Requirements catalogue
AI data protection requirements for your RFP
By Maskflare · Published · Updated
Short answer
This catalogue lists 78 requirements to put in an RFP or security questionnaire for a tool that stops personal data and secrets from reaching AI models. Each has an ID, a reason, a priority and a framework mapping. Pick the ones you need, mark them must, should or could, and export them.
How to use this catalogue
Buying a tool for AI data protection usually starts from a vendor's own feature list. This catalogue starts from the buyer's side instead. It covers 78 requirements in 9 groups, from discovering which AI tools people use to the evidence a vendor should give for GDPR, the Swiss nDSG, HIPAA, PCI DSS, the EU AI Act, ISO/IEC 42001 and the NIST AI Risk Management Framework.
Every requirement has an ID you can quote in a questionnaire, a description of what to ask for, a reason it matters, a default priority and the frameworks it relates to. The builder below lets you select the ones you need, change must, should or could, and export the result. It works in your browser, needs no sign-up and sends nothing anywhere.
Who wrote this. Maskflare wrote the catalogue and sells a product in this category, so we kept it free of product names and did not score any vendor, ourselves included. Maskflare does not meet every requirement here. Browser response rewriting (MSK-08), topic filtering (POL-08), screening chat prompts for injection (AGT-07) and completed SOC 2 Type II and ISO 27001 audits (CMP-08, CMP-09) are examples where we have gaps today. Use the list to ask us the same questions you ask everyone else. See how we compare with other tools.
Priorities and framework mappings
Default priorities reflect a typical European company with employees, applications and agents sending data to AI models. Must means a vendor without it should not pass. Should means it is expected and a gap needs a plan. Could is useful for some buyers.
The framework tags point to the article, rule or function where the topic is covered. They do not say a product satisfies the law. This is a starting point, not legal advice, and it reflects the texts as of November 4, 2026. EU AI Act dates follow Regulation (EU) 2026/1744: AI literacy under Article 4 applies now, and Annex III high-risk obligations apply from 2 December 2027. Where we could not confirm a clause number against a primary source, we cite the framework only. No ISO/IEC 42001 clause numbers are given because the standard is paywalled. See our guides on GDPR, the EU AI Act, the Swiss nFADP, HIPAA and PCI DSS for the detail.
- GDPR: GDPR (Regulation (EU) 2016/679)
- nDSG: Swiss Federal Act on Data Protection (nDSG / revFADP, SR 235.1)
- HIPAA: HIPAA Privacy and Security Rules (45 CFR Part 164)
- PCI DSS: PCI DSS v4.0.1
- EU AI Act: EU AI Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744)
- ISO 42001: ISO/IEC 42001 (AI management systems)
- NIST AI RMF: NIST AI Risk Management Framework 1.0 (Govern, Map, Measure, Manage)
78 of 78 selected
Must 38 · Should 34 · Could 6
Discovery (7)
You cannot protect or approve what you cannot see. These requirements cover finding which AI tools people use.
The product lists the AI apps and AI-enabled services used on managed devices, with the number of users and devices, the first date seen and the volume of data sent to each.
Why it matters: An inventory is the starting point for policy, AI literacy measures and risk assessment. Without it the other controls cover only the tools you already know about.
- GDPR: Art. 30
- EU AI Act: Art. 4
- ISO 42001
- NIST AI RMF: Govern (GOVERN 1.6, AI system inventory)
For the same AI app, the product distinguishes a company workspace account from a personal account, so that policy can allow one and block the other.
Why it matters: Most leakage through approved tools happens in personal accounts, which usually have no processor contract and different retention and training terms.
- GDPR: Art. 28
- nDSG: Art. 9
- NIST AI RMF: Govern (GOVERN 6, third-party AI)
Each discovered app can be set to approved, tolerated or blocked, and the decision is enforced and visible to users.
Why it matters: A written AI policy that the tooling cannot enforce leaves a gap between the policy and what staff actually do.
- EU AI Act: Art. 4
- ISO 42001
- NIST AI RMF: Govern
Each app gets a risk rating based on stated factors such as vendor jurisdiction, use of customer data for training, retention, certifications and security posture, and the factors are visible.
