How AI Works: An Executive Briefing

---

The Quarter That Changed Everything

In the fall of 2023, Mariana Delgado was 18 months into her role as CEO of Stratton-Byrne, a 2,400-person commercial insurance firm based in Chicago with $1.2 billion in annual premium revenue. By every traditional metric, the company was performing well -- retention rates above industry average, steady 6% year-over-year growth, a loyal distribution network. Her board was satisfied. Her leadership team was executing.

But Mariana had a problem she couldn't articulate in an earnings call.

Every week, her inbox filled with pitches from AI vendors. Her CTO had launched a "proof of concept" with a claims processing tool that nobody outside IT understood. A competing firm -- half Stratton-Byrne's size -- had announced an AI-powered underwriting engine and seen its stock jump 14% in a single day. Two of her best underwriters quit for an insurtech startup. At a CEO roundtable in Aspen, she found herself nodding along to conversations about "foundation models" and "inference costs" without understanding what those terms meant or why they mattered to her P&L.

The pivotal moment came during a board meeting in October. A director -- a former Google executive -- asked a straightforward question: "Mariana, what is our AI strategy?" She gave the answer most CEOs give. She talked about the CTO's proof of concept, mentioned the word "transformative" three times, and referenced a McKinsey report about AI's trillion-dollar potential. The director pressed: "No -- what specifically can AI do for us, what can't it do, and how much will it cost?"

Mariana didn't have the answer. Not because she wasn't smart -- she'd built her career on analytical rigor -- but because nobody had given her a mental model for AI that mapped to how she actually made decisions.

That weekend, she canceled her Saturday plans, sat down with her laptop, and spent eight hours getting her arms around what AI actually is. Not the hype. Not the marketing materials from Salesforce and Microsoft. The actual technology -- what it does, how it works, where it breaks, and what it costs.

By Monday morning, she had a framework. By Q2 of the following year, Stratton-Byrne had three AI initiatives in production -- not dozens of unfocused experiments -- that were collectively saving $23 million annually in claims processing and underwriting efficiency.

This lesson gives you the same framework Mariana built for herself. In 25 minutes, you'll have a working mental model for AI that lets you make decisions, ask the right questions, and lead with clarity instead of buzzword bingo.

---

What AI Actually Is -- The Executive Mental Model

Let's strip away the jargon and build the understanding you actually need.

The Three Layers Executives Must Understand

Think of AI as a technology stack with three layers. You don't need to understand the engineering of each layer -- but you need to know what each one does and how it affects your business decisions.

Layer 1: Machine Learning (The Foundation)

Machine learning is software that learns from data instead of following hardcoded rules. Traditional software is like a recipe -- a developer writes step-by-step instructions for every scenario. Machine learning is more like training an employee by showing them thousands of examples until they recognize patterns on their own.

Your fraud detection system, your recommendation engine, your demand forecasting model -- these all use machine learning. They've been running in enterprises for a decade. Nothing new here.

Layer 2: Large Language Models (The Breakthrough)

Large language models -- LLMs -- are the technology behind ChatGPT, Claude, Gemini, and the explosion of AI you've been hearing about since late 2022. Here's what happened in plain terms:

Researchers trained neural networks on essentially the entire written output of human civilization -- books, articles, code, financial filings, legal documents, medical records, academic papers, conversation transcripts. Trillions of words. The training process is simple in concept: the model reads a passage and tries to predict the next word. It gets feedback on whether it was right. It adjusts. Then it does this billions of times.

The result is a system that doesn't just predict words -- it has absorbed the patterns of human reasoning, argument, analysis, and communication. When you prompt an LLM, it doesn't look up answers in a database. It generates a response by drawing on all the patterns it absorbed during training.

Layer 3: Generative AI (The Application)

Generative AI is the application layer -- the products and tools built on top of LLMs and similar models. ChatGPT is a generative AI product built on OpenAI's LLM. Microsoft Copilot is a generative AI product integrated into Office 365. When your marketing team uses AI to draft campaign copy, or your legal team uses it to summarize contracts -- that's the generative AI layer.

What AI Can Do Today -- An Honest Assessment

The gap between what AI vendors promise and what AI actually delivers is where billions of dollars go to die. Here's the honest breakdown.

