In early 2024, a finance employee in Arup's Hong Kong office joined a video call with the company's UK-based CFO and several colleagues, then made 15 transfers totaling HK$200 million, about US$25 million. Every other person on the call was a deepfake. CNN reported the case after Hong Kong police described it in February 2024, and Arup confirmed it was the victim that May.
Nobody in that chain needed to understand transformer architecture. They needed a payment rule that a convincing face on a screen could not override. That is the test to apply to any AI for CxOs course: does it change what your executives approve, or does it hand them a certificate?
What an "AI for CxOs certification" certifies
Many AI for CxOs courses advertise a certification. Read the fine print and it is usually a certificate of completion: proof of attendance from the provider or a university's executive education office. That has value on a LinkedIn profile. It has none in an audit.
The two things auditors and regulators do look at are different:
- The EU AI Act's literacy duty. Article 4 of Regulation (EU) 2024/1689 has applied since 2 February 2025. It requires providers and deployers to ensure "a sufficient level of AI literacy" for staff who operate or use AI systems, taking into account their knowledge, experience, and the context of use. It names no credential. The question will be what you did for each role, not what anyone holds.
- ISO/IEC 42001. ISO/IEC 42001:2023 is a certifiable standard for an AI management system. It certifies the organization, not an executive.
So the useful output of executive AI training is a record: which executive, which decision, what changed.
The decisions that land on each chair
Generic AI for CxOs training organizes the day around capabilities: what large language models can do, use cases by industry, a maturity model. Security problems arrive organized by who signs. This is the table I would build a C-suite day around:
| Role | Decision | What goes wrong | Reference |
|---|---|---|---|
| CFO | Payment and vendor bank-change approvals | A cloned voice or video of an executive authorizes a transfer | ATT&CK T1656 Impersonation, T1657 Financial Theft |
| CIO or COO | Rollout of an enterprise assistant over mail and file shares | The assistant retrieves everything a user can already open, including files nobody meant to share | ATLAS AML.T0057 LLM Data Leakage |
| General counsel | AI vendor contracts and acceptable use | Staff paste confidential material into a consumer tool | Samsung's 2023 source code upload, reported by Bloomberg |
| CEO or CMO | Customer-facing chatbot | The bot invents a policy and the company is held to it | Moffatt v. Air Canada, 2024 BCCRT 149 |
| CTO or product | Shipping LLM features | Text in a user message or a retrieved document tells the model to do something else | OWASP LLM01:2025 Prompt Injection |
Two rows deserve more time than the rest.
The Air Canada decision is short, and every executive should read it. The airline's website chatbot told a grieving customer they could claim a bereavement fare refund after travel. The written policy said otherwise. Air Canada argued the chatbot was responsible for its own statements. British Columbia's Civil Resolution Tribunal rejected that and ordered the airline to pay CA$650.88 in damages. The amount is small. The finding is not: the company owns what its bot says.
The enterprise assistant row is the one that gets approved fastest. An assistant connected to mail and file shares inherits your permission model, so if a "board compensation 2026" spreadsheet sits in a folder open to the whole company, the assistant will hand it to anyone who asks the right question. Before rollout, have three people from different departments ask the pilot for "salary", "acquisition", "layoffs", and "board minutes", and log what comes back. It takes an hour and tells the CIO more than a vendor briefing.
Make the CFO's control measurable
The control that stops an Arup-style transfer is old: verify bank-detail changes and unusual payment requests by calling back on a number from your own records, never a number or meeting link supplied in the request. Most finance teams have this written down. Fewer check whether anyone follows it.
If your accounts payable system logs vendor bank-change requests, the check is a few lines of pandas:
import pandas as pd # Export of vendor bank-detail changes, last 12 months (column names illustrative) chg = pd.read_csv("vendor_bank_changes.csv", dtype=str) chg["called_master_number"] = chg["callback_number"] == chg["vendor_master_phone"] chg["self_approved"] = chg["requested_by"] == chg["approved_by"] print(f"changes: {len(chg)}") print(f"verified on master-file number: {chg['called_master_number'].mean():.0%}") print(f"no callback recorded: {chg['callback_number'].isna().mean():.0%}") print(f"approved by the requester: {chg['self_approved'].mean():.0%}") # The rows to pull and read by hand print(chg.loc[~chg["called_master_number"], ["change_id", "vendor_id", "requested_via"]]) A missing callback counts as unverified, which is the right default. If the second number is not close to 100 percent, the control exists on paper only, and the deepfake does not need to be very good. This is the kind of exercise that belongs in an executive AI course: a control the executive already owns, checked against their own data, with the AI-enabled threat as the reason it matters now.
Questions to ask before booking an AI for CxOs course
- Does the agenda name decisions by role, or only capabilities and industry use cases?
- Does each attendee leave with one control to verify in their own area, with a date attached?
- Is AI-enabled fraud and impersonation on the agenda, or is security one slide near the end?
- Has the instructor built and attacked these systems? "Could this happen to us" gets a different answer from someone who has.
- What does the certificate certify? If the answer is attendance, price it as attendance.
Where this advice does not fit
If the goal is a credential for a board seat or a job search, a university executive program gives you one, and this post is the wrong lens. A C-suite course also will not secure an AI deployment. Testing the assistant, the chatbot, and the product features is engineering work; executives need to know that work exists, fund it, and read the findings. And if your company has deployed no AI at all, the CFO row still applies. Attackers use generative AI whether you do or not.
We teach A Cyber Executive's Guide for Artificial Intelligence as a one-day course on the security side of these decisions: governance frameworks, AI-enabled threats such as deepfakes, vendor evaluation, and what AI can and cannot do. It was written for security leaders, and we also deliver it as a private executive briefing when the CFO and general counsel need to be in the room with the CISO.