We moved our AI to a cheaper model. Nothing broke.

Here’s what we’re reading and thinking about this week:

We get a lot of questions about token usage, AI costs, and how to make sure what you are setting up is sustainable.

Well, that exact topic was in the news this week.

Microsoft just changed how it charges for its newest AI tools with their new Copilot super app. Instead of one flat price per person, you pay for how much you use.

But even if you aren't on Microsoft, we expect this is going to be a growing topic of discussion in 2027.

Why? We are all already “paying” this way.

• If you have an AI subscription —> you really have a quota of tokens you can use each month

• If you pay for AI usage —> you likely have a budget of what you can spend $ each month

And the reality is that a lot of what we are paying for is wasted...

Think about the office manager who re-explains the vendor list in every new chat, or the marketing lead who pastes in five old docs and regenerates six times. Or the accountant uploading the whole ledger to ask one question.

In today’s edition we are going to break this down and give you practical tips for how to manage your token usage better:

Most of your token usage is the AI getting oriented

Think about a new hire on day one. Hand them a pile of old email threads and say "figure out what we need," and they'll spend the morning just reading. Hand them a one-page brief and they start in five minutes.

The AI is that new hire every single time you open a chat... and it reads the whole pile again.

Turns out that's where most of the bill goes. Researchers tracking what AI agents spend found most of the cost comes from the AI reading, not writing, and the runs where it wandered around cost the most and succeeded the least.

That’s one reason to use playbooks. (By playbook we just mean a written-down process: the steps, the inputs, and what good looks like.) So the AI spends less time reading and guessing, and gets it right in fewer tries. In one study of 87 tasks, giving the AI a written playbook raised its success rate from about one in three to about one in two.

And in that study, shorter playbook did better than long ones. So you don't need a 12-page manual. Clear instructions is what you need.

The hidden benefit of playbooking: the cheaper model suddenly works

In the same study we linked above, smaller models with a written procedure matched bigger models without one. Which makes sense when you think about it... a lot of what you're paying the top model for is working out what you meant.

That's why one of the hidden benefits of playbooking is that when you're clear about what you want the AI to do, you can actually use cheaper AI models, which spend less of your budget.

Let me give you an example:

We used to run our email triage (sorting what needs a reply from what can wait) on the most expensive model. It already had a written playbook, so we tried it on the mid-tier model instead.

Did it get worse? Nope. Quality held, and it actually got faster.

At a fraction of the cost.

What not to do: have the AI write the playbook for you

As people start to see the benefit of playbooking, they can be really tempted to want to run as fast as they possibly can. Part of that is having the AI write playbooks for you…

But this can be a trap.

In that same study, procedures the AI wrote for itself did worse than giving it nothing. Kind of wild, right? What makes a playbook work is your judgment about what good looks like, and the AI can't write that part for you.

Our take: a playbook AI wrote for you can be worse than no playbook at all.

"But AI keeps getting cheaper"

Fair, it does. Anthropic just released new models it says cost about 40% and up to 30% less to run than the ones before.

Both things are true, though. If a task takes six tries, you're still paying for six tries, just at a lower price. And honestly, the money is the smaller win. A good playbook gets you the result faster, and it gets you the same good result on Tuesday that you got on Monday.

What to do this week

1. Pick a playbook you've already written for something you run every week.

2. Run it on a cheaper model five times and put the results next to what you get now.

3. If it holds, switch. If it slips, look at where it slips. That's the line in your playbook that's too vague. Fix it and try again.

What's one task you'd trust to a cheaper model tomorrow? Hit reply and tell us - we'll tell you if we'd make the same call.

LINKS

For your reading list 📚

• An OpenAI agent got into Australia's Medicare stats portal because it "didn't accept 'no' for an answer." OpenAI told the government about three months later.

• Claude ran about 950 agents for 21 hours and found a CRISPR-like enzyme system. It burned through 210 million tokens. We hope somebody wrote a playbook….

• Steal this one: Pew published exactly what it will and won't use AI for. Two lists, one page.

• Insurers say hospital AI billing tools added $942M in costs with no change in care. AI hits whatever number you point it at.

• A Samsung update shut down AI fridges across Korea right before a major holiday. Smart fridge, rough week.

That's all!

We'll see you again soon. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected]

Cheers,

🪄 The AMP Team (formerly: the AI Exchange Team)