Security, privacy, reliability and accountability as leadership responsibilities
Updated
BeginnerWhat you'll learn
~15 min- Ask the four leadership questions (security, privacy, reliability, accountability) of any capability before routine use
- Decide the depth of review by the consequence of a failure instead of applying one process to everything
- Name the record to request from the accountable reviewer, and confirm which policy sources apply to your organization
You leave this lesson with: a consequence level, a named accountable reviewer and a dated decision for each capability.
Four questions that belong to the leader
Security, privacy, reliability and accountability sound like technical subjects. Each also contains a question only a leader can answer, because the answer depends on the organization’s purpose and its obligations.
Security. What can this capability reach, and what is the worst thing it could do with that reach? The leader decides what access the organization is willing to grant.
Privacy. What information about people goes in, where does it go, and who may see the result? The leader decides what the organization promised those people.
Reliability. What happens to the work when the capability is wrong, slow or unavailable? The leader decides what failure the mission or the customers can tolerate.
Accountability. When it produces a harmful result, whose name is on the decision to use it? The leader decides who signs, and makes sure that person has the authority to refuse.
Security, privacy and compliance reviewers exist to protect the organization from exactly these harms. The leader’s job is to see that each review is sized to the risk and has a named owner and a date.
Match the review to the consequence
Two uses in one organization can need very different review. A draft that a person checks before it is used carries a different consequence from an output that decides who loses a contract, a benefit or a place on a list. For each use, ask what happens if the checker misses an error. That consequence sets the use’s consequence level, which helps determine how much review is needed, and the accountable reviewer records the level and its reasons.
What to ask the accountable reviewer for
This course does not teach the method for setting a consequence level. It teaches what to ask for.
The accountable reviewer is whoever your policy names to approve uses at that consequence level, such as the privacy or security officer or the chair of the research ethics board (IRB); if no one is named, naming one is the leader’s first decision. A disagreement about the level goes to whoever already settles risk disputes in your organization, or, where no one does, to the executive who owns the program.
Ask the accountable reviewer how serious a failure could be and why, including whether the output is the main basis for a decision that affects people’s rights or safety. Before routine use, get a written approval naming the approved use, who approved it and when it must be reviewed again. AI Assurance (Lessons 1.3 and 1.4, a separate training in this series) shows how engineering and assurance teams set that level of seriousness and assemble that approval record.
🔍Federal programs only, as of 2026-11-07click to expand
None of these instruments authorizes your proposed use. Each sets a governance floor, an acquisition requirement or a policy direction; your agency’s own authorization, privacy, records, security and legal review still decides each use. For commercial and academic leaders, your organization’s legal, procurement, research ethics board (IRB) and security reviews play this role.
Secure Software Delivery for Federal Environments (a separate training in this series) covers the authorization process for software; its one-page leadership brief is the place to start.
OMB M-25-21 (agency AI governance). In force. The agency head designates a Chief AI Officer and sets the generative AI use policy; an annual use-case inventory is required (defense and intelligence excepted). The program leader confirms the use follows the agency policy, registers it in the inventory, and checks whether the output is a principal basis for a decision that affects rights or safety (the high-impact test, which triggers pre-deployment testing, an impact assessment and ongoing monitoring, or a documented waiver). Does not authorize: a use outside agency policy or without the high-impact practices.
OMB M-25-22 and OMB M-26-04 (AI and language-model procurement). Both are in force as of this note. The program leader works with the contracting officer to confirm they apply. M-25-22 covers federal contracts solicited on or after September 30, 2025, including options exercised on them later. It requires contract terms that stop the vendor from training publicly or commercially available AI on your agency’s non-public inputs and outputs, and terms that let you move your data and work to another vendor later. Everyday commercial products that include only minor AI features are exempt. M-26-04 requires contract terms carrying its Unbiased AI Principles and a way for users to report problems; national security systems are excluded, and free open-source licensed models require due diligence rather than the procurement requirements. M-26-04 ceases to have effect on December 11, 2027, unless extended. Does not authorize: accepting default vendor terms.
Agency authorization, FedRAMP, OMB M-26-05 and M-26-10. FedRAMP covers a cloud service’s security at an impact level, not whether a workflow is approved; its 2025 AI prioritization closed in April 2026. Under M-26-05 the agency sets its own software assurance policy, so AI-written code is not exempt from it; under M-26-10 the CIO approves IT contracts at CFO Act agencies. The authorizing official and the CIO act; the program leader supplies the use and the data description. Does not authorize: an authorized tool on unapproved data.
Voluntary frameworks: NIST AI RMF 1.0 and AI 600-1 (under revision); CISA’s May 2026 agentic AI guidance. The program team may use them as a shared vocabulary for risk review. CISA advises limiting agent access, avoiding broad permissions on sensitive systems, and starting with low-risk tasks. Does not authorize: treating guidance as a safe harbor.
Direction: EO 14179 (January 2025) and the July 2025 AI Action Plan. EO 14179 sets the policy of removing barriers to AI adoption and revoked EO 14110; the Action Plan is not binding. The agency head and the program leader act within it. Does not authorize: any specific deployment, or a waiver of M-25-21 practices.
Terminology: EO 14434, Inaugurating the Era of Super Intelligence (September 29, 2026) directs agencies to write “Super Intelligence” or “SI” for “AI” in non-statutory documents; it changes no duty above.
Vendor availability. Availability on federal contract vehicles can change by directive, litigation or supply-chain determination. The program leader confirms current status with the contracting officer before buying or deploying any vendor’s tool. Does not authorize: a purchase or deployment that assumes a vendor’s status is unchanged.
A mid-sized manufacturer wants to deploy an AI assistant that drafts maintenance notes, and a second tool that ranks suppliers for removal from the approved list. The security team proposes one full review process for both, which would take the same calendar time for each.
- Apply the full review to both tools using the existing review checklist, since consistency is safer.
- Start both tools now, and review whichever one produces a complaint.
- Ask the accountable reviewer to assess each intended use and record the required review before either proceeds.
Pick one before you open What happens.
ℹWhat happensclick to expand
- Apply the full review to both tools: It is consistent and leaves nothing unreviewed, but it fixes the depth before anyone has assessed the use, and the maintenance notes wait as long as the supplier ranking.
- Start both tools now: It saves review time now, but it leaves the supplier tool with no record of who accepted its risk, and a wrong removal may surface only after a supplier has been harmed.
- Ask the reviewer to assess each use: The reviewer assesses each use, and the leader receives a named, dated decision for each.
Choice 3 holds up, because each use gets a named, dated decision instead of the assumption behind Choice 2 that checked drafts are low consequence.
What should a leader ask for to size the review to the consequence?
Key takeaway
- LEADERSHIP DECISION
- Set review depth by consequence, and name one accountable reviewer for each capability
- IMMEDIATE ACTION
- Ask each pilot team for the reviewer's view of how serious a failure could be, and a written approval with a review date, before routine use
- SUCCESS MEASURE
- Days from a pilot team's request to its dated approval, and review issues found before routine use and resolved by their due date