Independent thinking. Informed defence.

CISO Times

Intelligence for the people behind the defence.

The CISO Decision Brief

How do we…

How do we decide whether a proposed AI use case can go ahead?

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

Add to board memo

The short answer

Recommendation
Run every proposal through the same screen before it starts: a complete intake form, a stop check, a tier from the highest trigger, then the holds. Let only the tier's named approvers approve, and never the person who owns the use.
Reasoning
One method means two people given the same answers reach the same result. The screen stops a prohibited practice or special category data first, because neither has a route to approval. The tier then sets how much scrutiny the use gets: Tier 1 a screen, Tier 2 an assessment, Tier 3 an assessment and a decision by the executive decision-maker your procedure names. Holds stop a new use until something is put right; for a use already running they become dated conditions, because stopping a service people rely on can do more harm.
Considerations
  • The tiers, decision times and review intervals are the pack's defaults. Your own standard may set different ones.
  • A new use built on a supplier's model API, or one you would brand or supply, is held until Legal records your role (AG-17). A deployer role sets the tier again; a provider or other role keeps it at Tier 3.
  • Tier 3 is decided by executive management by default, or by the board or its equivalent where your governance puts it. A Critical risk outside tolerance goes to the board or its equivalent.
  • The routing is not a legal classification. Counsel confirms whether a use is prohibited, high risk or subject to a transparency duty.

The decision, step by step

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

  • Start
  • Step
  • Decision
  • Outcome
  • Loop back

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. 20 steps.
  1. Start

    Someone proposes an AI use

    A proposal is a new tool, an AI feature switched on in software you already run, a new purpose for an approved tool, or a new kind of data or audience. No such use starts before an intake decision (AG-09). Propose one use at a time: one purpose, one owner, one set of data.

    In the pack

  2. Decision

    Has the owner answered all twelve intake questions?

    The AI use owner answers IQ-01 to IQ-12 in writing, from what the tool does with the data and not from the sales page. The owner must be a manager who answers for the use. If any answer is not known, the file is not complete, so take the second branch.

    In the pack

  3. Outcome

    Return the form and find the missing answers

    Send the request back to the owner. Nothing is screened on a guess, and the working days to decide count from a complete file. Ask for the supplier's documents, a screenshot of the data and sharing settings, and the name of the person who will read the output.

    In the pack

  4. Decision

    Might it match a prohibited practice, or take special category data?

    Check the prohibited-practice list (PP-01 to PP-10) and the answer on special category personal data (health, trade union membership, biometric data and the others of GDPR Article 9). An Unsure on the checklist counts as a match and goes to Legal. Take the first branch if you do not know.

    In the pack

  5. Outcome

    Stop. No route to approval under this procedure

    Tell the requester at once, in plain words, which of the two applies and what would have to change. Tell Legal for a prohibited practice and the Data Protection Officer for special category data. Rewording the form does not turn a prohibited practice into a permitted one. A changed use that keeps the data out is a new request. Neither case takes an exception (AG-20).

    In the pack

  6. Decision

    Annex III area, or a new use we build, brand or supply?

    Annex III areas include employment and the management of workers (HA-4). A new use built on a supplier's model API (for staff or customers), or one the organisation would brand, supply or change the purpose of, is Tier 3 and goes to Legal, who record the role (AG-17). A finished product used as supplied is not building. Once Legal has recorded the deployer role, answer from the other triggers; if Legal recorded the provider role or another role, the use stays Tier 3. A running use is not raised by the signal: answer No and the holds check picks it up as H-4. Otherwise take the first branch if you do not know.

    In the pack

  7. Decision

    Does it decide or shape a decision about a person, or act alone?

    A use that decides or shapes a decision about a person (job, pay, discipline, credit, access to a service) is Tier 3, as is a system that sends output or acts without a person reading it first (AG-06, AG-16). A use that only informs such a decision is Tier 2, and that comes at the next question. No AI output decides such a matter alone; a named person decides after reading the evidence. If you do not know, take the first branch.

    In the pack

  8. Decision

    Confidential or personal data, customers, or material effect?

    Any one of these gives Tier 2: Confidential information could go in; personal data could go in or come out; customers or the public deal with the system or see its output; the use informs a decision about a person; a wrong output or a stoppage would affect a core service or a material sum. None of them, with a person reading every output, is Tier 1. Take the first branch if you do not know. The tier is the highest any trigger gives.

    In the pack

  9. Step

    Tier 3: add Legal's worksheet, HR and any privacy assessment

    Tier 3 needs everything Tier 2 needs, plus the EU AI Act worksheet confirmed by Legal, a data protection impact assessment where GDPR Article 35 applies, and HR where staff are affected. The worksheet records your reasoning against the text. It does not classify the system, and counsel confirms. Record every trigger that applied, not only the highest.

  10. Step

    Assess the six factors, and get the supplier and privacy advice

    Tier 2 and Tier 3 uses are assessed on six factors: security, privacy, fairness, safety, operational, and legal and contractual. Each is rated before and after controls, and the result is a count by band, never a score. The supplier needs a documented decision under the third-party pack before Internal or Confidential data goes in, and for any Tier 2 or 3 use even with Public data (IS-3, H-1). Where personal data is used, the Data Protection Officer's advice is recorded.

    In the pack

  11. Decision

    Does the assessment give a higher tier than the screen did?