Why it matters: Reviewers need a consistent basis for deciding between hundreds of tools. A score with hidden inputs cannot be audited.
- GDPR: Art. 28, Art. 44 to 49
- nDSG: Art. 16 and 17
- NIST AI RMF: Map, Govern (GOVERN 6)
Discovery records metadata only: which app, which user or department, how much data. It does not record prompts, page contents or browsing history.
Why it matters: A discovery feature that collects content is itself a data protection and works council problem.
- GDPR: Art. 5(1)(c), Art. 25
- nDSG: Art. 6, Art. 7
The product reports AI assistants and AI features built into SaaS tools, browser extensions and plugins, not only dedicated AI websites.
Why it matters: A growing share of AI use arrives as a feature switch inside tools that were approved before the feature existed.
- EU AI Act: Art. 4
- NIST AI RMF: Map
The vendor states what is and is not covered on unmanaged personal devices and mobile apps, and offers a control or a compensating measure where coverage is limited.
Why it matters: Few tools cover unmanaged devices. The requirement is an honest statement of the boundary, so you know what to close with policy or identity controls.
- GDPR: Art. 32
- ISO 42001
- NIST AI RMF: Map
Policy and coaching (8)
Blocking everything pushes use to personal devices. These requirements cover how rules are defined, rolled out and explained to users.
A rule can allow, alert, warn, mask, redact, partially mask, hash or block, chosen per data type and per destination.
Why it matters: Different data deserves different handling. A card number may be blocked while a name is masked, and one action for everything produces either leaks or constant interruptions.
- GDPR: Art. 5(1)(c), Art. 25, Art. 32
- nDSG: Art. 6, Art. 7
- PCI DSS: 3.4.1
Users can be shown a prompt before sending, with options such as send anyway with a reason, mask and send, or cancel. The reason is stored with the event.
Why it matters: A short moment of friction changes behavior more than a silent block, and the recorded reasons show where the policy does not fit real work.
- EU AI Act: Art. 4
- ISO 42001
- NIST AI RMF: Govern, Manage
Warnings and block messages can carry your own wording, a link to your AI policy and a pointer to the approved alternative.
Why it matters: People follow a rule when they know where to go instead. A generic denial creates support tickets.
- EU AI Act: Art. 4
- NIST AI RMF: Govern
Every rule can run in alert mode first, counting matches without changing traffic, then be switched to enforcement.
Why it matters: False positives found in production traffic erode trust in the tool. Tuning on real data before enforcing avoids that.
- ISO 42001
- NIST AI RMF: Measure, Manage
Rules can differ by user group, department, environment and application, with inheritance from a default.
Why it matters: Legal, HR, engineering and support handle different data. One global policy is either too loose for some or unusable for others.
- GDPR: Art. 25
- nDSG: Art. 7
- NIST AI RMF: Govern
Changes to rules can require a second person to approve, and the change history shows who changed what and when.
Why it matters: A policy that one administrator can weaken silently is a weak control for audit purposes.
- GDPR: Art. 32
- HIPAA: 45 CFR 164.308
- ISO 42001
Known safe values and approved exceptions can be allowlisted by value, pattern or user, with an owner and an expiry date.
Why it matters: Permanent exceptions accumulate. Expiry forces a periodic review.
- ISO 42001
- NIST AI RMF: Manage
Policy can act on the subject of a prompt, for example unreleased financial results or a merger, without relying on a specific identifier format.
Why it matters: Confidential business information often has no pattern to match. Identifier-based detection will not catch it, so buyers with this risk should ask whether topic rules exist.
- GDPR: Art. 32
- ISO 42001
- NIST AI RMF: Map, Manage
Detection depth (12)
A vendor claim of PII detection says little. These requirements test what is detected, how it is validated and whether the result can be explained.
Detection covers national identification, tax and social security numbers for the EU member states you operate in, using each country's official format and check digits.
Why it matters: A bare digit pattern flags order numbers and misses formatted values. Country-specific validation is what separates a usable detector from a noisy one.
