Skip to main content
Clause 3.3 of the RICS standard requires firms using AI with a material impact on service delivery to create and operate a risk register. Two words in that sentence carry weight. Create and operate: a register written once and left alone does not satisfy the clause. Material impact: if AI has no material impact on how your firm delivers services, the duty does not bite. Most firms using AI in practice will find it does.

​
The risks the standard names

Whatever else your register holds, the standard expects it to document these four:

Inherent bias

Bias in the system and in the outputs it produces.

Erroneous outputs

The risk of the system simply being wrong.

Limited information

Limits on the quantity and quality of what you can find out about the system and its training data.

Data retention and use

What the supplier retains and does with the data your firm puts in.
Most firms will have more than four. These are the floor, not the ceiling.

​
What each risk has to carry

Clause 3.3 asks for six things per risk:
  1. a description of the risk
  2. its likelihood of materialising and its likely impact
  3. the plan to mitigate and manage it
  4. the firm’s risk appetite
  5. regular updates to status and progress
  6. a RAG categorisation, or something similar

​
How guardbook maps to that

The register is at Risk Register in the left-hand navigation. Each row shows Title, Tool, Category, Inherent score, Residual score, Within appetite, Status and Review. Review shows whether the risk is on track, due soon or overdue for review. Select a title to open the full record. A risk can also be linked to a specific tool from your Tool Register, which is worth doing wherever the risk is tool-specific rather than general. For each tool you add, guardbook adds four starter risks with no scores, one for each of the four risks the standard names: Data exposure; Hallucination and inaccuracy; Bias; and Limited information about the system. Score and complete them.
Only Title and Status are required to save a risk.Status defaults to Open, so in practice a title is all you need. Everything else the standard asks for, including the description and the scores, is optional in the form. A register of titles with no scores, no controls and no response will save without complaint and will not satisfy clause 3.3. Treat the rest of the fields as required even though guardbook does not.

​
Keeping it operating

The clause requires the register to be reviewed and updated at least quarterly, by the staff responsible for decisions about the firm’s AI use. That is an ongoing duty, and it is the part firms most often let slip. Quarterly review covers what that involves and how guardbook helps. Use Start quarterly review on the Risk Register to record it.