Start Lesson
A startup I advised built an AI customer support bot. The demo was great. The team was thrilled. Three months into production, they discovered each conversation cost $0.85 — but only saved $0.50 in staff time. They were losing thirty-five cents on every single interaction. At 10,000 conversations a month, that is $3,500 per month spent to make their business less efficient.
The technology worked. The economics did not. And nobody had done the math before building.
In the last lesson, you learned that cost is one of the three project killers. This lesson gives you the tools to calculate whether an AI feature is worth building before you build it. After this lesson, you will be able to estimate the cost of any AI feature and compare it to the value it creates.
AI models do not process words — they process tokens, small chunks of text roughly four characters or three-quarters of a word long.
"Economics" is two tokens: "econ" and "omics." "Hello, how are you?" is about six tokens. A full page of text is roughly 500 to 700 tokens.
Why does this matter? Because you are billed per token. Every token the model reads (your input) and every token it writes (its output) has a price. This is the meter running every time you use AI.
AI pricing has two components:
Here is what pricing looks like for popular models (per million tokens, as of early 2026):
| Model | Input Cost | Output Cost | |-------|-----------|-------------| | GPT-4o | $2.50 | $10.00 | | Claude Sonnet | $3.00 | $15.00 | | Claude Haiku | $0.25 | $1.25 | | Gemini Flash | $0.10 | $0.40 | | GPT-4o Mini | $0.15 | $0.60 |
The most capable models can cost 100 times more than the smallest ones. The critical insight: for many tasks, the cheaper model works just fine. Sorting customer feedback into categories? A small model handles that. Writing a nuanced strategy document from complex source material? You probably want a larger model.
This connects directly to the adoption levels from Lesson 3. Level 1 tasks (human reviews everything) can afford a top-tier model because volume is low. Level 2 tasks (automated, high volume) need to use the right-sized model or costs spiral.
Here is where AI projects live or die. The technology can work perfectly and still be a bad investment.
Worked example. You build an AI customer support assistant. Each interaction involves:
With a mid-range model (like GPT-4o Mini), that costs roughly $0.01 to $0.03 per interaction. At 15,000 monthly interactions, your AI bill is $150 to $450. If each interaction saves $2.00 in staff time, your monthly savings are roughly $29,500. Strong economics.
Now run the same math with a top-tier model. Each interaction costs $0.15 to $0.30. At 15,000 interactions, that is $2,250 to $4,500. Still profitable — but margins shrank by an order of magnitude.
Now add complexity: four AI calls per interaction instead of one (classify the question, search the knowledge base, draft the response, check the response for accuracy). Costs quadruple. At $0.60 to $1.20 per interaction with the top-tier model, your monthly bill hits $9,000 to $18,000. Suddenly the savings are thin or gone.
The formula is simple: Cost per AI interaction multiplied by monthly volume. Compare that number to the value created. If cost exceeds value, the feature is a liability — no matter how impressive the demo was.
You are not stuck with the first price you see. These are the levers that teams use to make AI economics work.
1. Model selection. The biggest lever. I have seen teams cut costs by 90% by moving classification tasks from a flagship model to a smaller one — with no meaningful quality drop. Match the model to the task, not the other way around.
2. Caching. If customers ask the same twenty questions repeatedly, store the response instead of recomputing it every time. This can eliminate 30-60% of AI calls overnight.
3. Prompt optimization. A 2,000-token prompt that could be 800 tokens costs 2.5 times more on every single request. Trimming unnecessary instructions, removing redundant context, and being precise with what you include is one of the highest-return optimizations.
4. Batching. Send multiple tasks together instead of making separate calls. Many providers offer batch pricing at a 50% discount for workloads that do not need real-time responses.
5. Tiered routing. Use a cheap, fast model for the first pass. Only escalate to the expensive model for complex cases. Think of it like a triage system — a nurse handles routine questions, and the specialist only sees the difficult ones.
Pick one task where you use (or plan to use) AI. Estimate the economics using these steps:
You do not need exact numbers. A rough estimate is enough to tell you if the economics are in the right ballpark or wildly off. If the numbers do not work with a top-tier model, try a cheaper one — that is lever number one.
This is the final lesson of "What AI Actually Is." Here is the full picture you have built across five lessons:
Lesson 1: AI is a prediction engine — not a thinking engine. It predicts the next word based on patterns learned from the internet. This is the foundation that explains everything else.
Lesson 2: Prediction excels at pattern-based tasks (drafting, summarizing, classifying) and fails at precision tasks (math, real-time data, novel reasoning). Hallucinations are a built-in consequence, not a bug.
Lesson 3: There are three levels of AI adoption — assistance, automation, autonomy — and most businesses should start at Level 1 and earn their way up deliberately.
Lesson 4: The gap between an impressive demo and a working production system is where most AI projects die. Cost, hallucinations at scale, and lack of evaluation are the three killers.
Lesson 5: Every prediction costs money. Understanding tokens, pricing, and unit economics is the difference between a profitable AI feature and an expensive experiment.
Together, these five ideas give you something most people do not have: a grounded, practical mental model for thinking about AI. Not the hype. Not the fear. Just a clear understanding of what the technology does, where it works, and what it costs.
You are ready for the next course: Prompt Engineering That Works — where you will learn to write inputs that get reliably useful outputs, every time.
