# AI Workshop for Executives — Edited Workshop Transcript *27 October 2026 · Facilitated by Mustapha Benkalfate, Revenu-ia* This is a **structured reconstruction**, not a word-for-word transcript. It follows the workshop’s sequence, preserves the main exchanges and includes selected verbatim excerpts. ## Welcome and purpose **Mustapha:** This was the first session in a proposed series of small-group executive AI workshops. The aim was open discussion rather than a conventional course: participants could interrupt, test ideas and relate them to their own work. > “I am more like I'm in between a trainer and the facilitator today, okay?” Mustapha introduced Revenu-ia’s focus on AI literacy and fluency, its work across 17 countries, and his research into how business leaders make AI-transformation decisions. He said a recurring problem was that members of the same board or leadership team often did not share the same understanding of AI. ## Introductions: different starting points Mustapha invited participants to describe their roles and recent AI use. - **Pau.**, involved with Da. and several other organizations, described herself as a beginner. She uses ChatGPT to prepare presentations and panel-session points. - **Si.**, recently retired from a full-time role, had used Copilot extensively, tried Claude to build a small app, and used Gemini through Google Search. - **CC**, visiting a humanoid factory in Shenzhen, raised AI ethics, governance and human control as his principal concerns. - **Se.**, working in early-career development from Switzerland, uses Claude, ChatGPT and Lovable to prototype ideas. - **To.**, based in Kuala Lumpur, uses Meta AI in WhatsApp for questions and wants to learn to automate tasks. ### CC’s question: who remains in control? **CC:** As agents receive more authority, how can organizations ensure a human remains the decision-maker? He raised examples involving autonomous vehicles and financial transactions. **Mustapha:** The difficult question is not simply whether there is a human in the loop: > “My opinion is, is— Which human?” He argued that effective oversight requires people who understand an agent’s capabilities, how it was built and the objectives it has been given. In his view, unrealistic or contradictory objectives can produce behavior people did not intend. He also stressed that sometimes the sound decision is **not** to use AI. These were Mustapha’s assessments, not conclusions established by the group. ## What participants had already achieved with AI **CC: contract review.** CC said he asks counterparties to run draft agreements through AI *before* sending them, to assess whether the proposed terms are fair and reasonable. He felt this could surface inequitable terms without making the objection come directly from his side. **Si.: financial-statement review.** In a board-training simulation, Si. used Copilot to make a first pass over a 120-page financial statement. Si. specified the audit-committee role and asked it to examine disclosures, numerical inconsistencies and credit-related provisions. It was particularly good at finding inconsistencies—but Si. found further issues on a subsequent read. **Mustapha:** Si.’s example illustrated why expertise still matters. The AI did useful initial work; Si.’s own judgment identified what it missed. Mustapha connected this to a broader concern: less-experienced employees may not yet have the knowledge needed to recognize omissions. ## From a single prompt to a maintained agent Mustapha presented three levels of AI capability: 1. **Practice:** use tools, understand how they work and use them safely. 2. **Design:** build processes and agents, maintain them, and recognize when AI is unnecessary. 3. **Orchestrate:** design end-to-end fleets of agents. His practical distinction was: prompt for a one-off task; consider a reusable skill for a recurring personal task; consider an agent when the capability needs to be shared across a group. He warned against treating the *number* of agents created as evidence of successful transformation. Every additional agent can bring maintenance and security obligations. ### The three foundations of an agent Mustapha organized agent creation around **data, instructions and security**. He said teams often spend time refining instructions when the underlying problem is conflicting, outdated or poorly labeled data. His advice was to begin with a small amount of clean, relevant material and clear ownership, rather than grant access to every available folder. **Si.:** Could AI clean the data? **Mustapha:** Yes, but it is not always the optimal approach. People still need to recognize when the result requires their own review or intervention. He illustrated simple “file hygiene” with explicit document names that identify a policy and its date or version. > “Showing constantly beats explaining.” For Mustapha, an example of a good report, presentation or email can convey the desired output more effectively than a long description. ## Live demonstration: a reusable design and voice At Pau.’s suggestion, Mustapha used **Da.** as the example for a design-system demonstration. He took several screenshots from its website and prompted an AI tool to infer reusable visual guidance for AI-generated presentations. He requested a plain-text output, explaining that he wanted a reusable reference while being mindful of token use. He then described a parallel **communication or voice system**: provide examples of a person’s past emails or speech so an AI-assisted follow-up better reflects that person’s style. This responded to feedback he had heard from employees that AI-generated messages from a manager felt dry and did not sound like her. The broader question, he said, was how people value work when a polished result no longer reliably signals the effort behind it. He offered this as a possible reason AI-generated material may be easier to abandon, while noting the challenge of demonstrating business value from AI activity. ## Participant exercise: the BASE method Participants opened a shared prompt and tried **BASE©**, Mustapha’s agent-creation framework: **Brief, Anchor, Specify, Evaluate**. Rather than immediately writing agent instructions, the method first asks questions about the task, the examples and data it needs, the intended output, and who will own it. **Se.