AI is genuinely excellent at:

• Synthesizing large volumes of unstructured information. A human analyst takes 40 hours to read 200 earnings call transcripts and identify themes. AI does it in 3 minutes. This isn't hype -- it's measurable and repeatable.

• Drafting structured written content. First drafts of memos, reports, communications, analysis frameworks, meeting agendas, board materials. AI won't produce final-quality output, but it can produce a 70–80% draft in seconds that a human polishes in minutes.

• Analyzing patterns in text and data. Sentiment analysis across thousands of customer reviews. Anomaly detection in financial data. Categorizing support tickets. Pattern recognition at scale.

• Translation and reformatting. Taking technical content and making it accessible. Converting a 50-page report into a 2-page executive summary. Translating across languages. Adapting tone for different audiences.

• Code generation and technical work. AI can write, debug, and explain software code. This has massive implications for your technology organization's velocity and cost structure.

AI is mediocre at:

• Quantitative analysis with precision. LLMs are language models, not calculators. They can reason about numbers in broad strokes but make arithmetic errors. Never trust AI-generated financial calculations without verification.

• Sustained logical reasoning across many steps. AI can handle 3–5 step logical chains well. Beyond that, error rates climb. For complex strategic analysis with many interdependent variables, AI is a thinking partner -- not an autonomous analyst.

• Creative work that requires genuine originality. AI is very good at recombination -- mixing existing ideas in new ways. It is not good at paradigm-breaking creativity. It can write a competent marketing campaign. It cannot invent a new business model.

AI fundamentally cannot do (today):

• Access information it wasn't trained on or given. AI doesn't know your company's Q3 pipeline, your board dynamics, or what your CFO said in the hallway yesterday -- unless you explicitly provide that context.

• Exercise judgment about organizational politics, culture, and relationships. The most important decisions executives make involve weighing factors AI cannot see: team morale, board sentiment, regulatory relationships, competitive signaling.

• Guarantee factual accuracy. This is the critical one. AI generates plausible text, not verified text. It will state fabricated statistics with the same confidence as real ones. It will cite studies that don't exist. It will attribute quotes to people who never said them.

The Cost Structure Executives Need to Understand

AI costs are unlike traditional software costs, and most executives price it wrong. Here's the breakdown.

API Costs (Variable)

If your organization uses AI through APIs -- calling models like GPT-4, Claude, or Gemini programmatically -- you pay per token. Tokens are chunks of text, roughly 3/4 of a word. Costs range from $0.25 to $30 per million input tokens depending on the model. Output tokens (what the AI generates) cost 3–5x more than input tokens.

For context: processing 1,000 customer support emails through an advanced AI model might cost $15–$50. Doing the same task with human agents costs $5,000–$15,000. The unit economics are compelling -- but they scale. If you're processing millions of documents, API costs become a material line item.

Talent Costs (Fixed, High)

The scarcest resource in AI is not the technology -- it's the people who know how to deploy it effectively. An experienced AI/ML engineer commands $250K–$450K in total compensation. A team capable of building and maintaining production AI systems -- a lead engineer, two ML engineers, a data engineer, and a product manager -- represents $1.5M–$2.5M in annual fully loaded cost. This is before you've written a single line of code.

Infrastructure Costs (Variable)

If you're training or fine-tuning models (most companies shouldn't be), GPU compute costs are significant -- $10K–$100K+ per training run. If you're running models in production on your own infrastructure, GPU server costs run $20K–$50K per month per server. Most organizations should use cloud APIs to avoid this entirely.

The Hidden Cost: Organizational Change

The most expensive part of AI adoption isn't the technology. It's the workflow redesign, training, change management, and organizational friction. A $500K AI tool that nobody uses because the rollout was botched costs you $500K plus the opportunity cost of 6 months of executive attention. This is where most AI initiatives actually fail.

Key Metrics Executives Should Track

You can't manage what you don't measure -- but most companies are measuring the wrong things with AI. Here are the metrics that matter.

Adoption Metrics (Leading Indicators)

• Active AI users / total eligible users. What percentage of your workforce is actually using AI tools? Industry average is 15–25%. Top-performing organizations hit 60%+.
• Use cases in production. How many AI use cases have moved past pilot into daily operational use? Beware "pilot purgatory" -- dozens of experiments, nothing in production.
• Time from pilot to production. Best-in-class organizations move a use case from pilot to production in 8–12 weeks. If your timeline is 6+ months, your process is broken.

