Model providers tell us AI gets cheaper every quarter. So why is it that the bills at the firms I advise say the exact opposite? Both are right, even if that sounds contradictory.

We see the unit cost of AI falling sharply. Stanford’s AI Index found that running a model at the level of the original ChatGPT went from about twenty dollars per million tokens to a few cents in under two years. Open source models are only accelerating it, and many of them perform well.

Three reasons the bills go up

So why do the bills go up? Three reasons:

• Reasoning models spend more compute per answer. They explore options and check their own work before they respond. Very helpful on harder problems, but every request costs more.

• Agents remove the natural ceiling. Ask a model a question yourself and there’s a limit to the work behind it. Hand the task to an agent, and that limit is gone.

• Cheaper AI just gets used more. When something useful drops in price, you reach for it constantly.

The three compound. A cheaper model invites you to use an agent, the agent runs a reasoning model many times over, and the total lands well above what the price per token suggests.

None of this means the price drop is fake. Your cost per unit of work really is falling. What rises is how much work you hand over.

Measure, but not everything

All of this is an argument for measuring AI use properly. Often I hear leaders say their token bill is steep, yet they don’t know what came out of the work.

The usual reflex is to start measuring everything. I’d be hesitant. Instead, I’d use three layers.

Measuring has a cost of its own, and most of it is paid by the people being measured. If the reporting takes longer than the work it is tracking, you will get bad data and some resentment with it.

Layer one: AI as a general assistant

When AI is a general assistant, treat it like any other workplace tool. Think someone asking Claude to clean up an email or talk through a difficult spreadsheet.

• Who has access.

• Who actually uses it.

• Whether you’re paying for overlapping tools.

• Whether people report saving time.

Don’t ask employees to tag every conversation just to get a read on usage.

Layer two: a recurring workflow

When a team runs the same workflow through AI again and again, you can be more precise. Say drafting the monthly variance commentary. Look at the frequency. Test how long the work took before, and how often it needs redoing. Then decide whether to expand it, redesign it, or drop it.

The redo rate is the number worth watching here. It tells you more than the time saved, because a draft that needs heavy correction every month is not saving anything, whatever the first measurement said.

Layer three: AI inside a production process

Full precision becomes essential when AI runs inside a production process. Processing invoices. Resolving customer cases. These generate a lot of activity without a person behind each call. Here you need to know how often it succeeds, which cases need rework, and what the whole thing costs you.

This is also the layer where cost can run away without anyone noticing, because no person sits behind each call to feel it. Put a ceiling on it and an alert when the ceiling is approached.

As a rule of thumb: measure loosely where AI assists a person, carefully where it reshapes a workflow, and precisely where it runs a process.

Originally posted on LinkedIn