INTAKE
Normalize the buyer's questions before drafting.
Keep the original wording and buyer identifier, but classify each question by system, topic, required evidence, risk, and owner. Mark duplicates and dependencies. This preserves the buyer's intent while exposing which answers can be reused.
Confirm the deadline, submission format, confidentiality terms, and whether the buyer wants current facts, roadmap commitments, or contract language. Those are different approval paths.
- Original buyer question and due date
- Named system and deployment boundary
- Topic, risk, owner, and required reviewer
- Requested attachment, link, or attestation type
SYSTEM / MODEL / DATA
Build a buyer-readable AI system record.
Describe the service, its intended use, customer-facing AI functions, models and providers, data sources, outputs, storage, regions, retention, subprocessors, and human checkpoints. Distinguish your controls from controls operated by upstream providers.
Capture versions and dates. A model or provider inventory without change history becomes stale the moment engineering changes the stack.
- Architecture and data-flow record
- Model, API, infrastructure, and subprocessor inventory
- Customer data-use and retention statement
- Intended use, prohibited use, and limitation record
TEST / PROTECT / MONITOR
Link security answers to operational records.
Buyer questions increasingly ask for testable AI controls. OWASP's AISVS 1.0 describes an open catalogue of testable requirements spanning the AI lifecycle and explicitly identifies procurement as a use case. Use a recognized source to challenge your coverage, but answer only with controls actually implemented in the named system.
Evidence may include access reviews, threat models, adversarial test plans, evaluation results, release thresholds, logging configurations, incident exercises, recovery tests, and change approvals.
- Identity, privilege, and production access review
- Model and prompt attack testing with dated results
- Evaluation thresholds, failed cases, and accepted exceptions
- Monitoring, escalation, incident response, and recovery records
OWNER / APPROVAL / LIMIT
Make governance claims traceable to decisions.
A committee name is not the same as evidence of governance. Link material claims to a charter, decision record, risk acceptance, review cadence, or approval history. Name the role that can approve the representation and the role that can fix a gap.
Record limitations beside the claim. If an evaluation covers only English prompts or one model version, the answer should not imply broader coverage.
- Policy or standard with effective date and owner
- Risk register entry and acceptance authority
- Release or model-change approval record
- Known limitation, exception, and target remediation date
FINAL REVIEW
Ship a frozen, reviewable response set.
Before submission, confirm that every link resolves for the intended reviewer, confidential attachments have the right access, answer language matches the evidence, and open gaps are not presented as completed controls.
Freeze the submitted version and keep the buyer's follow-up questions. They are inputs to the reusable answer library for the next deal.
- Security, product, privacy, and legal review where applicable
- Approved response version and evidence index
- Visible partial and missing states
- Buyer follow-up log and library update owner
FREE / NO EMAIL GATE
AI Procurement Evidence Inventory
A no-gate CSV with fields for the buyer question, framework reference, system scope, owner, approved answer, evidence URI, review date, gap, and target date.
PRIMARY SOURCES