Four AI Questions Every CISO Gets, One Without a Dashboard
Threats, Compliance, and Data Have Telemetry Behind Them. The Fourth Question Is About People, and Nothing Logs It
This article is for educational purposes and does not constitute legal advice or certification criteria. Organizations should consult qualified counsel about obligations that apply to them.
Of the four AI questions boards now ask CISOs, three have a dashboard behind them and the fourth has a hope.
The Four Questions
On August 4, 2026, TrustCloud's Sravish Sridhar published an essay, syndicated the same day on Security Boulevard, arguing that the CISO's entire AI problem comes down to four questions the board is already asking. How are CISOs protecting the business from AI-driven security threats? How are they meeting new AI compliance obligations? How are they protecting data and IP as AI adoption grows? How are they proving AI improves outcomes without adding unnecessary security risk or work?
The framing's honest: protection, compliance, data, and value were the CISO's brief before any model shipped. The essay's opening challenge is the one worth keeping: "Could you answer with evidence?" That's the right test, and it's the test the fourth question fails for structural reasons.
Three Have Telemetry
Take the questions in order and ask what evidence system sits behind each one.
AI-driven threats have the security stack. Prompt injection, data leaving for public models, AI-assisted phishing, and anomalous agent behavior all leave traces that SIEM, EDR, and continuous control monitoring are built to catch.
Compliance obligations have the GRC platform. Controls mapped to ISO 42001 and the NIST AI Risk Management Framework, tested on a schedule, with evidence attached to each control. The obligations themselves are a moving target (statutory references and obligation categories are indexed at EveryAILaw), but the evidence system for showing a control operated is mature and auditable.
Data and IP have DLP, CASB, and third-party risk management. Verdantix's Strategic Focus report of August 17, 2026, The Cyber Resilience Imperative For Firms Using AI, by analysts Mahum Khawar and Luis Niño, argues that most firms cannot yet answer basic questions about where sensitive data flows and how concentrated their AI dependency has become. Hard questions, but they're about systems and vendors, answered by mapping the architecture.
Each of the three has gaps, the tool the inventory hasn't caught, the control narrative that didn't survive an audit. But every one of those is a coverage gap over activity. The activity exists to be seen.
| Board question | Evidence system | What it cannot show |
|---|---|---|
| Protecting the business from AI-driven security threats | SIEM, EDR, continuous control monitoring | Whether an analyst trusted a wrong AI triage |
| Meeting new AI compliance obligations | GRC platform mapped to ISO 42001 and NIST AI RMF | Whether the control operated in someone's judgment |
| Protecting data and IP as adoption grows | DLP, CASB, third-party risk management | Whether people notice when the model is wrong about the data |
| Proving AI improves outcomes without adding unnecessary security risk or work | Nothing. Training completion, usage, self-report | Whether people catch what the model gets wrong |
The Fourth Is a People Question
Read the fourth question slowly. "Proving AI improves outcomes" is measurable in hours saved and cycle times; every vendor dashboard reports it. So is "without adding work." "Without adding unnecessary security risk" is the clause that matters, and it unpacks to something no system logs: that the people using the model catch what it gets wrong before it reaches a customer, a regulator, or a filing.
An analyst accepts an AI-generated incident summary that omitted the initial access vector. A compliance lead files a control narrative citing a framework clause that doesn't exist. A data steward approves a classification the model assigned confidently and wrongly. None of those fires an alert. The SIEM sees a normal session, the GRC platform a completed control, the DLP no exfiltration. The risk that got added is a person's uncorrected trust, and it's invisible to every system answering the first three questions.
The NIST AI RMF puts this squarely inside the GOVERN function. GOVERN 2.3 says executive leadership takes responsibility for decisions about AI risk. GOVERN 3.2 calls for policies that define and differentiate roles for people working alongside AI systems and overseeing them. GOVERN 4.1 asks for practices that foster a critical thinking and safety-first mindset in the use of AI. Those are the framework's words for the fourth question, and none of them names a telemetry source. Policies, roles, and mindset get asserted. Whether people actually think critically when the model's confident and wrong has to be observed.
The tools we have measure everything except what matters: behavioral patterns when AI is uncertain, incomplete, or wrong. Dashboards show activity. They do not show reliability.
How CISOs Answer It Today
When the fourth question comes up in a board deck, three proxies fill the slot.
Training completion rates. The AI acceptable-use module was assigned and completed. That proves attendance. Training ≠ Capability. Passing the quiz and catching a fabricated control reference under deadline take different instruments, and most organizations own only one.
Usage counts. Seats provisioned, prompts per week, adoption by department. These prove the model's in use, the precondition for the risk rather than evidence of it. Usage ≠ Quality.
Self-report surveys. People rate their own confidence in verifying AI output. Self-Assessment ≠ Behavior. The people most at risk are the ones whose confidence is least calibrated, and we're asking them to calibrate themselves.
When an AI-assisted control contributes to an audit finding, accountability travels upward to whoever owns the program, usually the CIO or CISO, unless ownership and evidence were defined in advance. That's the real cost of the fourth question: the CISO owns an outcome that depends on judgment across the business, holds no instrument that observes it, and reports proxies an audit committee can see through. When new systems are introduced, execution is audited. Behavior is not.
An Answer the Board Can Audit
The fourth question needs an evidence system of its own, and the evidence is behavioral. Put a person in front of confident, plausible, wrong AI output under time pressure and watch what they do. Do they catch the error, verify the claim that carries the decision, escalate when a source can't be confirmed, and correct rather than defend? That's observable, repeatable, and comparable across a cohort. It's the missing row in the table.
PAICE (People + AI Collaboration Effectiveness) produces that evidence. The assessment seeds errors of the kind that reach incident reports, control narratives, and classifications, and scores how each person behaves when the model gets it wrong. Fluent but missed the seeded error scores lower than terse and caught everything. I think of it as a credit score for AI-enabled work: not omniscience, but far better operational visibility than none. The board gets a behavioral reliability baseline to read next to the other three answers, explainable and re-measurable.
Privacy is architectural. Scores attach to hashed identifiers, no conversation text is retained, and individual results are structurally unavailable to the organization, so the CISO sees the cohort's reliability distribution and never a name. That's governance without surveillance, and the only form of this evidence a regulated organization should accept, because a people measure that could be turned on individuals would create the risk it exists to reduce.
PAICE is scoped to the People vector. Threat telemetry, control mapping, and data controls stay with the systems answering questions one through three. How the whole picture rolls up is an organization-level judgment, and AI Posture is one framework for thinking it through. The public matters and determinations that follow when AI output goes wrong are recorded case by case at AI Incident Law.
The board's going to ask all four questions again next reporting cycle. Three answers will come from dashboards. The fourth can come from evidence, or it can come from hope, and the difference shows the moment someone asks how you know.
Want an answer to the fourth question before the board asks it again? Learn about PAICE for organizations or take an individual assessment to see it firsthand.
Recommended Reading
- AI Collaboration for Cybersecurity Professionals - Verification practice for analysts and incident responders
- Building the Business Case for AI Collaboration Assessment - What to bring to procurement and the audit committee
- AI Governance Metrics Beyond Adoption and Training - Why the standard catalog measures systems and not operators
- Filling The Missing Trust Layer - How the portfolio's components divide the problem
Curious but short on time?
Take the 3-minute PAICE Pulse — a quick confidence check that maps how you see your own AI collaboration posture. No login required.