Mary Fung
noteJune 22, 2026

What are tokens?

Tokens are the hidden meter behind AI cost, speed, and context. Leaders do not need the math, but they do need the model.

Tokens are how most AI systems measure language.

You can think of a token as a small piece of text. Sometimes it is a word. Sometimes it is part of a word. Sometimes it is punctuation. The exact math is not the point for most leaders.

The point is simpler:

Tokens are the meter running in the background.

They affect cost, speed, and how much information the AI can consider at one time.

When you send a prompt to an AI model, that prompt uses tokens. When the model sends an answer back, that answer uses tokens too. Those are usually discussed as input tokens and output tokens.

Input tokens are what you send in.

Output tokens are what the model generates back.

Both matter.

Why executives should care

Tokens turn AI from a vague capability into an operating cost.

If a team pastes a long document into a model, that uses more input tokens. If the model writes a long answer, that uses more output tokens. If the same workflow runs hundreds or thousands of times, small choices start to matter.

This does not mean every AI interaction needs to be painfully short. Sometimes the model needs a lot of context to do the job well.

But it does mean leaders should understand the tradeoff:

More context can improve quality. It can also increase cost, slow the workflow, and make the system harder to manage.

More output can be useful. It can also create noise, review burden, and unnecessary spend.

Input cost: what you send in

Input tokens are driven by the amount of information you give the model.

Examples:

This is why good AI workflows do not just dump everything into the model.

They decide what the model actually needs for the task.

Output cost: what comes back

Output tokens are driven by the answer you ask the model to produce.

A one-paragraph summary costs less than a ten-page memo. A concise table costs less than a sprawling explanation. A terse internal note costs less than a polished executive brief.

This is one of the easiest levers to control.

If the job is internal triage, you may not need beautiful prose. You may need short bullets, labels, and next actions.

That is where formats like terse-response modes or "caveman" style instructions can be useful. The point is not to sound primitive. The point is to force the system to spend fewer words on ceremony and more on the actual answer.

For example:

Give me the decision, the risk, and the next action. No background unless it changes the decision.

That kind of instruction can reduce output length, speed review, and make recurring workflows cheaper.

The trap: treating tokens as a technical detail

Tokens are technical, but their consequences are managerial.

They affect budget. They affect latency. They affect whether teams can afford to run a workflow at scale. They affect how much context the AI can use before older information falls out of view.

If a team says, "We need more context," the leader's follow-up should be:

Which context improves the decision, and which context is just making the prompt bigger?

That question is not micromanagement. It is operating discipline.

Questions leaders should ask

When reviewing an AI workflow, ask:

The goal is not to starve the model of useful context.

The goal is to stop paying for avoidable noise.

← back to the field