- GDPR: Art. 4(1), Art. 5(1)(c)
- nDSG: Art. 6
Detection covers the Swiss AHV/AVS number, Swiss IBANs, the UID/CHE business identifier and Swiss address and phone formats in German, French and Italian.
Why it matters: Many global detectors treat Switzerland as an afterthought. Swiss banking, insurance and public bodies need these validated, not guessed.
- nDSG: Art. 6, Art. 8
Where an identifier has a checksum or check digit (payment cards, IBAN, national IDs), the detector validates it and the vendor lists which validators exist.
Why it matters: Checksums cut false positives without hurting recall. The number of validated identifier types is a more useful measure than the number of detectors.
- GDPR: Art. 5(1)(d)
Primary account numbers are detected with a Luhn check, and card verification codes and similar sensitive authentication data can be blocked outright.
Why it matters: Card data in a prompt can bring an AI provider into the scope of your cardholder data environment. Authentication data must not be stored after authorization.
- PCI DSS: 3.3.1, 3.4.1, 3.5.1, 12.8
Personal names, places and similar details in free text are found with a named-entity model that runs inside the vendor's own boundary, with no call to a third-party AI service.
Why it matters: Names have no checksum. If detecting them means sending text to another AI provider, the control creates the exposure it is meant to remove.
- GDPR: Art. 5(1)(f), Art. 28
- nDSG: Art. 8, Art. 9
Dates, addresses, phone numbers and context words are recognized in the languages your staff write in, and the vendor lists them.
Why it matters: A detector trained on English context misses a German address or an Italian date of birth.
- GDPR: Art. 5(1)(c)
- nDSG: Art. 6
Health identifiers and terms, and other special categories such as religion, union membership and criminal records, are detected, with a documented way to review matches before enforcing.
Why it matters: Special categories raise the legal bar. The 18 Safe Harbor identifier types in the US and Article 9 data in the EU both need more than a name list.
- GDPR: Art. 9, Art. 10
- nDSG: Art. 5, Art. 6
- HIPAA: 45 CFR 164.514(b)
API keys, access tokens, private keys, connection strings and passwords are detected across a documented list of provider formats, including in code and configuration files.
Why it matters: Secrets pasted into prompts are a routine incident in developer workflows, and a leaked key stays valid until someone rotates it.
- GDPR: Art. 32
- PCI DSS: Requirement 8
- NIST AI RMF: Manage
You can add custom patterns and term dictionaries, and match exact values from your own records such as customer or case numbers, without uploading the records in clear text.
Why it matters: Customer, employee and case numbers have no public format. Only your own data can define them.
- GDPR: Art. 5(1)(c)
- nDSG: Art. 6
For every match the product reports which rule fired and which evidence (pattern, checksum, context word, model) supported it, so that a false positive can be traced and fixed.
Why it matters: Security teams cannot tune what they cannot inspect, and auditors ask why a value was or was not masked.
- GDPR: Art. 5(2)
- EU AI Act: Art. 13 (high-risk systems)
- ISO 42001
- NIST AI RMF: Measure (transparency)
The vendor publishes recall and false positive figures with the test set size, the data origin (synthetic or real), the languages and how it can be reproduced, and states where results are weaker.
Why it matters: Accuracy claims without a method cannot be compared. A request for the test set and the failures is a fair RFP question.
- EU AI Act: Art. 15 (high-risk systems)
- ISO 42001
- NIST AI RMF: Measure
Text inside uploaded files (PDF, Word, Excel, PowerPoint, email attachments) and text in images is scanned, with the supported formats and the limits of image text recognition stated.
Why it matters: Users upload files rather than type. A control that sees only typed text misses a large share of the data.
- GDPR: Art. 32
- PCI DSS: 3.5.1
Masking and reversibility (8)
Removing data is only useful if the AI answer is still usable. These requirements cover how values are replaced and restored.
Values can be replaced by tokens before the request leaves and restored in the answer, so the user sees the real values again.
Why it matters: Without restoration, staff paste raw data back in to get a usable answer, which defeats the control.
