Clone Solutions
Template

AI Risk Register

A living register of AI risks by use case, scored for likelihood and impact, with controls, owners, autonomy limits and a review cadence.

What this is for

Every Clone workflow carries risk: a wrong answer, a leaked record, an action taken on stale data. The register makes those risks visible, scores them consistently, and ties each one to a control, an owner and a limit on autonomy. It is how InfoSec decides whether a workflow can be promoted, how steering sees the risk position each month, and how auditors see that risks were managed, not just listed.

When to use it

Create the register in the Week 3 governance design session. Add risks for every new use case in the Design week of each department rollout. Review open High and Critical risks at the weekly InfoSec sync, the full register monthly for the steering pack, and every risk before an autonomy promotion.

Who's involved

  • InfoSec lead: owns the register and the scoring method.
  • Business owners: own the risks for their departments' use cases and accept residual risk.
  • Implementation Engineer: proposes technical controls and tests them.
  • GRC Specialist and GRC Clone: facilitate, maintain, and report.

How to fill it in

  1. For each use case, walk through the risk categories below and ask "how could this go wrong here?"
  2. Write each risk as cause → event → consequence. One risk per row.
  3. Score inherent risk (before controls) for likelihood and impact.
  4. List controls. Mark each as Preventive, Detective or Corrective, and as In place or Planned.
  5. Score residual risk (after controls in place).
  6. Set the autonomy cap: the highest level this workflow may reach while this risk remains at its residual score.
  7. Assign an owner and a review date. Risks above appetite need a treatment plan and a target date.
  8. Log incidents and near misses against the relevant risk. Rescore after every incident.

The template

Scoring scales

Likelihood

ScoreLabelGuide
1RareNot expected in a year of operation
2UnlikelyCould happen once a year
3PossibleCould happen once a quarter
4LikelyExpected monthly
5Almost certainExpected weekly or more

Impact

ScoreLabelGuide
1MinimalInternal inconvenience, fixed in minutes
2MinorInternal rework, no external effect
3ModerateLimited customer or financial effect, reversible
4MajorSignificant financial loss, regulatory notification possible, customer harm
5SevereRegulatory action, material loss, serious harm to people or reputation

Rating = Likelihood × Impact

RatingLevelAppetite and action
1–4LowAccept; review quarterly
5–9MediumAccept with controls; review monthly
10–16HighTreatment plan required; review weekly; autonomy capped at Execute with approval
20–25CriticalDo not go live until reduced; steering informed

Risk categories (prompt list)

CategoryPrompt questions
Accuracy and hallucinationCould the Clone state something false with confidence? Would anyone catch it?
Data exposureCould sensitive data reach the wrong person, channel or system?
Unauthorized actionCould the Clone act beyond its approved autonomy level or scope?
Prompt injection and manipulationCould content from an email, document or web page change what the Clone does?
Stale or wrong knowledgeCould the Clone rely on an outdated policy, price list or procedure?
Bias and fairnessCould outputs treat people differently on protected characteristics?
Over-relianceWill approvers rubber-stamp because the Clone is usually right?
Integration failureWhat happens if a connected system is down, slow or returns bad data?
Regulatory and contractualDoes the use case touch rules on disclosure, consent, record keeping or automated decisions?
Third-party and vendorDoes the workflow depend on another AI vendor or tool with its own risks?
Adoption and changeCould users bypass the Clone, or misuse it, because of how it was introduced?
Cost and capacityCould usage spike unexpectedly in a way nobody is watching?

Register

IDUse case / CloneCategoryRisk (cause → event → consequence)LIInherentControls (P/D/C, in place / planned)LIResidualAutonomy capOwnerTreatment plan and dateStatusLast reviewed
R-01Open / Treating / Accepted / Closed
R-12AP Matching CloneAccuracyInvoice with ambiguous PO reference → Clone proposes the wrong PO → overpayment or duplicate payment3412 HighP: match only on PO + vendor + amount within tolerance (in place). D: approver sees match evidence side by side (in place). D: weekly duplicate-payment report (planned). C: ERP payment hold on variance (in place)236 MediumExecute within limits for variance ≤ 2%; Execute with approval otherwiseControllerDuplicate report live by Week 9Treatingwk 7
R-13AP Matching ClonePrompt injectionSupplier email contains instructions → Clone changes vendor bank details → payment fraud2510 HighP: Clone has no write scope on vendor master (in place). P: bank detail changes are prohibited in the charter (in place). D: injection test cases in test pack (in place)155 MediumNo change to vendor master at any levelInfoSec leadn/aAcceptedwk 7

Incident and near-miss log

DateRisk IDClone / workflowWhat happenedSeverityCaught by (control)FixAutonomy changeRescored?
None / Demoted to ___

Monthly summary for steering

MeasureThis monthLast month
Open risks: Critical / High / Medium / Low
New risks added
Risks closed or reduced
Incidents above agreed severity
Near misses logged
Workflows demoted
Treatment plans overdue

Review checklist (before any autonomy promotion)

  • All risks for the workflow reviewed within the last 30 days
  • No open Critical risk; any High risk has controls in place, not planned
  • Proposed level does not exceed any risk's autonomy cap
  • Incidents since last review logged and rescored
  • Business owner has accepted residual risk in writing

Good looks like

  • Risks are specific to a use case. "AI might hallucinate" is a category. "Clone proposes the wrong PO when the reference is ambiguous" is a risk.
  • Controls are tested, and each test is linked in the Test Scenario Pack.
  • Every risk caps autonomy. The register drives promotion decisions, not just reporting.
  • Near misses are logged. They are the cheapest lessons you will get.
  • Residual risk is accepted by a named business owner, not by the delivery team.
  • Common mistake: building the register once for the audit and never touching it again. Review it weekly while workflows are changing.

How Clone helps

The GRC Clone drafts risks for each new use case from the Clone Design Spec and the prompt list, reminds owners before review dates, logs incidents from the audit log review, and produces the monthly summary for the steering pack. The GRC Specialist chairs the risk discussion at the weekly InfoSec sync. In the portal, the register is the AI Risk Register deliverable under Security & GRC; High and Critical risks surface on the steering dashboard, and promotion requests show the linked risks and their caps.

Sign-off

RoleNameDateDecision
InfoSec leadRegister approved
Business owner (per use case)Residual risk accepted
Executive sponsor (any Critical risk)Accepted / Not accepted