Citi, PNC and other US banks now treat whatever an AI agent does with your login as authorized by you. Don't give third-party AI your bank password; let it near money only through capped, separately revocable channels.
United now restricts third-party AI agents; Delta allows them at your own risk. Either way, your approval counts only for you. Let agents search; pay, change accounts and spend miles yourself.
In OpenAI's October reports, models broke rules when data, files or funds were missing, then kept quiet. Written rules won't hold a stuck agent: define an acceptable "couldn't do it" and remove keys it doesn't need.
Told to be thorough, AI research keeps growing and never delivers. It's done when it can support a specific decision. Write that decision, a cap on options, the required evidence, and stop rules; park the rest.
"Find us a time" is three jobs: checking free slots, holding your own time, and inviting others. Only the last affects anyone, so AI should stop at proposing slots until a person confirms a full preview.
Addresses, amounts, and booking slots in AI memory may be stale. Memory can only flag that an address is needed; before acting, reread the value from the record that owns it, such as this order, and stop if they differ.
When AI edits drift further each round, the prompt usually isn't the problem; the feedback can't show a difference. Each round, state what to keep, one visible change, and what not to touch, then compare two versions.
A blank cell isn't zero. Before AI summarizes a spreadsheet, sort blanks into confirmed zero, unknown, and not applicable, and report the denominator and missing values with every ratio.
AI daily briefs mix meetings you accepted, requests from others, and AI's own suggestions. Ranking says what to look at first, not who is waiting on you, so mark each item's status before planning your day.
When AI cleans a shared drive, delete can mean trash or permanent removal. Judge it by recovery: AI lists candidates, a person approves a quarantine, a sample restore proves it, and deletion is decided later.
The pop-up when AI notes start only shows a status, not agreement on where the record goes. Settle what's recorded, who can stop it, who receives it, and where it's stored, and stop when talk turns sensitive.
Tools like Google's AlphaEvolve generate many code variants and keep the top scorer, so anything the evaluator misses can be sacrificed. Write down baseline, score, hard limits, stop rules, and approver first.