Imagine a company with an AI policy—employees complete training. Important outputs supposedly receive human review. The board sees an approval dashboard and assumes it’s managing the risk. Now ask a different question: What evidence shows that the review changed anything?
That is the central compliance challenge in ECI Compliance Week’s AI in the Audit Room. The report was prepared in conjunction with Workiva. The collection translates its AI Governance Framework into an audit program, connecting 14 controls across three tiers to specific tests and supporting evidence. AI-in-the-Audit-Room For Chief Compliance Officers (CCOs), the lesson is straightforward: AI governance must demonstrate how decisions are controlled, challenged, and improved. A policy describes expectations. An approval records an action. Neither, on its own, proves that a control works.
The Oversight Gap Is an Operational Problem
45.2 percent of the respondents reported training or awareness programs, while 33.8 percent reported usage monitoring or logging by IT or security. Separately, the collection reports that approximately one in four executives in Workiva’s 2026 Midyear Executive Benchmark Survey said internal audits had discovered AI errors after outputs reached external stakeholders or board members. Only 11 percent believed their data quality was sufficient for AI use.
These findings should prompt a practical examination of where assurance occurs. Finding an error after distribution is not the same as preventing unsupported information from reaching a decision-maker. Workiva’s separately labeled partner contribution proposes an “assurance architecture” built around governance, traceability, and defensibility. The useful compliance question is whether management can explain what data informed an output, which model produced it, and what an employee authorized.
For an AI-assisted disclosure, that means following the information back to its source rather than accepting a plausible explanation. Ask whether reviewers can retrieve the supporting material and distinguish verified inputs from unsupported assertions. Traceability should make challenges possible before release, not merely explain a failure afterward.
Tier One: Establish Foundations That Can Be Tested
The framework’s first tier contains five controls: an accountable executive, an AI policy, an AI use inventory, employee AI training, and an annual program review. The accompanying audit tests make those familiar concepts concrete. The accountable executive needs authority and budget. The inventory must cover licensed products, AI embedded in existing software, and tools employees adopted independently. Training must reach actual users and match the policy in force. The annual review must produce findings that lead to action.
For the CCO, start with the inventory. Reconcile it against procurement, identity, and expense records, as the framework recommends. Then examine whether each material use has an owner and an appropriate data classification. Do not stop at asking whether training occurred. Ask which users received it and whether the content addressed their activities.
The same discipline applies to program review. A dated report is not the endpoint. Look for assigned remediation, closure dates, and evidence that the organization changed something because of what it learned.
Tier Two: Make Governance Part of Daily Operations
The second tier covers third-party AI due diligence, an AI speak-up channel, incident response, and employee AI use attestation. The third-party control is particularly important. The framework calls for reassessment when AI functionality arrives through a software update, not merely when the original contract is signed. It also calls for identifying the underlying model provider chain and determining whether customer data is used for training.
For compliance teams, that means connecting procurement and vendor oversight to technology change management. The original supplier approval should not be treated as the final assessment of every capability subsequently introduced. The speak-up and incident controls should work together. An employee who identifies troubling output needs a route to someone with AI governance expertise. Where appropriate, a substantive response should lead to investigation, containment, root-cause analysis, and control changes. Employee attestations should also update the inventory rather than disappear into a training database.
My practical recommendation is to test that entire process with a realistic scenario: an approved tool produces an unsupported claim intended for an external report. Determine who receives the concern, who can stop distribution, and what evidence documents the resolution.
Tier Three: Govern the Use Case and the Authority
The third tier addresses pre-deployment approval, human oversight and output monitoring, AI agent access and security, bias and fairness considerations, and external transparency. A critical distinction is that approving a tool is not the same as approving every use of that tool. The framework calls for separate use-case approval records, including cross-functional review for uses affecting financial reporting, customers, or employees.
For agents, the control table calls for unique identities, least-privilege access, defined limits on autonomous action, reconstructable action logs, and a tested way to stop the system. It also calls for testing decisions that affect people and can produce disparate outcomes and for substantiating public statements about AI. Translate these expectations into demonstrations. Ask the system owner to show an agent’s permissions, reconstruct an action, and produce the results of a shutdown test.
Monitoring should also look beyond internal error rates. The framework treats vendor model changes, public incidents, and regulatory shifts as external signals that require scheduled attention. Ask for the record of those signals and the actions taken, not simply the latest system performance report. For transparency, select a public claim and request its evidence file. The test is not whether the statement sounds reasonable. It is whether the organization can support it.
The Signature Problem: Test Judgment, Not Clicks
The collection’s most useful challenge concerns the human review step. Its “signature problem” describes a reviewer who sees a polished answer but lacks the underlying evidence, discarded alternatives, or meaningful ability to challenge the result. Approval becomes easy; disagreement becomes burdensome. The collection proposes three questions: Can the reviewer change the available options? Can the reviewer see enough evidence to disagree? Does disagreement produce a change?
The report identifies four metrics: approval rate, edit distance on accepted output, time spent per review, and override consequence. Together, these examine whether reviews vary meaningfully, whether edits are captured and evaluated, whether reviewers have time to examine evidence, and whether overrides lead to action.
For CCOs, the fourth metric offers an immediate starting point. The collection recommends examining the last 20 overrides in an AI-assisted process affecting financial statements, customers, or regulators and asking what changed because of each. That exercise gets beyond a completion dashboard. It connects a reviewer’s judgment to remediation, revised instructions, or an adjusted control.
The collection also recognizes the limits of universal human review at scale. It describes moving human judgment into system boundaries, evaluation criteria, exception handling, sampling, and monitoring. It preserves individual review for specified consequential situations, including irreversible outcomes and cases outside anticipated policy boundaries. The practical lesson is not to remove people from oversight. Instead, place human judgment where it can actually influence outcomes, and give reviewers the evidence, authority, expertise, and time to exercise it.
Do Not Overlook the Inference Gap
Shruti Mukherjee’s contribution extends the discussion to AI’s environmental footprint. She distinguishes model training from inference, the recurring activity of running it, and frames the issue as a governance exposure involving ownership, measurement, third-party information, and substantiated claims. Her proposed audit questions are practical: Who owns the energy and water footprint of AI use? Can the organization measure ongoing consumption? What do supplier contracts disclose? Can the company support its environmental statements?
For the CCO, the broader application is to look beyond familiar risk categories. A benefit claimed in a presentation should not become an external assurance without supporting evidence. Where information is unavailable, document the limitation and identify who is responsible for addressing it.
Practical Takeaways: Bring Evidence to the Board
Start with one material AI-assisted process rather than rewriting an enterprise-wide policy. Trace it from the approved use case through source data, output review, escalation, and final distribution. Ask control owners to produce evidence at each stage.
Next, examine the review process itself. Use the collection’s four metrics to identify where approval activity may be substituting for judgment. Follow overrides through to their consequences and assign responsibility for unresolved weaknesses.
Finally, change the board conversation. Report not only how many tools are approved or how many employees are trained, but also what testing found, which controls failed, what management changed, and which risks remain unresolved.
The board should ask whether management knows where AI is used, whether reviewers can meaningfully challenge it, and whether the organization can reconstruct and support significant actions and public claims. This is where compliance connects governance directly to business execution.
The CCO should ask an equally direct question: When someone identifies that the AI is wrong, does the organization merely record the concern, or does it learn and change? That is the difference between an AI governance program that documents approvals and one that demonstrates effective oversight.
