Shadow AI in the Claims Team
Callisto Insurance's claims division employs 400 handlers. A DLP alert shows that over six weeks 61% of them used at least one public generative AI tool at work; one handler pasted 3,000 claim summaries, including medical details, into a free chatbot to "summarise the tone". Callisto has no AI acceptable-use policy. It did procure an enterprise generative AI tool eight months ago, but the information security review has been open for seven months waiting on a data residency answer, so only 20 licences were issued, all to head office. The risk committee meets in five days; its chair has tabled a motion to ban all generative AI use and to issue formal warnings to the 244 staff identified. The CISO supports the ban. The head of claims argues that handlers turn to the tools because the claims template takes 40 minutes to complete and the approved tooling never arrived.
1. Assess the risk posture and recommend what the risk committee should decide. Justify why a ban is or is not the right primary control.
Read the worked answer
A ban treats a symptom. Map the situation properly: the data exposure is a confidentiality and regulatory risk involving special-category data; the 244 users are a policy-deployment gap, not a discipline problem; and the 40-minute template is the root cause of the demand. A ban alone will push use further underground - the tools are free, reachable from personal devices, and the productivity gap is real. You would then reduce detected use rather than actual use, which is the worst possible governance outcome because you lose visibility as well as control.
Recommend a three-part decision. First, containment now: block the highest-risk tool categories at the network and DLP layer, and require that any pasted special-category data is reported so the incident can be assessed against breach notification duties. Do that before debating policy. Second, provisional permission with guardrails: publish a one-page acceptable-use standard this week (no special-category data, no customer identifiers, approved tools only) and fast-track the 20-to-400 licence rollout by granting a time-boxed data residency exception with compensating controls. Third, fix the cause: an improvement action on the claims template with a named owner and a date, or the pressure to use unapproved tools returns within a month.
On the 244 staff: blanket warnings are disproportionate and will suppress reporting. Differentiate - the handler who pasted 3,000 records is a conduct matter; the rest are a policy-deployment failure for which management is accountable. Saying that openly is what makes the policy credible.
2. Define the governance roles and artefacts Callisto needs so this does not recur, and explain how each one changes an actual decision.
Read the worked answer
Roles: an accountable executive AI policy owner; a product owner for each AI use case; a named risk owner per use case; and a decision forum with a fixed cadence, which the risk committee already is. Artefacts: a maintained AI use inventory including shadow use, an AI impact and risk assessment per use case, an exception register with expiry dates, and a monitoring report per deployed tool.
Each must change a decision, otherwise it is paperwork. The inventory converts unknown use into a governed queue - the committee then manages a portfolio rather than reacting to a DLP headline. The impact assessment forces the special-category data question to be answered before deployment instead of after an alert; it is the artefact that would have caught the free-chatbot case at submission. The product owner is the person answerable for monitoring outputs in production, which is precisely the gap between procurement and security here - eight months of ownership by nobody. The risk owner carries residual risk on the register with a review date, so accepted risk expires and is re-examined rather than fossilising.
The exception register is the one that would have changed this month's outcome most directly. The seven-month residency review is a process with no decision deadline; a time-boxed exception with compensating controls and an expiry date forces a decision either way, and an expired exception fails closed rather than silently blocking the tool everyone needs.