase 1 of the framework against any vendor’s privacy policy.