Same call, or a callback?

The question, and why vendors dodge it

I sat in a vendor meeting last month where the deck said "human in the loop" four times. I asked when the human gets the call. The answer took three minutes and the word callback arrived in minute two. The customer, it turned out, would get a message saying someone would ring back within the hour. A callback within the hour is a queue, whatever the slide calls it.

So the question I now ask first, before pricing, before languages, before the demo: same call, or a callback? It's a yes or no. If the answer has a dependent clause in it, it's a callback.

What a callback costs, in the customer's time

A callback means the customer explained the problem once to a machine, got told to wait, and will explain it again to a person who has a ticket number and not much else. On a renewal call the customer has already decided by the time the phone rings again. By then the promise to pay a collections customer was about to make has gone cold. On a support call they've opened the app and tried three things, and now they're angry at the person who rang back, who did nothing wrong.

Our own floor data is the reason we hold this line. The AI finishes 60 to 80% of calls on a mature programme (it starts lower; up to 40% by month 3, 50% by month 4, 70 to 80% by month 6). The other 20 to 40% are, by definition, the hard ones: the dispute, the hardship, the premium objection, the person who wants to close. Those are the calls where a callback loses the outcome. Making the easy 70% faster and the hard 30% worse is not a trade a head of collections would sign if it were put to them plainly.

What "same call" has to mean to count

Same call means the line never drops. The AI says it's bringing a colleague in, the colleague's screen already shows who this is, what was said, what the customer wants and what has been promised, and the first thing the person says is about the problem, not "can I have your registered mobile number". Three things have to be true for that to happen, and two of them are unglamorous.

• The AI and the person are on one queue and one record, so there is nothing to re-key.
• The person is on the floor from day one; nobody is hired in month four when the bot hits a wall. We call this the floor, and it's written into the pilot: nothing you measure today is allowed to drop.
• If every agent is busy, the AI says so and offers a choice (stay with the assistant, a callback at a named time, or hold) rather than dumping the call into a queue. A callback offered as a choice is fine. A callback as the only mechanism is the thing I'm arguing against.

The five tests, for the next vendor meeting

We put five tests on our own category page because buyers kept asking us what to compare on. They're short enough to use in a meeting. The person sees the transcript and context as the call arrives, which we call Handover. It's the same conversation with no re-verification, Continuity. There's a human bench from day one with your KPIs as the floor, which is the Floor test. The fee can be tied to the metric your board reads if you want it that way, Outcome. And you get a live call on your number, in your language, in the first meeting, not a slide: Proof.

A bot vendor passes none of them on their own. A traditional BPO passes Floor and fails the rest. A gig-caller model passes Proof and fails Handover and Continuity, because the person arrives later, on a different call. I don't think any of this is controversial once it's written down.

Ask it

Ask the question on your next vendor call and write the answer down in the vendor's words. If you want to hear what same-call sounds like in Hindi, the renewal cut below is one call, start to finish, with the person joining at the point the customer says the premium went up. Then ask us the same question. We'd rather you did.

Watch the renewal handover, 1 minute