- GDPR: Art. 4(5), Art. 25, Art. 32
- nDSG: Art. 7, Art. 8
- NIST AI RMF: Manage
Values can also be replaced irreversibly, with no stored mapping, for data that must never leave in any form.
Why it matters: For some data the right design is that nobody, including the vendor, can restore it.
- GDPR: Art. 5(1)(c)
- HIPAA: 45 CFR 164.514(b)
- PCI DSS: 3.5.1
The same value gets the same token across a conversation or document, so the model can reason about relationships between entities.
Why it matters: Random replacement breaks questions such as which of these two customers owes more.
- NIST AI RMF: Measure
Restoration works on streamed responses, including when a token is split across chunks.
Why it matters: Streaming is the default in chat interfaces. A control that only works on complete responses is slow or breaks the stream.
The mapping from tokens to real values is encrypted, held for a stated period, then deleted, and the retention period is part of the contract.
Why it matters: The mapping is the sensitive part of a reversible design. Its lifetime and protection define the residual risk.
- GDPR: Art. 5(1)(e), Art. 32
- nDSG: Art. 8
- HIPAA: 45 CFR 164.312
When the protection layer is unavailable, you can choose to fail closed (traffic stops) or fail open (traffic passes), per application, and the default is documented.
Why it matters: A silent bypass during an outage is the kind of exception an auditor asks about.
- GDPR: Art. 32
- PCI DSS: Requirement 12
- ISO 42001
Masking applies to embedding requests, batch files and retrieval pipelines, with stable tokens so search still works.
Why it matters: Personal data embedded in a vector store is hard to find and delete later.
- GDPR: Art. 5(1)(c), Art. 17
- EU AI Act: Art. 10 (high-risk systems)
- NIST AI RMF: Map, Manage
In browser-based AI assistants, sensitive values or restored content in the model's answer can be rewritten or masked in the page the user sees.
Why it matters: Responses can repeat or invent sensitive details, and some buyers want the screen controlled as well as the request. Many tools act only on the outgoing prompt.
- GDPR: Art. 32
- NIST AI RMF: Manage
Data handling (9)
The protection layer sees your data before it masks it. These requirements cover what the vendor stores, where and for how long.
The product does not store prompt, response or file content. Logs and events hold metadata and counts only, and this is stated in the contract and not only in marketing copy.
Why it matters: A tool that stores the text it is meant to protect becomes a new concentration of the same risk.
- GDPR: Art. 5(1)(c), Art. 5(1)(e), Art. 25
- nDSG: Art. 6, Art. 7
- HIPAA: 45 CFR 164.502(b)
Your data is never used to train or improve models, including the vendor's own detection models, unless you opt in in writing.
Why it matters: Data used for training is hard to recall and can surface in other customers' outputs.
- GDPR: Art. 5(1)(b), Art. 28(3)(a)
- nDSG: Art. 9
You can choose where processing happens (EU/EEA or Switzerland), the vendor states every location where content or metadata is processed or backed up, and support access from other regions is described.
Why it matters: Transfer rules depend on where data is processed and who can reach it, not only on the primary region.
- GDPR: Art. 44 to 49
- nDSG: Art. 16 and 17
The vendor provides a current sub-processor list with locations and gives advance notice of changes, with a right to object.
Why it matters: You remain responsible for the processing chain, and the AI provider you route to is part of it.
- GDPR: Art. 28(2), Art. 28(4)
- nDSG: Art. 9
A data processing agreement covering GDPR and nDSG is available before the trial, with transfer safeguards where data leaves the EEA or Switzerland.
Why it matters: Without a processor contract the vendor cannot lawfully receive personal data on your behalf.
- GDPR: Art. 28(3), Art. 46
- nDSG: Art. 9, Art. 16 and 17
- HIPAA: 45 CFR 164.308(b)
Retention periods for event metadata, audit logs and the token store are stated and configurable within limits.
Why it matters: Storage limitation applies to the control's own records as much as to the data it protects.
- GDPR: Art. 5(1)(e)
- nDSG: Art. 6
- EU AI Act: Art. 26(6) (high-risk deployers, logs of at least six months)
Data is encrypted in transit and at rest, provider API keys are stored encrypted and never logged, and customer-managed keys are available or on the roadmap.
