Independent thinking. Informed defence.

CISO Times

Intelligence for the people behind the defence.

The CISO Decision Brief

How do we…

How do we assess the risk and impact of an AI use case before we approve it?

5 decision pointsThe pack behind this answer: AI Security Governance & Acceptable Use

Add to board memo

The short answer

Recommendation
Assess every Tier 2 and Tier 3 use on six factors: security, privacy, fairness, safety, operational, and legal and contractual. Rate each before and after controls on your own risk scales, from evidence you have seen, and report counts by band and the highest band, never a score.
Reasoning
Rating AI uses on the same impact and likelihood scales as your other risks lets the result sit beside them and be read the same way. Each factor is judged in the areas named for it, and you take the worse. A control lowers a rating and never raises it, so the band after controls is the one that counts. The highest band sets a floor for the tier: a High band needs at least Tier 2, and a Critical band or any physical safety impact needs Tier 3. A factor outside appetite goes to a higher acceptor.
Considerations
  • Cautious is the pack's default appetite: up to Medium accepted without escalation, High tolerated with a plan. Your leadership team may set another. Critical is never within tolerance.
  • Only the board, or a board risk committee it has delegated to, may accept a risk outside tolerance; the management board in a two-tier structure; executive management where there is no board.
  • The assessment is only as good as the evidence. Look at the settings, the contract and the records, and do not rely on what a supplier or colleague says. A second reader is the best defence against optimism.
  • The assessment routes a use to the right tier. It does not classify a system under the EU AI Act, which counsel confirms.

The decision, step by step

Select any step for the reasoning behind it and where the pack covers it.

  • Start
  • Step
  • Decision
  • Outcome
StartA proposed AI use has been through the intake screenDecisionWhat did the screen set?OutcomeTier 1: no assessment, the screen is enoughOutcomeStop: do not assessStepCollect the evidence before anyone rates anythingDecisionDoes personal data go in or come out?StepRecord the Data Protection Officer's advice on a DPIADecisionCould the output or action physically harm a person?StepRate safety impact 4 (Severe), make it Tier 3 and tell LegalStepRate each of the six factors before controlsStepName the controls and who checks each, then rate againStepRead the result as counts: factors by band, and the highest bandDecisionWhere do the factors sit against our appetite for AI-use risk?StepTake the factor to the higher acceptor named for its bandOutcomeOutside tolerance: reduce the risk, or declineDecisionIs the highest band Critical, or does the safety rule apply?StepTake the use to the approvers of the higher tierOutcomeFile the assessment and send the use for decision

Keyboard: Tab to a step, arrow keys move along the flow, Enter pins its explanation, + and − zoom, 0 fits the diagram, Esc closes. The text version below has the same content.

Where a step quotes a number, it is the default in the CISO Times pack. Your own standard may set a different one.

Read as text

The same flow written out: each step, its reasoning and where the pack covers it. 18 steps.

  1. Start

    A proposed AI use has been through the intake screen

    The assessment follows the screen. Keep the twelve intake answers, because the assessment starts from them. The same scales you use for other risks rate the AI use, so the result sits beside your other risks and is read the same way.

    In the pack

  2. Decision

    What did the screen set?

    Tier 2 and Tier 3 uses are assessed. A Tier 1 use needs only the screen. A Stop has no route to approval under the procedure, so nothing is assessed. A new use built on a supplier's model API is Tier 3 until Legal records the role, so it is assessed. When the deployer role is recorded the tier is set again and the screen is repeated; when the provider role or another role is recorded the use stays Tier 3. A use already running is not raised by the signal. Judgement: if the tier is not yet set, finish the screen and start the evidence from the intake answers meanwhile.

    In the pack

  3. Outcome

    Tier 1: no assessment, the screen is enough

    A Tier 1 use has no assessment, only the screen and the baseline conditions. Review it at the tier's interval or on a trigger. If a review finds a higher tier, the use is a new decision at that tier and is assessed then.

    In the pack

    See also

  4. Outcome

    Stop: do not assess

    A prohibited practice or special category personal data stops the use at the screen. Assessing it would only suggest a route to approval that does not exist. Legal or the Data Protection Officer is told, and a changed use is a new request.

    In the pack

    See also

  5. Step

    Collect the evidence before anyone rates anything

    Look at the tool's settings, the contract, the supplier's decision under the third-party pack and the sign-in and logging records. Do not rely on what a supplier or colleague says. The assessment is only as good as the evidence behind each rating. The AI use owner answers for the assessment and the Head of Information Security does the work (RS-04).

    In the pack

  6. Decision

    Does personal data go in or come out?

    Privacy is one of the six factors. Where personal data is involved, the Data Protection Officer's advice is recorded. The officer and the use owner decide together whether a data protection impact assessment is needed. If you do not know whether personal data is involved, take the first branch.

    In the pack

  7. Step

    Record the Data Protection Officer's advice on a DPIA

    A data protection impact assessment is considered where the processing is likely to result in a high risk to people, and Tier 3 needs one where GDPR Article 35 applies. Systematic and extensive evaluation of people by automated processing is a case that needs one. Record the advice and the decision with the file.

