Last week I called my bank about my debit card. The automated system didn’t recognize my voice. I got put on hold. It took 18 minutes just to speak with someone.
(For the love of God, please stop telling me to go to the website for faster service. I’m calling because the site doesn’t let me do what I need. Do you think I want to be on the phone? I’m a millennial.)
I’d gladly hand that wait to an AI assistant. I’ve got a shiny new one. Mine’s a Dot named Beaumonster. I’d happily let it call a business and book an appointment for me, too.
Beaumonster
The person answering the phone gets a vote too. If they hear an automated voice and hang up, I still don’t have an appointment.
On September 22, Reuters reported that a Meta employee’s insurance company kept hanging up when it heard Muse was AI. Meta was testing a “human concierge” to get requests like that through: trained contractors would place some of the calls.
This was an internal test. Employees raised concerns about sensitive information reaching those contractors. A Meta executive acknowledged that starting without proper disclosure “was a miss,” and said the human-calling feature had been rolled back for now.
Meta told employees that some internal tests brought call success into the 95% to 98% range. That sounds useful. I’d still like to know when that person is getting involved, and what they’re allowed to see.
Meta has been here before. Facebook’s M assistant, shut down in 2018, reportedly had people handling around 70% of its requests.
Hidden help is tempting for whoever builds the product. If a person quietly rescues the call, the assistant looks brilliant. It feels like magic until you find out. Then it feels like a trick.
This week, OpenAI introduced Dots. Amazon has Alexa for Shopping. Meta has Muse. The demos start to blur together: an assistant goes off and does stuff for you. You get to focus on being human and presumably never sit on hold again.
What I want to see in those demos is what happens when the errand gets stuck.
Who steps in? And how much work is waiting for them?
I spent 14 years at Apple. On a manufacturing line, 95% first-pass yield can sound like you’re almost there. Five of every hundred devices still need attention. Across millions of units, “almost there” keeps entire teams busy.
You investigate the failures, figure out what can be reworked, and feed what you learn back into the process. You want fewer devices needing that help tomorrow.
AI assistants need some of that discipline. An errand that stops halfway through should be easy for somebody to pick up. If the customer has to start over, you’ve given them another chore.
The demo is the first 95%. The product is what happens to the other five.
Some of those failures will happen because the other side has rules of its own. As I wrote last week, Amazon blocked Muse from shopping in its store. A business can refuse an automated call. A bank can insist on speaking with the account holder.
I can give Beaumonster permission to act for me. The business still gets to decide what it will accept.
Back to my bank. If it will accept an assistant, Beaumonster can sit through the hold music. When the bank needs me to verify my identity, I want something like:
“Someone’s on the line about your debit card. They need you to verify your identity. Join the call?”
Let me do what the bank needs, then hand it back. If I have to read the transcript and explain why I called all over again, I’m still doing the errand.
And if the bank always needs me at the same point, the assistant should know to bring me in there next time. Same as the factory: learn from the failure so fewer errands need help tomorrow.
Dots includes rules for approvals and human takeover. I like that direction. We’ll see how well it handles the moments when a task stalls. Asking at the right time, with the right context, takes care.
So does choosing which human to involve. The employee in Reuters’ report was calling an insurance company, exactly the kind of call where a stranger shouldn’t hear more than the errand requires. I can verify my own identity with the bank. A contractor needs a narrower brief: enough information to do the errand, with clear limits on what they can agree to.
“Book an appointment” shouldn’t quietly become permission to share whatever I’ve told my assistant.
A person can make mistakes too. Someone still needs to check that the appointment was booked, that it fits what I asked for, and that I hear the result. Putting a human on the phone creates another handoff to get right.
I’m happy for Beaumonster to have help. I’d like to know who’s on the line.
AI won’t replace ranchers. But it will change who wins. DIT AgTech puts Physical AI to work alongside ranchers to produce more with less. Now building for a massive market. Be one of the first.