Why it matters: Encryption is a named measure in the main data protection laws, and key handling decides who can read stored values.
- GDPR: Art. 32(1)(a)
- nDSG: Art. 8
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 3, Requirement 4
On request and at contract end, your configuration, events and token store are deleted or returned, and deletion is confirmed in writing.
Why it matters: An exit plan is part of processor oversight, and an AI tool is easy to adopt and hard to unwind.
- GDPR: Art. 28(3)(g)
- nDSG: Art. 9
The vendor states in which component and location unmasked text is handled before masking, and offers a deployment in which it never leaves your infrastructure.
Why it matters: The masking layer is itself a recipient of the unmasked data. This answer decides whether it is a processor and whether a business associate agreement is needed.
- GDPR: Art. 28, Art. 44 to 49
- nDSG: Art. 9
- HIPAA: 45 CFR 164.308(b)
Coding agents and MCP (8)
Agents read files, run commands and call tools without a person typing each prompt. These requirements cover that traffic.
Prompts, file contents and tool results handled by the coding agents you allow are inspected, through hooks, a proxy or a gateway, and the vendor lists which agents and versions are supported.
Why it matters: Coding agents send source code, configuration and logs to a model by default. That traffic often bypasses the browser controls.
- GDPR: Art. 32
- NIST AI RMF: Map, Manage
Credentials in source files, environment files and command output are detected before they reach the model.
Why it matters: A key in a repository that an agent reads ends up in a provider's request log.
- GDPR: Art. 32
- PCI DSS: Requirement 8
Arguments sent to Model Context Protocol (MCP) tools and the results coming back are masked, so data is protected on the way to the model and on the way out of internal systems.
Why it matters: An agent connected to a database or ticketing system can pull personal data into the model without any user typing it.
- GDPR: Art. 5(1)(c), Art. 25, Art. 32
- nDSG: Art. 7, Art. 8
- NIST AI RMF: Manage
Which MCP servers and tools a group may use, and whether read or write, is set by policy, with a request and approval process for new ones.
Why it matters: Least privilege for agents is the same principle as for people, and the number of available tools grows quickly.
- GDPR: Art. 32
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 7
- ISO 42001
The product lists the MCP servers and agent tools in use across the organization.
Why it matters: Shadow AI now includes shadow agents. The inventory requirement from the discovery group applies to tools as well as apps.
- EU AI Act: Art. 4
- NIST AI RMF: Govern (GOVERN 1.6), Map
Tool descriptions and tool results are checked for instructions that try to hijack the agent, and flagged or blocked.
Why it matters: Content fetched by a tool is untrusted input. A poisoned document can instruct the agent to send data elsewhere.
- GDPR: Art. 32
- EU AI Act: Art. 15 (high-risk systems)
- NIST AI RMF: Measure, Manage
Prompts typed into chat assistants are screened for injection and jailbreak attempts, with the detection method and its false positive rate described.
Why it matters: This is a different problem from data protection. Some buyers want one tool for both, and should say so explicitly in the RFP.
- EU AI Act: Art. 15 (high-risk systems)
- NIST AI RMF: Measure, Manage
Agent tool invocations are recorded as metadata (tool, user, time, outcome) without storing content, and can be exported.
Why it matters: When an agent acts, someone needs to reconstruct what it did.
- GDPR: Art. 5(2), Art. 32
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 10
- EU AI Act: Art. 12 (high-risk systems)
Identity and SIEM (8)
The console holds policy and event data. These requirements cover who can use it and how events reach your security team.
Console access uses SAML 2.0 or OpenID Connect with your identity provider, and local passwords can be disabled.
Why it matters: Central identity control gives you joiner, mover and leaver handling and conditional access.
- GDPR: Art. 32
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 8
Users and groups can be provisioned through SCIM or directory sync, and policy can target synced groups.
Why it matters: Policy that follows your existing groups avoids a second user directory to maintain.
- GDPR: Art. 32
- PCI DSS: Requirement 7
Roles are separated, at least into administrator, security analyst, auditor (read only) and developer, with least privilege by default.
