Skip to main content
AI GRC guide

AI GRC: AI proposes, your people approve

AI GRC means using AI inside governance, risk and compliance work to draft, map, summarise and propose. It helps most when a person still decides what is accepted, attested and signed off.

This guide is for MSPs, MSSPs, consultancies and vCISOs weighing AI in their GRC tooling. Most of it applies to any vendor. The GetCybr section says plainly what our AI does today.

What AI GRC is (and is not)

The term is used two ways. Here it means AI inside GRC tools. The other meaning is governing a client's own AI systems, which is a separate job covered by standards such as ISO/IEC 42001.

AI in GRC is a fast assistant, not a decision-maker. It does well at work that is slow to write and quick to check:

  • Drafting. First versions of policies, procedures and assessment narratives that a person then edits.
  • Mapping. Matching an existing policy or control to the requirements of one or more frameworks, and showing where coverage is thin.
  • Summarising evidence. Turning long documents, interview answers and exports into a short account a reviewer can check.
  • Proposing actions. Suggesting remediation steps and next questions for a practitioner to accept, change or reject.

People must keep three decisions, because each carries accountability:

GRC decisions that stay with people
DecisionWhy it stays with a person
Risk acceptanceAccepting a risk is a business decision with an owner who answers for it. A model cannot hold that accountability.
AttestationSaying a control operates is a statement a named person signs, so a person has to have checked it.
Audit sign-offThe auditor or assessor forms the opinion. Software can prepare the evidence, not give the assurance.

AI output can be wrong or incomplete. Treat every draft as a proposal to review, however it is produced.

What MSPs need from AI in GRC

An MSP works across many clients, so four things matter more than the list of AI features.

  • Per-client data separation. One client's records must never become another client's context. Ask how the AI is kept inside the client it is working on.
  • Approval before anything is saved. AI output should arrive as a proposal. A reviewer applies it, so the record shows who decided.
  • Choice of model, including self-hosted. Some clients accept a hosted model with your own key. Others need the model to run on infrastructure you operate.
  • An audit trail. You should be able to show a client or auditor what was proposed, who approved it and when.

For the wider platform questions, see our guide to GRC software for MSPs.

What GetCybr's AI does today

GetCybr is a compliance platform for MSPs, consultancies and vCISOs. Here is what its AI does now, and where it stops.

  • Run security assessments as a conversation: GetCybr's AI interview asks the questions and drafts control maturity, and nothing is saved until your reviewer applies it.
  • A built-in vCISO assistant answers questions about each client and proposes actions. Nothing runs until you confirm.
  • Turn a one-line prompt into a control-mapped policy document, then send it through your normal approval workflow.
  • Upload an existing policy and AI maps it to framework controls, showing coverage and gaps.
  • Self-hosting is part of GetCybr Enterprise: run GetCybr in your own environment with the local LLM, so AI processing runs on infrastructure you operate.

Every account can also bring its own key for OpenAI or Anthropic. Enterprise is quoted rather than priced publicly, and our guide to self-hosted GRC covers what running it yourself involves. GetCybr is platform software; a human vCISO is supplied and priced separately.

How to evaluate AI GRC software

Put these questions to every vendor, including us, and get the answers in writing.

Checklist for evaluating AI GRC software
CriterionWhy it matters for an MSPQuestion to ask the vendor
Human approvalAn unreviewed AI write to a client record becomes your liability the moment a client relies on it.Does AI output wait for a named reviewer to apply it, or can it change records and run actions by itself?
Client separationMixing one client’s data into another’s answers breaks confidentiality and your contracts.What stops the AI using one client’s data when it works on another?
Model choiceClient contracts differ. One may accept a hosted model, another may rule it out.Which models can I use, can I bring my own key, and is a local model an option?
Where processing runsData-location terms usually cover the AI step as well as storage.Where does inference run, and which parts still call out to the vendor or a third-party model provider?
Audit trailYou may need to show who proposed and who approved each change.Can I export a record of what the AI proposed, who approved it and when?
Visible reasoningA reviewer can only approve what they can check, such as which control a policy was mapped to and why.Does the output show its mapping, coverage and gaps, or only a score?
Use of your dataClient material may be covered by confidentiality terms that exclude model training.Is client content used to train any model, and where is that written down?
Pricing of AI featuresUsage-metered AI can make per-client cost hard to predict.Is AI included, metered or an add-on, and does cost change with the number of clients?

Compare platforms side by side in our vCISO platform guide. AI is only one part of the choice.

Where AI fits across a client programme

AI speeds up the paperwork, but the programme is still built on the compliance frameworks your clients answer to and on their supply chain. See our guide to third-party risk management software for MSPs for the supplier side.

Help Center

FAQs have moved to the Help Center

Find current answers for the topics covered on this page in our consolidated FAQ.

See human-approved AI in your own workflow

Book 15 minutes to see how the AI interview, vCISO assistant and policy tools fit your practice.