Six rules for running AI for real
Checked answers, long runs, agent storage, pricing, privacy and customer data: six days of notes on one page, one rule each, with a link to the full note.
Abhijit Chatterjee Chattopadhyay

The last six days were about running AI for real: past the first demo, into long runs, real prices and real customer data. Here is each note as one rule, with a link if you want the full version.
The six rules
- A valid answer is not a right answer. A schema fixes the format. Your code still checks the facts. Your AI's answer can be valid and still wrong
- Make every action safe to run twice. Long runs get interrupted, and a careless restart repeats what was already done. An agent that works all day will be interrupted
- Decide what an agent keeps, who can change it and when it goes. Where your agents keep their work
- Price the result, and price for your heaviest users. Not your average one. Your AI costs grow with use. Does your price?
- If sending data out is the risk, keep the AI where the data is. The most private AI never leaves the phone
- Know where customer data goes before it goes into an AI tool. Where does your AI send your customers' data?
What ties them together
Each rule is a question someone should have answered before the AI went live. What does it check, what does it repeat, what does it keep, what does it cost, where does the data go, and should it leave at all. None of them needs a better model. They need an owner.
Your question for the week
Which of the six could nobody on your team answer today?
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.