The Economics of AI — ROI Frameworks and Cost Structures

Sarah's $81,000 Surprise

Sarah is VP of Finance at a 2,000-person SaaS company. Her engineering team wanted to launch an AI chatbot for 100,000 users. They estimated the cost at $10,000/month. They put it on a board slide. The board nodded.

Then Sarah actually ran the numbers.

$81,000 per month.

Eight times what engineering promised. The slide was already in the board deck. The project was already approved. And nobody had done the math.

This chapter makes sure you never get blindsided like that. By the end, you'll be able to calculate any AI bill on the back of a napkin — and you'll actually understand what you're paying for.

AI Costs Work Like a Phone Bill

Seriously. If you understand your phone bill, you understand AI pricing. Here's the analogy:

That's it. Three things control your AI bill:

1. Input tokens — the text you send to the AI (your question, background info, instructions)
2. Output tokens — the text the AI sends back (its answer)
3. Model tier — which AI "brain" you pick (cheap and fast, or expensive and smart)

Wait, what's a token?

A token is a small chunk of text — roughly three-quarters of a word. A common word like "hamburger" is 1 token; a long word like "uncharacteristically" can be 5–6 tokens. The sentence "What is the weather today?" is around 6 tokens.

Every token costs money. Input tokens and output tokens have different prices. And those prices change dramatically based on which model you pick.

The Price Menu

Before you look at the numbers: Guess how much more expensive the most powerful AI model is compared to the cheapest one, for the same task. Is it 2x? 5x? 20x? Write down your estimate — the real number tends to surprise people.

Here's what the three model tiers actually cost (using Anthropic's Claude as our example):

Prices are illustrative. AI token costs have dropped significantly and change frequently — verify current rates at provider pricing pages.

Look at those numbers. The premium tier costs 60x more than the fast tier for input tokens — model names and pricing tiers change; use the provider's current pricing page for the exact model lineup. Same question, same answer format — 60x the price. That's the single biggest cost lever in your entire AI budget, and it's an architecture decision your engineering team might make in a five-minute Slack thread.

Let's Do Sarah's Math — Step by Step (using the illustrative prices from the table above)

No skipping ahead. We're going to walk through Sarah's calculation the way she should have before that board meeting.

The setup:

• 100,000 users
• Each user asks 5 questions per day
• Each question sends ~800 input tokens and gets back ~200 output tokens
• Engineering picked Sonnet for everything

Step 1: How many queries per day?

Step 2: How many tokens per day?

Step 3: What does Sonnet charge?

Step 4: Monthly bill

Engineering estimated $10,000. The real number is $81,000. That's an 8.1x miss.

The answer: Haiku costs $6,750/month for the same workload. That's 12x cheaper than Sonnet.

But here's the catch — Haiku is less capable. It might fumble your complex queries. So you can't just swap everything to Haiku and call it a day. You need a smarter approach.

Model Routing = Call Center Staffing

Think about how a well-run call center works:

• Simple questions ("What's my balance?" "How do I reset my password?") go to junior agents. They're fast, cheap, and perfectly capable.
• Complex questions ("I was double-charged on a disputed transaction across two accounts") go to senior agents. They cost more, but they get it right.

You wouldn't pay senior-agent rates for every single call. That would be insane.

Model routing works the same way:

The blended cost math

Let's recalculate Sarah's bill with a 70/30 routing split:

Haiku handles 70% of queries (350,000/day):

Sonnet handles 30% of queries (150,000/day):

Blended monthly cost: ($157.50 + $810) x 30 = ~$29,000/month

That's a 64% reduction from the all-Sonnet bill, with quality preserved where it matters.

The Secret Weapon: Prompt Caching

There's one more trick that can slash your bill — and it requires almost no engineering effort.

Think of it like a photocopier. Every time your chatbot answers a question, it re-reads the same set of instructions (the "system prompt"). That's like typing the same cover letter from scratch every time you apply for a job. Ridiculous, right?

Prompt caching stores a copy of those instructions so the AI doesn't re-read them every time. The cached tokens cost 90% less.

Here's a real example:

Your engineering team can enable this in a single sprint. It's free money.

The Cost Diagram — How It All Fits Together

Three levers, one bill. You control all of them.

The Board-Ready Playbook

Sarah revised her board slide, engineering added routing logic in one sprint, and the board approved the project at $29K/month instead of $81K. Here's what she learned — and what you should do before approving any AI budget:

Step 1: Get the real numbers — users, queries/day, tokens per query, model tier. Do the multiplication yourself. It takes ten minutes.

Step 2: Ask engineering about model routing. If they're using one model for everything, that's like paying senior-agent rates for password resets.

Step 3: Ask about prompt caching. If the system prompt is longer than 1,000 tokens and they haven't enabled caching, you're leaving money on the table.

Step 4: Stage the investment with gates. Don't release the full budget based on pilot results — tie funding to production milestones with kill criteria at each gate.

Back to Sarah

She walked back into the board meeting with new numbers.

"Engineering estimated $10,000 a month. I ran the actual model — at our user volume, with the responses that system generates, the real cost is $81,000. I'd like to propose we approve the project with three conditions: we route routine queries to the cheaper model tier, we add prompt caching for the system prompt, and we set a $20,000 monthly hard cap while we validate usage in the first 90 days."

The board approved it. The project shipped. First-month costs: $18,400.

The $81,000 surprise turned into a $63,000 saving — not because the engineering team was wrong about AI, but because someone ran the numbers before the commitment was made.

Key takeaways

• You can prevent 10-100x budget surprises by modeling input tokens, output tokens, and model tier before committing to any AI architecture.
• Every time you route simple queries to a cheaper model, you save 60-80% of AI spend without users noticing any quality difference.
• You can cut input costs by up to 90% for repeated instructions by enabling prompt caching on any app with a large, fixed system prompt.