In a previous post, I said don't let an AI click around your bank. At my short-term rental, I let one take over a screen on purpose.
One more thing at turnover
Both TVs at my rental go back to guest mode at every turnover. That signs the last guest out of their apps and gives the next guest a clean start, ready to sign in to their own.
By hand it took about five minutes. Not hard, just clunky, and one more thing to remember during changeover.
Is there an API? No.
I asked the question I ask first. The streaming account behind the TVs doesn't have one.
So Grok Bot does it the way I would: it opens a browser, signs in to the TV account, and clicks through the settings. That's the screen-clicking I warned about with my bank. My rental bot calls it at each turnover, so I don't have to remember it at all.
Why I'm fine with it here
- Its own browser, in the cloud. It runs in Grok Bot's cloud browser, not on my computer. That bot has one web page and nothing else. It can't see my email, my bank or my files.
- No credit card on the account. There's nothing to buy.
- The worst mistake is small: a TV that isn't in guest mode. I'd catch it at the next turnover.
It's simple, and it hasn't gotten anything wrong.
Match the guardrail to the risk
Each of my previous posts set the guardrails by what's at risk.
| Post | What's at risk | The guardrail |
|---|---|---|
| Bank | Money moved or account settings changed | Read-only API; it never gets my password |
| Sprinklers | Too much or too little water for a day | Low stakes, so none needed |
| Bills | It can see everything in my inbox, not just bills | Read-only email access |
| TVs | A TV left out of guest mode until the next turnover | Its own cloud browser, one web page, no credit card |
The bigger the risk, the tighter the guardrail. The TVs get the most control because there's the least at stake.
The catch
Clicking through screens could break when the vendor moves a button. An API would still be better. Until there is one, this works.
Same question for your business
When a system has no API, ask:
- What's the worst an agent could do on that screen?
- Does it get its own browser and login, or does it use a person's computer?
- Is there anything there to buy, send or delete?
- Would you notice a mistake, and how soon?
If the answers are "not much," let it drive.
What I took away
First question I ask now: is there an API? When the answer is no, the next question is what a mistake would cost.