Why it matters: People who tune policy should not be the only ones who can read the audit trail.
- GDPR: Art. 32
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 7
- ISO 42001
Administrator accounts require multi-factor authentication, enforced by the product or by your identity provider.
Why it matters: The console can weaken protection for the whole organization, so it is a high-value target.
- GDPR: Art. 32
- PCI DSS: Requirement 8
Every change to policy, users and integrations is logged with who, what and when, and the log can be exported and cannot be edited by administrators.
Why it matters: Evidence of control changes is a standard audit request.
- GDPR: Art. 5(2), Art. 32
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 10
- EU AI Act: Art. 12 (high-risk systems)
A per-event stream goes to your SIEM or any HTTPS endpoint, carrying metadata and counts only and never prompt or response content.
Why it matters: Security operations need the events. Putting prompt content in the SIEM would spread the exposure to another system.
- GDPR: Art. 5(1)(c), Art. 32, Art. 33
- HIPAA: 45 CFR 164.312
- PCI DSS: Requirement 10
- NIST AI RMF: Measure, Manage
Alerts can go to chat and incident tools, and webhook payloads are signed so the receiver can verify them.
Why it matters: Unsigned webhooks can be spoofed to hide or fake incidents.
- GDPR: Art. 33
- NIST AI RMF: Manage
Reports can show departments or pseudonymous identifiers instead of names, with re-identification limited to defined roles and cases.
Why it matters: Monitoring staff use of AI raises employee privacy and works council questions. Showing groups by default reduces them.
- GDPR: Art. 5(1)(c), Art. 25
- nDSG: Art. 6, Art. 7
Deployment (8)
A control that is hard to roll out does not get rolled out. These requirements cover how the product reaches devices and applications.
The device component installs through your MDM or software distribution tools and enrolls without manual steps per user.
Why it matters: Per-user installation does not scale past a pilot.
- GDPR: Art. 32
- ISO 42001
Windows, macOS and Linux are supported, with versions listed, and unsupported systems are stated.
Why it matters: Developers often run Linux, and a gap there is a gap for the highest-risk data.
Protection can work at the network or device level without a browser extension, and the extension is optional.
Why it matters: Extensions can be removed by users, restricted by browser policy or unavailable in some environments.
- GDPR: Art. 32
A deployment in your own cloud account or data center is available, so that unmasked text never leaves your infrastructure.
Why it matters: Regulated sectors and public bodies often need this, and it removes the question of the vendor as recipient.
- GDPR: Art. 28, Art. 44 to 49
- nDSG: Art. 9, Art. 16 and 17
- HIPAA: 45 CFR 164.308(b)
Users cannot disable or remove the device component without an administrator, and removal or bypass is reported.
Why it matters: A control that a user can switch off in a minute is advice, not enforcement.
- GDPR: Art. 32
- ISO 42001
Applications can be protected by changing a base URL, using an SDK or calling an API, without redesigning the application, and the supported providers are listed.
Why it matters: Engineering teams adopt what takes an afternoon, not a quarter.
- GDPR: Art. 25
- NIST AI RMF: Manage
The vendor publishes added latency with the conditions of the test, and an availability commitment with the failure behavior from MSK-06.
Why it matters: A median figure for small requests says little about your largest prompts. Ask for the conditions.
- GDPR: Art. 32(1)(b)
- PCI DSS: Requirement 12
You can start with a pilot group in monitor mode, widen it in steps and roll back a policy or version quickly.
Why it matters: Staged rollout limits the impact of a bad rule.
- ISO 42001
- NIST AI RMF: Manage
Compliance mapping and assurance (10)
These requirements cover the evidence a vendor gives you for your own compliance work, and the assurance behind the vendor itself.
The vendor gives you material for your records of processing and data protection impact assessment, and states which GDPR duties remain yours.
Why it matters: The tool can supply evidence. The controller still owns the lawful basis, the assessment and the record.
- GDPR: Art. 25, Art. 30, Art. 32, Art. 35
The vendor documents how its processing fits the Swiss act, including processor terms and disclosure abroad, and offers Swiss-specific detection from DET-02.
