Skip to content

Checklist

Regulated AI procurement checklist

What to ask an AI vendor before you approve them: where the data goes, who operates it, what the certificate covers and how you would leave.

Updated 21 Aug 2026

Definition

Procurement should ask an AI vendor for the workflow and professional boundary, evidence that each part of the product is implemented, a data-flow map, identity and permission tests, named operating owners, current certification scope, and the commercial and exit terms. A useful checklist turns each unknown into a named evidence request or a blocking decision.

01

Workflow and professional boundary

  • Users, inputs and output
  • Decision and human owner
  • Prohibited uses
  • Failure behaviour and consequences
02

Product evidence

Ask for proof that a surface is implemented and for representative evaluation results covering retrieval, citations, uncertainty, permissions and your specific workflow. A feature list answers none of that.

  • Implemented-surface proof
  • Representative evaluation results
  • Retrieval, citations and uncertainty
  • Not a generic feature list
03

Architecture and data

Map every place data rests or passes through for the deployment you are proposing, and who holds it there.

  • Stores, processors and models
  • Logs, backups and support routes
  • Location and retention
  • Deletion and subprocessors
04

Identity and governance

  • Organisation separation and roles
  • Source permissions and revoked access
  • Administrator powers
  • Trace coverage and evidence export
05

Operations

Assign monitoring, incident, recovery, change, update, vulnerability, support and business-continuity responsibilities.

  • Monitoring and incident response
  • Recovery and change control
  • Updates and vulnerability response
  • Support and business continuity
06

Corporate and certification evidence

Review product controls separately.

  • Legal operator and certificate holder
  • Scope and issuer
  • Accreditation and dates
  • Product controls reviewed separately
07

Commercial and exit

  • Scope and dependencies
  • Service, support and price terms
  • Renewal and termination
  • Data export and portability
08

Decision record

Mark each requirement evidenced, open, accepted with a named risk owner, or blocking. Keep hard constraints outside any weighted preference total.

  • Evidenced or open
  • Accepted with a risk owner
  • Or blocking
  • Hard constraints kept separate

Procurement checklist

What to see before you approve

Mark each row evidenced, open, accepted with an owner, or blocking. A blank is not a low-risk answer.

Review areaEvidence needed
WorkflowNamed users, decision, documents, failure cost and human owner
Evidence qualityKnown, synthesis, conflict, stale, no-answer and permission tests
Data flowStores, processors, locations, retention, deletion, support and backups
IdentityOrganisation, role, source permission, revoked access and administrator tests
OperationsMonitoring, incident route, recovery, change, update and vulnerability ownership
CommercialScope, dependencies, service terms, exit, export and evidence commitments

What this page does not prove

  1. B1The checklist is not legal advice or certification.
  2. B2Completion does not prove a deployment safe or compliant.
  3. B3Evidence must match the exact product release and architecture.
  4. B4Material changes trigger re-review.

Ask for the current Marella procurement route

The public checklist remains available without a form. Use this only to request current Marella-specific evidence categories or an approved route for a confidential questionnaire; do not paste one here.

Contact the Marella AI team

Please keep it non-confidential.