We spent years selling voice bots to contact centres and watched the same thing happen on every floor: the bot took the easy calls, plateaued somewhere around two thirds, and the hard third got worse because the people who used to take it were gone. So we changed what we sell. The AI still makes the calls, and our people, on the same queue, take the ones it can't finish.
The room
The first contact-centre floor I spent a full day on had well over a thousand people in it. Headsets, three screens each, a wallboard with the queue count in red. I was there to watch our bot, which was taking the front of the inbound queue. It was doing fine. The number on the wallboard was going down. And the people on the floor were having a worse day than before we arrived, because every call that reached them was one the bot had already failed to finish, and the customer on the other end had already told the whole story once.
Nobody had designed it that way, and I'd signed the deal. The bot's dashboard showed containment going up. The floor's dashboard showed handle time going up. Both were true, and neither was the number the business cared about, which was how many of those customers renewed.
The plateau
Every voice bot we have run, ours included, plateaus. On a mature programme the AI finishes 60 to 80% of calls. It gets there over months (up to 40% by month 3, 50% by month 4, 70 to 80% by month 6 is the shape we now put in writing). The remaining 20 to 40% is the customer who disputes the amount, the one who lost a job, the one who says the premium went up, the one who wants to close. The money sits there, in both directions.
The industry's answer to the plateau has been to sell a better bot. I think that's the wrong unit. The unit is the floor: the AI and the people together, on one queue, measured by one number. If the people are a different vendor on a different queue with a different dashboard, the join between them is where the outcome leaks, and nobody owns the join.
What we changed
We stopped selling the bot alone. We put trained people on the same queue as the AI from day one of a programme, in our console or on the client's own desk, and we priced it the way the work is shaped: the AI by the minute, the people by the seat, the handover between them free, setup once. We wrote the client's current KPIs into the pilot as a floor. And we started saying the plateau number out loud in the first meeting, because a buyer who hears 60 to 80% from us before they hear 95% from someone else asks better questions.
The product is the handover, the moment the call moves from one to the other with nothing lost. The person's screen shows who this is, what was said and what was promised, and the first thing they say is about the problem. Customers don't thank you for it, they just stay on the line.
What it is called
We call it a Human + AI BPO. Some people say agentic BPO; the press has started saying agentic process outsourcing. I'm not precious about the noun. I am precious about the five tests on the category page, because they're what a head of collections can use in a meeting where everyone claims a human in the loop, and they fit on one line of a notebook.
We run about a million calls a day across forty-odd enterprises, and I still think the interesting part is the 30% the AI hands over. If you run a floor and you have a bot that plateaued, I'd like to see your number before and after the bot arrived. Send it and I'll tell you what I think it says.