Eight rules from eight days of AI in production
Budgets, approvals, access, outages and silent model changes: eight days of notes on one page, one rule each, with a link to the full note.
Abhijit Chatterjee Chattopadhyay

The last eight days were about what breaks once AI leaves the demo. Here is each note as one rule, with a link if you want the full version.
The eight rules
- Cap, wall and log any agent that pays. Limits live in the payment layer, not in the prompt. Your AI agent can now pay
- Price the task, not the token. Cheaper tokens came with bigger bills. Cheap tokens, expensive bill
- Approve by risk, not everything. A reviewer who checks everything stops catching what matters. Approving everything means catching nothing
- Narrow, scored and high-volume? Test a smaller model. The small model that beat the LLM
- Give each agent only what its job needs. Then check what came bundled. Your AI agent has more access than you think
- Decide what users see when the provider is down. Degrade, switch, time out smart. Your AI provider will go down
- Keep a test set. When the scores move and you changed nothing, the model did. Your model changed and nobody told you
- Measure whether it works every time, not on average. Right three times out of four is not reliable
What ties them together
Almost none of these fixes is a better model. They are limits, tests and routes you put around the model. That is the part you control, and the part that decides whether it still works on Monday.
Your question for the week
Which of the eight would fail first in your setup?
Your next step
I write one practical note like this every day. Sign up just below to get them by email.
Drafted with AI help, edited and approved by me.