** adapted the exercise to a prospective agent that could check business plans against grant or loan eligibility criteria. In reviewing her AI conversation, Mustapha highlighted the value of first asking whether a reusable agent was justified at all. The questions on screen included who would use it, what one session should produce, and who would own the agent and its outputs. **Mustapha:** A request such as “I’m sick of this task; automate it” leaves essential context unstated. Unlike a new colleague, a generative-AI tool may proceed without saying the brief is too vague. Taking time to answer basic questions at the start may prevent rework, wasted usage and unsafe autonomous behavior later. ### Why begin with only a few use cases? Mustapha asked why a manager and employee should identify two or three tasks where AI would be *obviously* useful, instead of building many agents. **Se.** suggested an early success could encourage further investment. **Si.** suggested automating individual components before joining them into a larger process. Mustapha agreed Si.’s approach was valuable, then clarified the psychological point he was pursuing: people often experience an initial “wow,” stop using what they built, and later restart the same task. An unmistakably useful application is more likely to be maintained. ## After the break: internal and external “brains” Mustapha proposed a recurring organizational use case: compare an **internal brain**—well-structured company information—with an **external brain** of selected outside sources. An agent could then help expose differences, gaps and potential improvements, provided the underlying data and sources were managed. For AI governance, he gave examples: | Internal information | External information | | --- | --- | | Ambassador reports, tool spending, performance reports, incident logs, roles and owners | Vendor changes, security developments, regulation, competitors and market news | Mustapha said this comparison could support more grounded governance reporting and board discussions. He also introduced **PROBE©**, a method that asks the AI to question the user before executing a task—clarifying the real objective, relevant nuances and expected format. For outside information, he emphasized named sources, recency, links and outputs a person can audit. ## Risks: security, overreliance and abandoned work Mustapha’s risk discussion covered several categories: - **Confidentiality and access:** He recounted a biotech example in which, according to his account, employees were told an enterprise AI service was available but were not given adequate connection instructions; they used personal accounts with confidential information. - **Shadow AI:** He described a press-group example involving journalists’ use of personal ChatGPT accounts. - **Unmaintained agents:** He warned that agents can remain in use after their creators move on or their underlying services change. - **Prompt injection:** Instructions planted in material an agent reads may redirect its behavior. - **“Prompt forcing”:** Mustapha used this term for a more persistent form of external control through a compromised browser or integration. These were cases and risk descriptions **presented by Mustapha** during the workshop; the session did not independently verify their figures or technical details. He also raised a human risk: overreliance may weaken the very expertise needed to check AI outputs. He referred to research and examples involving cognitive activity and professional deskilling, and called for **“intentional cognitive friction”**—keeping people actively involved in defining, testing and judging the work. ## A proposed order for organizational rollout Mustapha argued for building capability before deploying autonomous agents widely: 1. Establish shared AI knowledge among leaders. 2. Select and support ambassadors based on capability, security awareness and willingness to help others—not age or managerial preference alone. 3. Establish data ownership, hygiene and reporting. 4. Let trained employees surface practical use cases from their work. 5. Deploy autonomous agents selectively, with maintenance and risk ownership. The intended outcome was not “one agent for the whole company.” It was a smaller number of relevant, supported agents rather than unrestricted creation without evidence of value. ## Audience feedback and closing clarification **Si.:** The session offered a useful overview, but introduced many unfamiliar terms. He wanted more time to understand concepts such as prompt injection, plugins and prompt forcing. **Pau.:** She was still digesting the material and felt that trying the methods would help make them clear. **Mustapha:** He acknowledged the density of the first workshop. His reason for covering the risks was that leaders asking for “automation” also need to understand what autonomy, access and maintenance entail. Pau. then tested the central recommendation: **Pau.:** Was Mustapha suggesting that companies should have one shared agent rather than every employee creating their own? **Mustapha:** No. His objection was to agent creation as a goal in itself. Train people first; let useful ideas emerge from their work; then filter for applications that bring value and can be maintained. > “The logic of creating more and more and more, and that's— this is the way you measure quality of your AI transformation is probably the opposite of what should be done.” Se. connected the discussion to universities, students and SMEs, where she saw a need for basic training, security understanding and workable frameworks. To., drawing on a financial-industry background, emphasized that interest in AI must be matched by compliance, security and the ability to give boards confidence in the approach. ## Follow-through mentioned in the session Mustapha said he would share the workshop materials and methods, the transcript, and information on Revenu-ia’s training offerings. He also offered to share the Da. design-system output with Pau. and discussed connecting Pau. with Se. concerning a Davos event planned for June 2027. These were **stated follow-ups**, not confirmations that they were completed.