    In the pack

  8. Decision

    Could the output or action physically harm a person?

    Think of a system whose output reaches a physical process, such as machinery, vehicles or care. If you cannot rule out physical harm to a person, take the first branch.

    In the pack

  9. Step

    Rate safety impact 4 (Severe), make it Tier 3 and tell Legal

    The safety rule applies whatever the other areas say. The use is Tier 3 and Legal is told.

    In the pack

  10. Step

    Rate each of the six factors before controls

    The factors are security, privacy, fairness, safety, operational, and legal and contractual. Give each an impact and a likelihood from 1 to 4 on your own risk scales, judged as if nothing you plan were in place. Each factor is judged in the areas named for it, and you take the worse. Judgement: where the evidence is missing, take the less favourable value and say so.

    In the pack

  11. Step

    Name the controls and who checks each, then rate again

    A control lowers a rating and never raises it. Name a person who checks each control, as the worksheet asks. Judgement: a control nobody checks should not lower a rating. A second person who reads the factors and the controls is the best defence against optimism. The band after controls is the one that counts.

    In the pack

  12. Step

    Read the result as counts: factors by band, and the highest band

    The result is the number of factors in each band after controls, such as 2 Low, 3 Medium, 1 High and 0 Critical, and the highest band. The product of impact and likelihood only finds the band for one factor. Never add the factors together or turn the counts into a score, a percentage or a traffic light.

    In the pack

  13. Decision

    Where do the factors sit against our appetite for AI-use risk?

    Cautious is the pack's default: a band up to Medium is accepted without escalation and High is tolerated with a plan. Averse accepts up to Low and tolerates Medium. Open accepts up to High. Critical is never within tolerance. Use the highest band of any factor, and take the more cautious branch if it is unclear.

    In the pack

  14. Step

    Take the factor to the higher acceptor named for its band

    Outside appetite but within tolerance needs more than the band alone. A Medium factor needs the risk owner and the Head of Information Security, and executive management as well. A High factor needs the accountable executive and the Head of Information Security, and executive management as well. Record who accepted it and the review date.

    In the pack

  15. Outcome

    Outside tolerance: reduce the risk, or decline

    Only the board, or a board risk committee it has delegated this to, may accept a risk outside tolerance. In a two-tier structure that is the management board. Where there is no board, such as in an owner-managed company, it is executive management. The band's approvers recommend. Judgement: the useful step is usually more controls or a narrower use, then a new rating.

    In the pack

    See also

  16. Decision

    Is the highest band Critical, or does the safety rule apply?

    The assessment sets a floor for the tier. A Critical band after controls, or the safety rule, needs Tier 3. A High band needs at least Tier 2. Medium and Low bands set no floor. The tier for the decision is the higher of the tier from the intake answers and this floor.

    In the pack

  17. Step

    Take the use to the approvers of the higher tier

    If the floor is above the tier the screen gave, the higher tier applies and the screen is repeated. Tier 2 goes to the AI governance lead and the Head of Information Security. The AI use owner is consulted and signs the conditions it accepts, but nobody approves a use they own, so the executive decision-maker takes the place of an approver who owns it (IS-4). Tier 3 goes to the executive decision-maker the procedure names, with Legal's confirmed worksheet and HR where staff are affected. Otherwise the screened tier stands.

    In the pack

  18. Outcome

    File the assessment and send the use for decision

    Keep the workbook with the intake file, and put the counts, the highest band and any condition you recommend on the decision record. The approvers decide with the screen, the assessment, the supplier's decision and the advice given. Further conditions are each written so that someone else can check them, and the use owner signs the ones it accepts.

    In the pack

    See also

Practical checklist

  1. Confirm the screen set Tier 2 or Tier 3; a Tier 1 use needs no assessment and a Stop is not assessed.
  2. Collect the tool's settings, the contract, the supplier's decision and the sign-in and logging records first.
  3. Record the Data Protection Officer's advice, including on an impact assessment, where personal data is involved.
  4. Rate safety impact 4 (Severe), make the use Tier 3 and tell Legal if it could physically harm a person.
  5. Rate the six factors before controls, then name each control and who checks it, and rate again.
  6. Read the result as counts of factors by band after controls, and the highest band.
  7. Take any factor outside appetite to the acceptor its band needs; reduce or decline one outside tolerance.
  8. Apply the tier floor, and send the use to the approvers of the higher tier.

Evidence to have ready

  • The six factors with impact and likelihood before and after controls, and the evidence behind each rating.
  • Each control relied on and the person who checks it.
  • The counts by band, the highest band, the tier floor and the tier for the decision.
  • Who accepted any factor outside appetite, with the review date, and the Data Protection Officer's advice where recorded.

Assumptions

  • The intake screen is done and the use is Tier 2 or Tier 3. A Tier 1 use needs only the screen.
  • You have a risk assessment method with impact and likelihood scales, and a stated appetite for AI-use risk. The pack uses Cautious as its default, and your leadership team may set another.
  • You can see the evidence: the tool's settings, the contract and the supplier's assessment. The ratings are judgement on a four-point scale.
  • The assessment routes a use to the right tier. It does not classify a system under the EU AI Act, which counsel confirms.

Not covered here