The Next KPI Every AI Company Will Track: The LLM Token Expenditure Index

For the last two years, we've been asking the wrong question.

"What's the smartest AI model?"

It's an interesting question — but not the one that will matter most to businesses. The better question is: which AI gets the job done using the fewest resources?

We've been here before

Every mature technology eventually develops a metric that balances performance with efficiency. Early cars competed on horsepower; eventually fuel efficiency mattered more to buyers. Airlines compete on cost per seat-mile. Cloud providers compete on uptime and cost per compute-hour. Data centers measure Power Usage Effectiveness.

AI hasn't reached that stage yet. Today we compare models by benchmark scores, reasoning ability, context windows, and coding performance. Tomorrow, we'll compare them by something simpler: how much it costs to accomplish useful work.

Introducing the Token Expenditure Index (TEI)

A simple idea:

TEI = Total Tokens Consumed ÷ Successful Business Outcomes

Instead of measuring AI by intelligence alone, measure it by efficiency.

Now leaders have a KPI they can actually optimize — not "is our AI smart," but "is our AI efficient at the specific job we hired it for."

Why this matters

Traditional SaaS businesses are incredibly efficient. Once the software is built, serving another customer costs almost nothing. AI is different: every conversation consumes compute, every prompt has a cost, every retry costs more, every unnecessary paragraph generated costs more. Unlike traditional software, AI carries a real marginal cost every single time it's used.

As organizations spend hundreds of thousands — or millions — of dollars annually on AI, someone will eventually ask: where are all these tokens going?

The companies that win won't necessarily have the smartest AI

Imagine two AI systems solving the identical problem.

System A solves it using 20,000 tokens, needs two follow-up prompts, and occasionally requires human correction.

System B solves the same problem using 5,000 tokens, finishes in one interaction, and rarely needs review.

If the outcome is identical, System B is dramatically more valuable. The most intelligent model isn't always the most profitable one.

Token efficiency is more than cost

A useful index could include input tokens, output tokens, retry rate, tool/API calls, human review time, completion rate, response latency, and total AI cost per successful outcome. The goal isn't simply to reduce tokens — it's to maximize business value for every token spent.

The new optimization race

Today's AI teams focus on better prompts, better models, better RAG pipelines. Soon they'll also optimize for lower token usage, fewer retries, smaller context windows, smarter memory, better orchestration, and higher first-pass success rates. Call it what it is: AI efficiency engineering.

The dashboard every executive will want

Imagine opening an executive dashboard and seeing cost per AI interaction, tokens per successful outcome, AI spend by department, token efficiency trend over time, cost per employee, cost per customer, and AI ROI — laid out as plainly as cloud-spend dashboards are today. That's coming. It's the same fit-for-purpose thinking I explore in Your Business Probably Doesn't Need GPT-6.

The bigger shift

For years we've asked "which model is the smartest?" Soon we'll ask "which model generates the most business value per token?" That shift will change how companies choose models, design workflows, write prompts, and measure ROI.

The future of AI won't be won solely by bigger models. It will be won by more efficient systems. And when that happens, I believe every organization will begin tracking a metric like the Token Expenditure Index.

Not because tokens matter. Because business outcomes do.