Where the money really goes when AI does the work.

You pay for reading and writing #

Most AI models charge by the token, which is roughly three-quarters of a word. You pay for every token the model reads (your instructions, the documents, the conversation so far) and every token it writes. Writing usually costs more per token than reading.

The surprise for most people is that reading dominates. An agent doing a long task rereads its whole working context on every single step. A brief with a 50-page document attached gets read again and again, dozens of times.

The five things that make bills balloon #

When a bill surprises someone, it's almost always one of these:

• Bloated context. Stuffing every document and the full history into every step. The agent rereads it all each time, and usually thinks worse for it, too.
• Retries. An agent stuck on a problem can burn through a lot of attempts. Retrying the same approach is the expensive kind.
• Agents talking to agents. Every handoff means one agent writing a brief and another reading it, plus everything they send back and forth.
• Supervision nobody needed. A manager agent restating a clear brief or approving work that already passed its checks costs a lot and adds little.
• The wrong model for the job. Using the most powerful model to rename files is like hiring a surgeon to put on a plaster.

How to keep it sane #

None of this is complicated. It's mostly the same discipline as running a team of people:

• Brief tightly. Give each worker what it needs for its job, not everything. It's cheaper and the work is better.
• Use cheaper models for routine work. Save the most capable model for judgement, taste and hard problems.
• Let stable instructions be cached. Many providers charge much less for text they've seen before, so keeping the unchanging part of a brief at the start and identical each time saves real money.
• Keep teams small. One capable worker under one owner is the sensible default. Add more only when they genuinely save time or catch problems.
• Set a budget the system enforces. A monthly cap with an early warning, broken down by agent and model, so you can see which job is expensive before the bill arrives.

The second budget: your attention #

Here's the cost that never shows up on the invoice. Every time the AI asks you something, you pay in interruption, in catching up on what happened, and in actually making the call. A system that saves a few dollars of model time by asking you a question every ten minutes is a very expensive system.

So when you compare options, count both budgets. The right question isn't "what does the model cost?". It's "how much useful finished work do I get, for the money and the attention it takes?"

Comparing it to people #

People often ask how AI compares to a contractor or a virtual assistant. It's a slightly odd comparison, because they're good at different things. AI is fast, tireless and cheap per hour of work, but needs clear briefs and checking. A good contractor brings judgement, relationships and accountability.

The fairer comparison is per finished piece of work, including the time you spend briefing and checking. For a lot of routine research, drafting and production, AI wins comfortably. For anything that needs a relationship or real accountability, you still want a person, maybe with AI doing the legwork for them.