AI in GRC software: 12 questions to ask your vendor

Back to Blog

Last updated: 3 November 2026 · By Gaute Wangen

AI is arriving in GRC software. Platforms offer to draft policies, map controls across frameworks, suggest risks and fill in vendor questionnaires. Some of this is genuinely useful. But a GRC platform holds some of the most sensitive information an organisation has, so before you switch AI features on, or choose a platform because of them, you should know what happens to your data. This article is about the security and privacy of AI in GRC software: twelve questions to ask your vendor, what to look for in the answers, and what to decide on your own side.

Why AI in GRC software deserves extra scrutiny

A GRC platform is not an ordinary business tool. Your risk register is a map of your weaknesses. It lists your critical systems, known vulnerabilities, incidents, open audit findings and the measures you have not yet put in place. Your records of processing and DPIAs describe the personal data you hold. Supplier assessments say which suppliers you depend on and where they fall short.

When a GRC platform uses AI, some of that information may be sent to an AI model. Often that model is run by another company. That raises questions any security or privacy lead will recognise:

  • Confidentiality: who can see the data, where is it processed, and how long is it kept?

  • Privacy: if your GRC vendor uses an external AI provider to process personal data on its behalf, that provider may be a subprocessor. Under GDPR, a processor needs the controller's prior written authorisation to engage another processor. Ask how the provider is authorised, what agreement applies and where the data is processed.

  • Accountability: if the AI suggests a risk rating or writes a policy, who is responsible for it? For organisations covered by NIS2, management must approve the cybersecurity risk-management measures and oversee how they are implemented. Whatever rules apply to you, someone in the organisation has to own the result, whatever tools were used to produce it.

None of this means you should avoid AI. It means you should be the one deciding how it is used.

12 questions to ask your GRC vendor about AI

Your data

1. Which features use AI, and what data does each one send?

Ask for a list, feature by feature. "We use AI to improve the product" is not an answer. You want to know whether the feature sends the text you selected, a whole record, or your entire register.

2. Which AI providers receive our data, and where is it processed?

If the vendor uses an AI provider to process personal data on its behalf, the provider should be on the vendor's subprocessor list, with the country where processing takes place. Ask how you will be told before a new AI provider is added.

3. Is our data used to train or improve AI models?

This applies both to the vendor's own models and to the AI provider's. Ask whether the answer is written in the contract or the data processing agreement, or only set in a menu that could change.

4. How long are prompts and responses kept, and by whom?

The vendor may keep a log, and the AI provider may keep data for a period for abuse monitoring. Ask for both, and ask whether you can delete them.

5. Is our data kept separate from other customers' data?

Many AI features run on models shared by all customers. That can be fine, as long as your data is isolated, is never shown to other customers, and is only used as your agreement allows. Ask how this is done.

Your control

6. Can we switch AI off, and is it off until we switch it on?

You should be able to turn AI off for the whole organisation, and ideally feature by feature. AI that is on by default means data may have been sent before anyone decided anything.

7. Can we use our own AI instead of yours?

Whether this matters depends on your needs. If your organisation already has an approved AI service with agreed terms, being able to use it can save you from assessing a new AI provider for every supplier.

With "bring your own AI", you choose the AI service, for example the assistant your organisation already uses, and connect it to the GRC platform. The connection often uses MCP (Model Context Protocol), an open standard for connecting AI assistants to other systems. MCP is only the connection. What the assistant reads is processed by the AI service you chose, under your agreement with that service, not under your agreement with the GRC vendor. So you decide which data the assistant may access, and you assess the AI service's terms and data handling as you would any other supplier.

8. What can the AI do, and with whose permissions?

Can it only read, or can it create and change records? Ideally it works within the access rights of the user it works for, so that a user without access to the HR risks cannot reach them through an AI assistant. If the feature uses a service account instead, ask how its access is limited, controlled and logged.

Trust in the results

9. Does a person approve AI output before it is saved?

A suggested risk, control mapping or policy text should be a draft until a person accepts it. This is often called keeping a human in the loop. Ask whether AI-generated content is marked as such.

10. Is AI activity recorded in the audit trail?

When an auditor asks who changed a risk assessment, "the AI" is not a good answer. The trail should show what the AI suggested, who approved it and when.

11. How do you deal with wrong answers?

AI models can produce confident, wrong answers. Ask how the vendor tests the features, whether the AI shows the sources it used (for example the control or regulation text), and how you report errors.