Why it matters: Swiss law has its own article numbers, breach notification route and personal criminal fines, and a GDPR-only statement does not cover them.
- nDSG: Art. 8, Art. 9, Art. 16 and 17, Art. 22
A profile covers the Safe Harbor identifiers, and the vendor will sign a business associate agreement or offers a deployment that makes one unnecessary.
Why it matters: Masking reduces what an AI provider receives but is not a substitute for the agreements HIPAA requires.
- HIPAA: 45 CFR 164.502(b), 164.514(b), 164.308(b)
A profile detects primary account numbers and sensitive authentication data and blocks or masks them before the AI provider sees them.
Why it matters: Keeping card data out of AI requests keeps the provider out of your scope.
- PCI DSS: 3.3.1, 3.4.1, 3.5.1, 12.5.2, 12.8
The vendor explains how its features support AI literacy and an inventory of AI tools, which AI Act duties it does not address, and the dates that apply to you.
Why it matters: For organizations that only use AI, the main duties are AI literacy (Art. 4) and, for high-risk uses covered by Annex III, deployer duties that apply from 2 December 2027 under Regulation (EU) 2026/1744.
- EU AI Act: Art. 4, Art. 26, Annex III
The vendor can show how its features give evidence for an AI management system, and says whether it holds or is pursuing certification itself.
Why it matters: Organizations running an AI management system need inputs such as an inventory, risk assessment records and monitoring evidence.
- ISO 42001
The vendor maps its features to the Govern, Map, Measure and Manage functions and states which it does not cover.
Why it matters: A function-level mapping is easy to publish. Be cautious of claims of full alignment.
- NIST AI RMF: Govern, Map, Measure, Manage
A completed SOC 2 Type II report is available under NDA, covering at least security and confidentiality, with the audit period stated.
Why it matters: A Type II report tests controls over a period. Ask whether it is complete or in progress, and what the period is.
- HIPAA: 45 CFR 164.308
- PCI DSS: 12.8
- ISO 42001
The vendor holds an ISO/IEC 27001 certificate with a scope that covers the service, and shares the statement of applicability.
Why it matters: Check the scope and the certification body. A certificate for a different entity or product does not cover this service.
- GDPR: Art. 32
- ISO 42001
You can export a report of what is detected and enforced in your configuration, and the vendor documents what the product does not do.
Why it matters: Evidence for auditors and honest limits are both part of the product. A vendor that lists no limits has not looked for them.
- GDPR: Art. 5(2)
- EU AI Act: Art. 13 (high-risk systems)
- ISO 42001
- NIST AI RMF: Measure
Preview and export
Files are built in your browser from the selection above and saved to your device. Nothing is sent to Maskflare or anyone else, and there is no form to fill in. Exports need JavaScript.
Frequently asked questions
How do I use this in an RFP?
Select the requirements that apply to you, set each one to must, should or could, and export the list as Word, CSV, JSON or Markdown. Paste it into the RFP or questionnaire and ask vendors to answer each ID with yes, no or partly, plus evidence. Remove what does not apply and add your own.
Is this a feature list for Maskflare?
No. The requirements are written to apply to any product in this category, and some are ones Maskflare does not meet today. Maskflare wrote the catalogue, so read it with that in mind and check every claim, including ours, against evidence.
Are the framework mappings legal advice?
No. They are a starting point. Each mapping points to the article, rule or function where the topic is covered, not to a statement that a product meets the law. ISO/IEC 42001 is cited at framework level only because the standard is not freely available. Ask your counsel or data protection officer what applies to you.
Is anything sent anywhere when I export?
No. The selection, preview and export run in your browser. The files are built on your device and saved locally. There is no form and no account.
What do must, should and could mean?
The default priority is our suggestion for a typical European company. Must means a vendor without it should not pass. Should means it is expected and a gap needs a plan. Could means it is useful for some buyers. Change them to match your risk.
How often is it updated?
The date at the top shows the last update. We review the framework references when the underlying texts change, for example the EU AI Act timeline set by Regulation (EU) 2026/1744.