    The highest band after controls sets a floor: a High band needs at least Tier 2, and a Critical band or any physical safety impact needs Tier 3. If the floor is above the screened tier, the higher tier applies and the screen is repeated. At Tier 3 already, nothing can be higher, so answer No. Take the first branch if the band is not yet settled and the tier is below 3.

  12. Decision

    Is a hold open: supplier, training, processor terms or role?

    H-1: no supplier decision yet and Internal or Confidential data would go in, or the use is Tier 2 or 3 even with Public data (no supplier, no H-1). H-2: Confidential data and no contract term barring training (a verified setting suffices only for Internal data). H-3: personal data and no processor terms. H-4: a provider signal, such as a build on a supplier's model API, with no role recorded by Legal. If the role recorded is deployer, the record clears it and the tier is set again from the answers; if it is provider or other, the record clears the hold but the use stays Tier 3. A running use carries it as a dated condition. A hold is not a refusal. If unsure, take the first branch.

  13. Decision

    Is the use already running?

    A use found already running does not wait. Stopping a service people rely on can do more harm than carrying on briefly under control. Each hold becomes a condition with a date, and the tool is switched off if the date is missed. A use that is a Stop is never in this group; it stops now.

    In the pack

  14. Outcome

    Hold: the use does not start

    Record the reason, the hold that is open and the date it is due to clear. The use does not start, and the approvers decide again when it clears. Each hold names what clears it, for example H-2 clears when the contract carries the term barring training on your data. Any approver may hold.

    In the pack

  15. Step

    Write each hold as a dated condition with a named person

    The approvers set the date. Name the person who answers for it, and say in the decision that the tool is switched off if the date is missed. A dated condition is part of the approval itself. The tier and the approvers are the same as for a new use.

    In the pack

  16. Decision

    Does the assessment leave a factor outside appetite?

    Outside appetite means the band after controls is above what the organisation accepts without escalation (Cautious is the pack's default). Tier 1 has no assessment, so the answer there is no. If you cannot say, take the first branch.

    In the pack

  17. Outcome

    Reduce the risk, or send it up to who may accept it

    The approvers do not approve a use at this point as it stands. They reduce the risk, send the decision up under the risk acceptance rules (P05), or decline. A High risk outside appetite needs executive management to sign as well as the band's approvers. A Critical risk is outside tolerance and goes to the board or its equivalent, not to executive management as such (RM-04). A departure from a rule is a separate exception, and the prohibitions and the special category data rule take none.

    In the pack

  18. Decision

    What do the tier's approvers decide?

    Tier 1: the AI governance lead. Tier 2: the AI governance lead and the Head of Information Security, two different people, with the Data Protection Officer's advice recorded where personal data is used. Tier 3: the executive decision-maker the procedure names, on the lead's recommendation. The use owner is consulted, signs the conditions it accepts and approves nothing it owns. Where an approver owns the use, the executive decision-maker takes that approver's place; if that person owns a Tier 3 use, the board or its equivalent decides (D-42). Any approver may hold or decline.

    In the pack

  19. Outcome

    Approved: record it, enable access, set the review date

    Enter the use in the inventory the same day, with tier, approver, conditions and review date. Connect single sign-on, apply the settings, and give access only to people who have done the AI essentials module. Put the supplier terms in the contract. Reviews fall at the tier's interval (12 months for Tier 1, 6 for Tiers 2 and 3 by default) and on any trigger. Approve with conditions writes each further condition so someone can check it. Pending a review the AI governance lead may suspend an approved running use, and the tier's approvers decide whether it resumes (AG-10, RS-15).

    In the pack

  20. Outcome

    Declined: give reasons the requester can act on

    Record why, in words the requester can use. They may propose again when the reason no longer applies. Keep the declined file: it shows what was asked and why the answer was no. Any approver may decline.

    In the pack

Practical checklist

  1. Return the form until the use owner, a manager, has answered all twelve intake questions.
  2. Stop the use if it might match a prohibited practice or take special category data, and tell Legal or the Data Protection Officer.
  3. Set the tier from the highest trigger and record every trigger that applied.
  4. Assess Tier 2 and Tier 3 uses on the six factors, with the supplier decision and privacy advice.
  5. Raise the tier and repeat the screen if the assessment's band after controls calls for it.
  6. Hold a new use while any hold is open; write each hold as a dated condition for a running use.
  7. Reduce or escalate any risk outside appetite before anyone approves.
  8. Have the tier's approvers approve, approve with conditions, hold or decline.

Evidence to have ready

  • The complete intake form and the screen result, with every trigger and its rule.
  • The assessment, the supplier decision, the Data Protection Officer's advice and, for Tier 3, Legal's confirmation of the worksheet.
  • The decision record: approvers by name and role, outcome, conditions with any dates and owners, and the review date.
  • The inventory entry made the same day, and the reasons for any hold or decline in plain words.

Assumptions

  • The organisation uses other people's AI and is a deployer in the ordinary sense. Building on a supplier's model API counts as building. If you build, brand or supply a system, Legal decides the role first.
  • You have, or will name, an AI governance lead, and a manager who answers for each use. Judgement: where there is no Head of Information Security, give that role to whoever assesses security.
  • The tiers, working days and review intervals are the pack's defaults. Your own standard may set different ones.
  • The routing is not a legal classification. Counsel confirms whether a use is prohibited, high risk or subject to a transparency duty.

Not covered here