Secure AI assistant claims: what should buyers ask?

Last reviewed May 24, 2026

Secure AI assistant claims usually combine product security, data handling, model use, identity controls, and AI-specific risk in one sentence. This guide turns those claims into evidence needed and buyer questions for vendor evaluation.

Evidence buyers verify

  • The exact secure AI assistant claim and the product plan or feature it applies to.
  • A data-flow map for prompts, files, outputs, logs, connectors, support access, and admin exports.
  • Identity, access, and role-control evidence for users, admins, apps, and integrations.

Opens the checker for this claim type. Paste your vendor's exact wording there. Evidence questions only — not a blacklist or fraud detector. Not sure what a result looks like? See a sample receipt.

Sources this guide draws from

  1. · January 26, 2023

    Source for AI risk-management context and trustworthy AI characteristics.

  2. · Excerpt from NIST AI RMF 1.0 (2023)

    Source for secure and resilient, privacy-enhanced, accountable, transparent, and human-oversight evidence questions.

  3. · Accessed May 24, 2026

    Public company source for business-data, encryption, access-management, retention, and training-boundary security wording; used as claim wording evidence, not independent validation.

  4. · Updated May 21, 2026

    Public company source for enterprise-grade privacy and security controls claim wording; not a third-party security audit.

Public claims with documented evidence gaps

"Enterprise-grade privacy and security controls"

Compliance / Safety
Source and date
OpenAI Help Center: What is ChatGPT Enterprise? · Updated May 21, 2026
Evidence signal
Security-grade wording without the specific controls, product scope, or admin boundary in the claim itself.
Evidence gap
A buyer needs the exact controls, supported plan, admin scope, identity integration, logging, data retention, and excluded surfaces.
Buyer question
For the enterprise-grade privacy and security controls claim, which controls apply to our plan, data types, and assistant features?

"data always remains confidential, secure, and entirely owned by you"

Compliance / Safety
Source and date
OpenAI business data page · Accessed May 24, 2026
Evidence signal
Broad data-control wording that needs product, retention, training, and support-access boundaries.
Evidence gap
A buyer needs the data categories covered, training-use default, retention settings, support access, subprocessors, and customer responsibility limits.
Buyer question
For the data remains confidential and owned by you claim, what data categories and product surfaces are covered or excluded?

"encrypted at rest and in transit"

Compliance / Safety
Source and date
OpenAI business data page · Accessed May 24, 2026
Evidence signal
Encryption wording that may not explain key control, logs, attachments, integrations, or admin-export paths.
Evidence gap
A buyer needs encryption method, key management option, covered storage locations, integration path, audit logging, and exception handling.
Buyer question
For the encrypted at rest and in transit claim, which data stores, logs, files, and integrations are included?

"We don't train our models on your organization's data by default"

Compliance / Safety
Source and date
OpenAI business data page · Accessed May 24, 2026
Evidence signal
Training-boundary claim qualified by 'by default', which means opt-in and configuration choices can change it. The claim also depends on product plan, API usage mode, fine-tuning, and subprocessor practices.
Evidence gap
A buyer needs the exact product surfaces and plans covered by the default no-training commitment, what opt-in or configuration options change that boundary, which subprocessors have access to inputs and outputs, and what the DPA says about training scope for the specific plan.
Buyer question
For the no-training-by-default claim, which product surfaces, plans, API configurations, fine-tuning workflows, and subprocessors are covered or excluded from the training boundary?

Match each claim pattern to the evidence buyers need

Claim pattern Evidence needed Buyer question
Secure AI assistant or enterprise-grade AI security Security architecture, identity controls, role model, admin settings, audit logs, incident process, and product-plan scope. Which assistant features can access sensitive data, and which controls restrict that access?
Private AI, confidential data, or customer-owned data Training-use boundary, zero-retention option, deletion controls, support-access process, DPA scope, subprocessors, and export paths. Is our prompt, file, output, log, or feedback data used for model training by default, and which retention or DPA terms apply?
Encrypted, protected, or secure by design Encryption at rest and in transit, key management, tenant isolation, integration security, and exception documentation. Where does encryption stop: browser, API, storage, logs, connectors, analytics, or support tooling?
AI-specific secure or resilient system claim Threat model for prompt injection, data leakage, tool access, model behavior, abuse handling, monitoring, and degradation path. What AI-specific risks were tested, and what happens when the assistant receives untrusted instructions or sensitive data?
AI data residency, audit log, or subprocessor disclosure claim Hosting region, data movement path, log coverage, subprocessor list, cross-border transfer basis, and customer configuration boundary. Which assistant data, logs, files, and connector events stay in the stated region or appear in the audit trail?
No training on your data, zero data retention, or DPA-backed data claim Training opt-out scope per product and plan, data-retention settings and defaults, DPA availability and terms, subprocessor list with data-use roles, and what changes if a connected app, API access, or fine-tuning is enabled. What does the vendor's DPA say about training, retention, subprocessors, and how each of those obligations changes if we use an API key, connect a third-party app, or enable a fine-tuning workflow?

Evidence to request

  • The exact secure AI assistant claim and the product plan or feature it applies to.
  • A data-flow map for prompts, files, outputs, logs, connectors, support access, and admin exports.
  • Identity, access, and role-control evidence for users, admins, apps, and integrations.
  • Retention, deletion, data-residency, DPA, subprocessor, and model-training boundaries in writing.
  • AI-specific risk controls for prompt injection, data leakage, tool actions, monitoring, and human escalation.
  • For training-boundary and zero-retention claims: the DPA text, subprocessor list, and which product surfaces, plans, and connected-app configurations are covered or excluded.

Questions to put in front of the vendor

  • For this secure AI assistant claim, which data categories are protected and which are outside the stated scope?
  • What evidence shows the assistant is secure and resilient for our intended workflow, not only for the vendor's general platform?
  • Can admins control access, connectors, retention, export, and model-training settings for this deployment?
  • What logs and audit events would help investigate an incorrect answer, unauthorized access, or data exposure concern?
  • What DPA, subprocessor, and data-residency evidence supports the assistant claim rather than the vendor's general platform claim?
  • Which wording boundary should replace the broad security claim if only selected controls or plans are covered?

Wording boundaries to compare against

  • Includes defined enterprise controls for identity, retention, encryption, and audit logging on specified plans.
  • Business data is not used for model training by default for named products and configurations.
  • Encrypts specified data stores and transfer paths; key-management options and exclusions are documented separately.
  • Designed to manage AI-specific risks such as untrusted inputs, tool access, and data leakage through documented controls.

Have your vendor's exact claim wording ready?

Check a secure AI assistant claim How the evidence method works