top of page

How a board can tell whether AI controls work

Nov 2
4 min read

One question per AI control gives a board evidence instead of assurance.



For boards and CEOs

A board can test whether its AI controls work by asking management five questions, one for each control that governs an AI system making consequential decisions. The answer to accept is evidence: a log, a test result, a number. A slide saying a policy exists doesn't count.

Directors already suspect they aren't getting that evidence. PwC's 2026 Annual Corporate Directors Survey asked 599 US public company directors to rate the AI information their boards receive. On metrics linking AI outcomes, risks, and business performance, 82 percent rated it fair or poor, or said they don't receive it at all.[1] Fewer than half rated even the information on AI strategy as good or excellent.

Each control should produce a visible record

The questions follow The Control Stack, my brief on the five controls any AI system making consequential decisions needs. Each control exists to produce a record. Each question asks management to show it.

Role definition comes first. The board should know what each AI system is allowed to do on its own.

Ask: What may each AI system do without a person signing off?

A good answer is a written list per system: the decisions it makes alone, the ones it only recommends, and the threshold, in dollars or risk, where a person takes over. "The vendor set it up" means nobody on the management team owns the answer.

Review gates put a person between the AI output and the decision, so the board needs to know whether that person catches mistakes.

Ask: Who checks the output, and how do we know they catch errors?

The evidence is a catch rate from a planted-error check. Someone seeds known mistakes into the review queue and counts how many the reviewers find. A reviewer who approves nearly everything may be careful or may be a rubber stamp; the planted errors tell them apart.

Immutable logging lets the company explain a challenged decision after the fact, from records nobody could have edited.

Ask: If a customer or regulator challenges an AI decision, can we reconstruct it?

Have management pick a decision from last quarter at random and walk the board through it: the input, the model version, the output, and who approved it. If that takes a week of engineering work, the log can't support a dispute.

Continuous testing matters because hosted AI models change on the vendor's schedule, and a change can alter outputs the business depends on.

Ask: When the vendor changes the model, how do we find out?

The answer to hear is a fixed set of test cases run on a schedule, with results compared release over release. Ask when the tests last ran and what changed.

Monitoring with alerting covers what drifts between tests. Someone has to hear about it.

Ask: What alerts fire, who receives them, and has one fired?

A good answer names the thresholds, the people who receive each alert, and the last alert that fired, with what was done about it. An alert that has never fired is either well tuned or disconnected, and management should be able to say which.

A good answer is a number, a log, or a test result

"We have a policy" fails all five questions. A policy describes what should happen, and the board needs to know what did happen. Each question has an answer management can bring to the next meeting as an exhibit, the same way the audit committee receives control-testing results for financial reporting.

The audit committee is the natural home for these questions, and its members already want the time. In the same PwC survey, 45 percent of audit committee members said their committee should spend more time on AI-related risks and disclosures.[1]

When the answers come back thin, ask one more question: does anyone skip the control because it slows them down? I tell clients that controls work only when following them is easier than skipping them. A review that adds a day gets skipped under deadline, and an honest log will show it.

Controls cost money, so they belong in the budget

Each control needs people and tooling: reviewer time, log storage, a test suite someone maintains, and an owner for every alert. If the AI budget priced the licenses and left these out, the controls run on goodwill until a deadline makes someone choose. Before you approve the 2027 AI budget shows where those costs hide.

The controls exist because AI systems produce plausible output without knowing the rules the business operates under. Your AI doesn't know the rules explains why that gap doesn't close with a better model.

A board that asks these five questions before it approves next year's AI spending learns two things: whether the controls work today, and what it will cost to keep them working.

If your board wants these answers as evidence instead of assurance, I can help you get them.

Notes

[1] PwC, "2026 Annual Corporate Directors Survey: Governing what's next," Governance Insights Center, September 2026 (released September 23, 2026). Survey of 599 US public company directors; 81 percent of their companies have annual revenues above $1 billion. PwC does not publish fieldwork dates; Bloomberg Law reports the directors were surveyed in the spring. AI information quality: question 11, base 558 to 560 directors; 46 percent rated information on AI strategy and value creation goals good or excellent. Audit committee time: question 16, base 434 to 439 audit committee members. https://www.pwc.com/us/en/services/governance-insights-center/library/assets/pwc-2026-annual-corporate-directors-survey.pdf. Fieldwork timing: Bloomberg Law, "Most Board Directors Say AI Skills Are Lacking, Survey Shows," September 23, 2026. https://news.bloomberglaw.com/esg/most-board-directors-say-ai-skills-are-lacking-survey-shows




© 2026 David Nix. Licensed under CC BY-NC-ND 4.0. Questions: dave@dave-nix.com

Recent Posts

See All

© 2026 David Nix.

bottom of page