The vendor's own AI governance

12. How do you govern your own use of AI?

Ask whether the vendor has an AI policy, whether AI is covered by its risk assessments, and whether its staff use AI tools with customer data when they develop and support the platform. ISO/IEC 42001, the management system standard for AI, is a useful reference, and the EU AI Act sets duties for providers and deployers of AI systems. A vendor does not need a certificate to give a clear answer.

Summary: what to look for

#

Topic

Look for

Ask more if

1

Features and data

A list per feature

The answer is "AI is everywhere"

2

AI providers

On the subprocessor list, with countries

The provider or country is unclear

3

Model training

Stated in the contract

It is only a setting

4

Retention

Specific periods

"As long as necessary"

5

Separation

Isolated data, used as agreed

It is unclear how data is isolated

6

Switching off

Off by default, or per feature

It can only be switched off per user

7

Own AI

Options that fit your needs

You need your own AI and cannot use it

8

Permissions

User rights, or tightly limited access

Wide access without logging

9

Human approval

Drafts until approved

AI writes directly to registers

10

Audit trail

Suggestions and approvals logged

No record of what the AI did

11

Errors

Sources shown, testing described

"The model is very accurate"

12

Own governance

An AI policy, AI in risk assessments

No clear answer

Questions to ask yourselves

The vendor's answers are only half the work. Before you switch AI on, agree internally on:

  • Your AI policy. Which information may be processed by AI, and with which services? Many organisations allow AI for public information but not for personal data or security details.

  • Whether a DPIA is needed. GDPR Article 35 requires a data protection impact assessment when processing is likely to result in a high risk, in particular when new technology is used. AI that processes personal data from your registers is a typical case to assess.

  • Who approves AI output. Decide who may accept AI suggestions in the risk register, and make sure they have the knowledge to judge them.

  • Your own risk assessment. Treat the AI feature like any new supplier or system. A risk and vulnerability analysis of the AI use is a good place to start.

How Cyrigo answers these questions

We built the Cyrigo platform so that you decide whether and how AI is used on your data.

  • We never send your data to AI services unless you choose to. The platform does not send your registers to an AI provider of its own. (Questions 1–6)

  • Bring your own AI. You can connect your own AI assistant through our MCP server. You choose the AI service, and what the assistant reads is processed under your agreement with that service, so assess it as you would any other supplier. (Question 7)

  • The AI works within your access rights. Connecting an assistant is off until your administrator switches it on. The assistant reads only what the user behind its access token may read, and it cannot write anything to a register itself. (Questions 6 and 8)

  • A human in the loop. Changes suggested through the MCP server arrive as drafts, and nothing reaches a register until a person with the right to do so publishes it in the platform. (Question 9)

  • Our own use of AI. We use AI tools to support development and operations, but they never receive personal data or your content. Read more under Use of AI on our Security page. (Question 12)

  • Our subprocessors are listed in our privacy policy. The platform is developed in Norway and hosted in Scandinavia.

See how the platform works on the product page, or book a demo.

AI in GRC: frequently asked questions

What is AI GRC?

AI GRC usually means governance, risk and compliance software that uses AI, for example to draft policies, map controls between frameworks, suggest risks or answer security questionnaires. The term is also used for the governance of AI itself, meaning how an organisation manages the risks of the AI it uses. This article is about the first meaning.

Is AI in GRC software GDPR compliant?

It can be, but it depends on the roles and how the data flows. All processing of personal data needs a lawful basis. If the vendor uses an AI provider on its behalf, that provider is usually a subprocessor, which needs your prior authorisation and an agreement with the vendor. If you connect your own AI service, you engage that provider yourself and assess its agreement. Where data is transferred outside the EEA, a transfer mechanism is needed. Questions 2 to 4 above cover this.

Can AI do a risk assessment?

AI can help: it can suggest threats, find similar risks and draft descriptions. But a risk assessment depends on knowing your own systems, people and priorities, and someone has to own the result. Treat AI output as a draft that a person reviews.

Will AI replace GRC analysts?

It is more likely to change the work than to replace it. AI can take over routine drafting and mapping, but accountability stays with people: for organisations covered by NIS2, the directive places it on management. Judgement about risk still needs people who know the organisation.

What is an MCP server?

MCP stands for Model Context Protocol, an open standard for connecting AI assistants to other systems. An MCP server lets an AI assistant you already use read from or work with a system such as a GRC platform, within the limits that system sets.

Sources

Back to Blog