This is a post about the AI agent handoff — the moment a bot decides it cannot help and passes the customer to a human. Everyone celebrates the part that comes before. Almost nobody designs the part that comes after.
The story aha tells about its WhatsApp service is one I have seen quoted four times in the last month. Sixty percent of incoming queries resolved autonomously by the bot. Average resolution time of two-and-a-half to five minutes. Five thousand queries a day. Clean number. Quotable number.
It is the other forty percent I keep thinking about.
A customer messages on WhatsApp with a problem. The bot does what bots do — checks intent, tries a template, asks a clarifying question. It cannot get there. So it hands off.
Now, here is what almost nobody writes about. Hand off to whom. With what context. Inside what response window. With whose name attached.
In most Indian SMB deployments I have seen up close, the answers are: to a shared inbox, with no context except the chat log, inside a window that depends on whether someone happens to be at their desk, signed by the brand. Which is to say — to nobody, with nothing, eventually, from a stranger.
This is not an AI problem. This is an operations problem dressed up as an AI problem.
The AI agent handoff is where bharosa breaks
The promise the bot made to the customer is implicit. I am the front door. If I cannot help, someone better will. The forty percent the bot cannot resolve are not the easy queries. They are, by definition, the harder ones. Higher emotional stakes. More money involved. More likely to leave if they do not get answered.
And those are the ones that get dropped into a queue.
A founder I spoke to recently was proud of his automation stack. Bots on WhatsApp, on the website, in the app. Then we looked at his cancellation reasons together. The pattern was not in the bot. The pattern was in the gap that opened up after the bot.
The customer’s last memory of the brand was not the friendly automated greeting at the top of the chat. It was the four hour silence after the bot said, let me connect you with our team.
What actually has to be designed
When SMBs buy an AI agent, they tend to spend ninety percent of the procurement time on the agent itself — its languages, its tone, its training data, its dashboard. Almost none on the handoff. The handoff is the part the vendor does not sell because the vendor cannot deliver it. It depends entirely on what the team behind the bot agrees to do, on what time, with what definition of done.
Three things need to exist for the forty percent to land softly.
The first is a named owner per channel, per shift. Not a team. A name. The customer’s escalated message lands with somebody who, when asked who is responsible, says I am.
The second is a context carry-over the human can read in twenty seconds. Not the full chat log. A two-line summary the bot generates as it hands off — what the customer wants, what has been tried, what the bot suspects is going wrong. This is a workflow design decision, not a feature.
The third is a hard response-time target with a real consequence inside the team. Not an SLA on a slide. An actual rule. If the response goes beyond X minutes, Y happens. The number can be anything you can defend. The point is that there is one, and the team knows it.
None of this is glamorous. None of it shows up in the demo. All of it decides whether the agent you paid for is worth paying for.
The honest number is not sixty percent
The number aha quotes is real and impressive. But the number that decides whether your customers come back is a different one. It is what percentage of the escalated forty percent gets a human response inside the window where the customer was still paying attention.
I have never seen an SMB measure that number. I have rarely seen one even define it.
The reason agents disappoint is not that the technology is weak. It is that the work around the technology — the rules, the names, the windows, the consequences — was never built. The team treated the agent as if it were a product, when what it actually is is the front door of a process the team still owns.
The agent does not fail. The handoff does. And the handoff is yours.