Alex Mead

THE HANDOVER RULE

When should AI hand a customer conversation to a person?

AI should hand over the moment a conversation stops being informational and starts being consequential: money is moving, a commitment is being made, judgement is needed, or the customer asks for a person. Hand over with the whole conversation attached, so the human opens on full context instead of asking the customer to start again.

Answered by Alex Mead. The dated record behind this answer.

How it is actually done

The operating line I run is short: AI first, human always one tap away. That is not a slogan about being nice. It is a design rule with four named triggers, and every one of them is checked on every turn of the conversation.

Trigger one: consequence

Money is moving, a booking is being changed, a commitment is being made on the company's behalf, or the customer is about to spend meaningfully more than they planned. Automation is excellent at the informational half of a conversation and should never be the last thing standing before a consequence.

Trigger two: judgement

The rules have to be interpreted rather than applied. Someone has to decide whether to make an exception, and to own that decision afterwards. A machine that makes exceptions is not being helpful, it is being unmanaged.

Trigger three: the customer asks

This one has no threshold and no clever exception. If a customer asks for a person they get a person, immediately, first time, without having to ask twice or type a keyword. Every hour spent tuning a model to talk a customer out of that request is an hour spent making the brand worse.

Trigger four: the read is not confident

The system knows how well it understood. When confidence drops, or the sentiment in the customer's language turns, the correct behaviour is to stop answering and start routing. An AI that is confidently wrong costs more than an AI that hands over early.

The standard a handover has to meet

The trigger is the easy half. The handover is where operations actually fail, so it gets a written standard.

  • The whole conversation goes with the customer, not a summary of it, and not a transcript the person has to read while the customer waits.
  • No repeated identification. If the AI knew who this was, the human knows too.
  • The person opens on context: who this is, what they own, what has already been tried, and what to do next.
  • The route back is built as well. When the consequential part is finished, the administrative tail can go back to automation rather than occupying a skilled person.
  • The handover is measured. The number that matters is how many customers had to explain themselves twice, and it should be zero.
D360 BANK, APRIL 2021Four fixed channels, no repeated identification, and the right to track, close and reopen any case, written into the operating doctrine of a licensed bank while the bank was still being certified to exist.

The question executives usually ask me is how much of this AI can take. That is the wrong question and it produces the wrong build. The right question is which moments in your operation must never be finished by a machine, and then everything else is fair game.

What most companies get wrong about this

  1. They set the handover threshold to protect the deflection number rather than to protect the customer, so the escape hatch is deliberately hard to find.
  2. They treat handover as a failure of the AI. Because it is framed as a failure it gets designed last, built cheaply, and it is the part every customer remembers.
  3. They pass the transcript and not the state. The person receives a wall of text and asks the customer to summarise it, which is the exact moment the customer decides the company is not serious.
  4. They make the customer ask three times. Once in the bot, once in the menu, once to the person who finally answers.
  5. They never build the route back, so a skilled human ends up doing twenty minutes of administration that automation should have taken, and the cost case for the whole programme quietly disappears.

Related questions, and the proof behind the answers

What is AI orchestration in customer service, and how is it different from deploying a chatbot?Read the answer How do you measure customer experience so it predicts revenue rather than reporting the past?Read the answer How do you cut the cost of a customer operation without damaging service?Read the answer The luxury cruise operating model, driven by youProof, not description
About the author

Alex Mead is a customer contact, service and experience executive who runs those operations as revenue engines and builds the AI that runs them. Twenty five years, six sectors and thirty countries, leading teams from 75 to more than 4,000 people, with full operating accountability rather than an advisory seat. He is available now for a permanent or interim Chief Customer Officer, VP, Director or Head of Customer Contact, Service and Experience role.

Read the CVGo to the homepageStart a conversation