Lesson 7 · 13 min
With an agent running all day, usage grows fast: on a subscription you hit the limit; paying per use, it becomes a big bill. In my case, almost everything runs on the subscription. As with any employee, you need to audit and watch the cost. First what applies to everyone: the right model and keeping track of the bill. Then, for those who pay per use, cache and batch.
Start the lesson
Goal: choose the right model for each recurring task, track a week of usage and know when cache and batch are worth it.
Tap or click each one.
Audit it the way you would audit an employee: what it delivers, how much it costs and whether the model is the right one for the job.
Three questions for each recurring task. Tap or click each one.
Simple and repetitive goes to the smallest; everyday text and analysis to the middle one; anything with risk or that decides money to the largest.
If in doubt, run the procedure from lesson 6 on both and compare.
Sorting emails by subject: smaller model. Proposal for a client: larger model.
If you only use the chat subscription, cache and batch don't show up for you: the provider takes care of that.
What counts is picking the right model, not repeating the context for nothing (the project already keeps it) and noticing when you hit the limit.
Hit the limit every Thursday? See which task used the larger model without needing to.
If nobody is waiting for the answer right now, the task can go in a batch for half the price.
Gather the day's tasks and send them all at once.
Weekly summary of 200 customer reviews: batch.
If every call starts with the same rules and documents, keep that start identical and at the beginning.
What changes (the question, the case) goes at the end.
Customer service: rules and price table always the same at the start; the customer's message at the end.
One table per agent. Replace the text between < > with yours.
# Cost per task · <agent> · week of <date> | task | times per week | model | reason | on the spot or in batch | fixed start (cache)? | usage noted | |---|---|---|---|---|---|---| | <task 1> | | smaller / middle / larger | <why, in one sentence> | on the spot / batch | yes / no | <hit the limit? or the amount> | | <task 2> | | | | | | | | <task 3> | | | | | | | | <task 4> | | | | | | | | <task 5> | | | | | | | Decision of the week: <what changes model or moves to batch>
Below is the list of tasks you do for me every week. For each one, tell me: 1. whether it really needs the larger model, or whether the smaller one can handle it (and how I test that); 2. whether it can wait and go in a batch; 3. whether the start of the request is always the same (and what I keep fixed to take advantage of the cache). Answer in a table and don't change anything yet. <list of tasks>
Decision: the summary goes to batch; the email sorting drops to the smaller model after it passes the 5-round test.
One week of real use.
List the five tasks you hand off to AI the most and fill in the cost-per-task table with model, on the spot or batch, and fixed start (if you only use a subscription, mark "on the spot").
Done when: all five rows have a model and "on the spot" or "batch" marked, with the reason in one sentence.
For one week, note the usage of each task: on a subscription, whether you used the larger model and whether you hit the limit; on the API, the amount.
Done when: the "usage noted" column is filled in for the whole week.
With the week in hand, pick one change (smaller model, batch or fixed start) and test it with the procedure from lesson 6.
Done when: the "Decision of the week" line has one tested change, with the result of the comparison.
AI agent without breaking the bank.
The explanation appears after your answer.
Question 1 of 3
Sorting 50 emails by topic every week. Which choice makes the most sense?skill
Question 2 of 3
You hit your subscription limit every Thursday. What do you look at first?
Question 3 of 3
Who do cache and batch apply to?
Choose an answer. The explanation appears after you try.