Value Metrics (Lagging Indicators)

• Hours saved per week per user. Measure actual time savings, not projected savings. Survey users monthly. Target: 3–5 hours per knowledge worker per week.
• Cost per AI-assisted transaction vs. manual transaction. Calculate the fully loaded cost of processing a claim, underwriting a policy, or drafting a contract with and without AI.
• Revenue influenced by AI. Track deals where AI-generated insights, content, or analysis contributed to the win. This is harder to measure but critical for justifying investment.

Risk Metrics (Guardrails)

• Hallucination/error rate in production outputs. Sample AI-generated content regularly and measure factual accuracy. Set a threshold and monitor it.
• Data exposure incidents. Track any instances where sensitive data was sent to AI systems inappropriately.
• Compliance flags. Count AI-generated content that required revision for regulatory compliance reasons.

Real Company Examples

A Global Law Firm (1,200 attorneys, $2.1B revenue) deployed AI for contract review and due diligence. Results after 12 months: 40% reduction in associate hours on routine document review, $18M in annual labor cost savings, and -- counterintuitively -- higher accuracy than human-only review because the AI caught boilerplate inconsistencies that fatigued associates missed. Key learning: they succeeded because they framed AI as a tool that made associates more valuable, not a replacement. Associate satisfaction actually increased.

A Mid-Market Manufacturer ($400M revenue) tried to build a custom AI demand-forecasting model. After 14 months and $2.8M in development costs, the model performed 3% better than their existing statistical model -- a negligible improvement that didn't justify the investment. They scrapped it and instead deployed a commercial AI tool for sales team enablement that cost $120K annually and increased quote-to-close rates by 11%. Key learning: the highest-value AI use cases are often not the most technically impressive ones.

A Regional Health System (8 hospitals, 12,000 employees) deployed AI for clinical documentation -- helping physicians complete notes faster. Physician satisfaction scores increased 22%, documentation time dropped by 35%, and the system recovered $8.4M annually in previously uncaptured billing codes. Key learning: the ROI came not from replacing anyone but from reducing the administrative burden on the highest-paid, most constrained resource in the organization.

---

Apply: Your Executive AI Workout

Now let's put this knowledge to work. The following exercises are designed for your actual workweek -- not hypothetical scenarios.

---

Reflect: The Executive's Real Advantage

Remember Mariana Delgado from the beginning of this lesson? Her breakthrough wasn't becoming technical. She never learned to write code, never trained a model, never read a machine learning research paper. Her breakthrough was building a mental model that let her lead AI adoption instead of reacting to it.

The framework she built -- and the one you now have -- comes down to four principles:

1. Know what AI is. It's a pattern-recognition system trained on human language. It generates plausible responses, not verified truth. It's a brilliantly well-read chief of staff who has read everything and experienced nothing.

2. Know what it costs. API costs are cheap. Talent is expensive. Organizational change is the hidden budget-killer. Most companies should buy, not build.

3. Know what to measure. Adoption rates, hours saved, cost per transaction, error rates. Not "models deployed" or "innovation initiatives launched."

4. Know your role. AI provides speed, breadth, and analytical horsepower. You provide judgment, context, and decision-making authority. The executives who treat AI as a replacement for thinking get burned. The executives who treat it as an accelerator for thinking build competitive moats.

Six months after that uncomfortable board meeting, Mariana presented her AI strategy again. This time, she didn't mention McKinsey. She showed three production use cases with measured results. She showed the cost structure and ROI timeline. She showed the risk framework and governance model. The former Google executive on her board -- the one who'd asked the question that stumped her -- sent her a note afterward: "That was the clearest AI strategy presentation I've heard from any portfolio company. Most CEOs still can't answer the question I asked you."

The difference wasn't that Mariana became an AI expert. The difference was that she stopped delegating her understanding of the technology to people who had incentives to make it sound more complicated -- or more magical -- than it is.

That clarity is now yours. The next lesson in this chapter takes you deeper into the AI business landscape -- who the key players are, how the market is structured, and what the competitive dynamics mean for your strategy. But the mental model you built here is the foundation everything else sits on.

You don't need to understand neural network architecture. You need to understand what AI can do, what it can't, what it costs, and how to lead it.

Now you do.