AI Fluency for Business Leaders in 2026: The Minimum Viable Understanding Every Director and VP Now Needs

You do not need to know how a transformer works. You do need to distinguish these five so vendor demos do not confuse you.

• Classical ML: churn, fraud, forecasting. Cheap, mature, interpretable. Runs on tabular data.
• LLMs (chat / completion): text summarisation, classification, drafting. Costed per 1,000 tokens.
• RAG (retrieval augmented generation): LLM + your private documents. This is where 80 percent of enterprise pilots land in 2026.
• Agents: LLMs that call tools and take multi-step actions. Higher value, higher risk, harder to evaluate.
• Multi-modal: image, audio, video in and out. Cost 5 to 20 times text-only.

In November 2026, GPT-5 costs roughly 0.005 USD per 1,000 input tokens and 0.015 USD per 1,000 output tokens. Claude 4.5 Sonnet is 0.003 / 0.015. Gemini 2.5 Flash is 0.0004 / 0.0012 for volume. A single customer support ticket answered end-to-end runs 0.02 to 0.15 USD depending on model, retrieval calls and reasoning depth. A leader who cannot ballpark cost per transaction will approve pilots that never scale.

A leader who walks into a vendor pitch with these 12 questions filters out 70 percent of hype in one meeting.

• What is the human baseline (accuracy, cost, latency) we are trying to beat?
• Which of the five system classes are we using and why?
• What is the eval set and pass criterion? (No eval, no launch.)
• What is the cost per 1,000 transactions at expected volume?
• What is the fallback when the model is uncertain or offline?
• Where does the data go? (Model provider, retention, region.)
• Is a human in the loop for high-stakes decisions?
• How will bias be measured and monitored post-launch?
• What is the roll-back plan if quality degrades?
• Which regulator, framework or standard applies? (EU AI Act, NIST AI RMF, ISO 42001.)
• Who owns the risk if the model is wrong? (Vendor contract clarity.)
• What is the training and change-management plan for affected staff?

The EU AI Act came into general applicability in August 2026. It classifies systems into unacceptable, high, limited and minimal risk, with fines up to 7 percent of global turnover for prohibited practices. High-risk systems (HR screening, credit, education, critical infrastructure) require conformity assessment, technical documentation, human oversight and post-market monitoring. ISO/IEC 42001:2023 is now the de facto AI management system standard, adopted by 4,200+ organisations globally by Q3 2026. NIST AI RMF is voluntary but expected by US federal contractors.

Leaders in regulated sectors (finance, healthcare, HR tech, edtech, gov) should sponsor an ISO 42001 gap analysis in Q1 and appoint an AI risk owner at director level.

MIT CISR's 2026 study of 218 GenAI programmes found that technical build was 25 percent of effort. Change management, training and workflow redesign was 60 percent. The remaining 15 percent was data prep and evals. Leaders who skip the middle bucket end up with brilliant models nobody uses. Budget a 1:2 ratio: for every dollar spent on the model or vendor, spend two on adoption.

• Week 1 to 2: complete a leadership AI primer (Learners Ink, Wharton or MIT Sloan Executive).
• Week 3 to 4: shadow one live AI project as a stakeholder, ask the 12 questions.
• Month 2: sponsor one narrow pilot with a real eval set and cost ceiling.
• Month 3: publish an internal AI usage policy (data, tools, IP, disclosure).
• Ongoing: read one primary source per month (arxiv preprint, Anthropic system card, EU AI Act update).