You don't need a PhD to understand where AI comes from, but you do need to know why 2023-2024 mattered so much. The difference between 1990 and 2020 was incremental. The difference between 2022 and 2024 was exponential. Here's the condensed history of AI for the executive who wants to understand the stakes.
The Long Slow Climb (1950s-2012)
Artificial intelligence as a concept goes back to Alan Turing in 1950 (the Turing Test: can you tell if you're talking to a human or a machine?). But the practical reality: AI was slow, brittle, and expensive for 60 years. Chess computers beat Kasparov in 1997, but chess is a game with clear rules. Teaching a computer to recognize a cat took until 2012.
In 2012, Geoffrey Hinton's team won the ImageNet competition using deep neural networks — a resurgence of a technique from the 1980s that had finally become practical thanks to GPU computing power. Suddenly computers could recognize objects in photos. This started the "deep learning" era.
For a CFO, the lesson: AI was an academic curiosity for 60 years. Companies that tried to monetize it (IBM expert systems, robotics, autonomous driving before 2020) mostly failed. The economics didn't work. Hardware was too expensive, algorithms were too fragile, and the use cases didn't justify the cost.
The Data Era (2012-2022): Machine Learning Goes to Work
Between 2012 and 2022, machine learning became practical and profitable in specific domains. Netflix recommendations, Google search ranking, fraud detection, credit scoring — these use cases had value, scale, and data to train on. Companies built machine learning teams, hired data scientists, and started to see ROI.
But here's what mattered: this was all supervised learning. You needed labeled data. To train a fraud detection model, you needed a million examples of "this is fraud, this is not." To train a recommendation engine, you needed billions of user interactions. The economics: if you had data and scale, ML worked. If you didn't, it was a cost center.
For a CFO, the lesson: machine learning was a game for big tech and large enterprises with massive datasets. The data moat was real. Smaller companies couldn't compete.
The Transformer Moment (2017-2020)
In 2017, Google researchers published a paper called "Attention is All You Need" introducing the transformer architecture. Transformers are the underlying tech behind GPT, Claude, Gemini — every large language model you've heard of. But adoption was slow. Transformers required enormous computing power. Google had it. Facebook had it. Most companies didn't.
The turning point: in June 2020, OpenAI released GPT-3 (175 billion parameters, trained on 300 billion tokens of text). For the first time, a language model could do things it wasn't explicitly trained for. Show it a few examples of "translate English to French" and it would translate. Show it a math problem and it would solve it. This is called few-shot learning, and it was revolutionary.
But GPT-3 was expensive. Each API call cost money. No one had a business model for it yet. It was a toy.
For a CFO, the lesson: GPT-3 proved that general-purpose AI was possible. But the economics were still wrong.
The Inflection Point (November 2022-Present)
On November 30, 2022, OpenAI released ChatGPT. It was free to use, easy to understand, and capable of things that shocked people. Within two months, 100 million people used it. The AI hype cycle began.
What changed: (1) User experience improved dramatically. GPT-3 was an API; ChatGPT was a product. (2) Cost economics shifted. Inference on LLMs became cheap enough that you could run it on a customer service chatbot. (3) The problem went from "can we build AI?" to "how do we make money from it?"
Between November 2022 and May 2026, the capability curve went vertical. GPT-4 (March 2023), Claude 1 (March 2023), Claude 2 (July 2023), GPT-4V (October 2023), Claude 3 (March 2024), GPT-4o (May 2024), Claude 3.5 (October 2024), reasoning models (December 2024). Every 3-6 months, a new model with higher accuracy, cheaper cost, or new capabilities.
Meanwhile, the adoption curve followed. By May 2026, 88% of enterprise organizations use AI in at least one function (McKinsey State of AI 2025). But only 39% see measurable EBIT impact. And only 5.5% are "AI high performers" — meaning the technology actually improved their business.
For a CFO, the lesson: capability is no longer the constraint. Economics and integration are. The question is no longer "can we do AI?" It's "can we make money doing AI?"
The Three Shifts That Changed Finance
Three things happened between 2022 and 2026 that should matter to you:
1. From scarce to abundant capability — In 2022, if you wanted AI, you built it yourself. That required data scientists, ML engineers, GPUs, six months of development. By 2026, you can rent it from OpenAI or Anthropic for pennies per inference. Capability is now a commodity. This flattened the competitive moat for big tech and opened the door for everyone else.
2. From custom models to prompt engineering — The old ML playbook was "build a custom model on your data." The new playbook is "use a general-purpose model (GPT-4, Claude) and tell it what to do via prompt." This is 100x cheaper to get started. But it's also 100x easier for everyone to do, which means competition is fiercer.
3. From technology problem to economics problem — In 2020, the AI constraint was "can we build this?" In 2026, the constraint is "does this make money?" That's a CFO problem, not a CTO problem. The finance team now owns the AI business case. This is new, and most finance teams aren't ready for it.
Where We Are in May 2026
By 2026, AI has three distinct operating zones:
The Commodity Zone — Tasks that any LLM can do (summarization, classification, content generation). Margin is razor-thin. Everyone competes on cost. This is where you use the cheapest model (Gemini, GPT-4o-mini, etc.) and hope for unit economics that work.
The Specialized Zone — Tasks that require domain expertise or proprietary data (medical diagnosis, insurance underwriting, technical support for specific products). These require fine-tuning, custom guardrails, and often reasoning models. Margin is better, but the cost is higher. You need ROI to justify it.
The Agentic Zone — Autonomous systems that make decisions and take actions (support agents, sales agents, process automation). These are the most valuable but also the riskiest. They require cost attribution, observability, and governance. If you get them right, they print money. If you get them wrong, they become shadow AI and blow out your budget.
Most companies in May 2026 are in the commodity zone, struggling to break even. The winners in 5 years will be the ones who move to the specialized and agentic zones and build real cost attribution.
For a CFO, the implication is clear: AI isn't a technology problem anymore. It's a financial operations problem. Runrate exists because the finance OS for AI doesn't exist yet. Your vendor cloud costs are tracked, your headcount is tracked, and your marketing spend is tracked. But nobody knows what your AI actually costs.
The next chapter of AI isn't about model capability. It's about cost clarity and business impact. That's your job. Curious where your team sits on the 5-Stage AI Cost Maturity Curve? Take the 15-question self-assessment and get a personalized report on your path to work-item-level cost attribution.
Where does your team sit on the maturity curve?
Take the 15-question self-assessment and get a personalized report.
Was this article helpful?