Agentic AI and GRC: The AI Proposes, Your Experts Decide

On a risk assessment, the AI starts from the documents you give it and prefills the fields of the pre-assessment and the workshops, quoting the passages it relies on. Each proposal lands in its place in the assessment, ready for review, not as text to be copied somewhere else. What separates it from an assistant is less the model than the scope of action: what the AI may read, what it may write, and what stays subject to human sign-off. In Vailor, nothing is approved without an expert.

No, and that is not a technical limitation. A regulator, an auditor or a board deals with an organisation and with named people, never with a tool: accepting a residual risk, choosing a control or signing off a scope remain human decisions, attributed and dated. The agent does the preparatory work, proposes, sets out its reasoning and flags whatever looks doubtful. The platform records who decided, when, and on what basis.

By going back to the source. When the AI relies on a document, it quotes the passage, and Vailor checks word for word that it is there; a quote that cannot be found is flagged, never erased. An answer you cannot verify is an answer to discard, and that is the test to apply to any tool of this kind. In practice the human review bears on the reasoning and the business context rather than on formatting, and that is where the time moves: out of data entry, into analysis.

No security framework forbids using software to run a risk assessment: what gets examined is the approach, not the tool. Is the method explicit, are the scenarios justified, is each decision owned by a named person, does the documentation reflect the real scope. The sensitive point is therefore not the AI but the audit trail: an assessment whose steps can be replayed and whose trade-offs can be retrieved holds up better than a file with no history behind it.

It puts a duty on your organisation to know and supervise the AI systems it uses. The EU AI Act entered into force in 2024, sorts systems by level of risk and phases its obligations in over time; it also expects the people deploying AI to have a working grasp of what they are running. Whatever classification applies to a given tool, the starting point is the same: knowing which AI systems are in use, for what purpose, on which data and under whose supervision. That inventory is a GRC object like any other, with a scope, risks, controls and evidence.

As SaaS, your documents and registers are stored in France (AWS Paris region), AI processing runs on AWS Bedrock in EU regions, and nothing is used to train a model. You choose the model used, including a self-hosted one. Beyond hosting, the questions worth putting to any vendor are about use: does the content you submit feed the training or improvement of a model, is it passed to a third party, how long is it kept, who can reach it on the supplier side. Insist that the answers sit in the contract and in the security documentation, not only on a web page. It is also what your own customers will ask you about your own supply chain.

A general-purpose assistant knows nothing of your risk register, your controls or the state of your evidence: it returns text that somebody then re-keys into a spreadsheet, losing the source and the date on the way. GRC work is made of links between a risk, a control, a requirement and a piece of evidence, and those links are exactly what breaks at every copy and paste. A platform holds the state, keeps the history and makes the output defensible. In Vailor, the AI prefills the risk assessment directly, where your experts approve it.