The question we are asked most often is not “can AI answer my customers?”, but “how far can I let it go?”. Our answer fits in one rule: the machine handles what is verifiable, the human keeps what commits the company.
What actually saturates a customer service team
On most of the projects we take over, between 60 and 75 % of inbound messages cover five subjects: product availability, order tracking, opening hours, payment terms and the returns procedure. These requests have one thing in common: the answer already exists somewhere, in a catalogue, a dashboard or a page of the website.
They require no judgement. They require fast access to up-to-date information. That is exactly the ground on which an automated assistant is reliable — and it is the only ground on which we deploy one in the first phase.
Draw the boundary before writing a line of code
An assistant with no written boundary always ends up inventing. So we start by listing, with the teams, what the tool may do alone, what it must have validated and what it must refuse.
- ✓AloneAvailability, opening hours, order tracking, returns procedure — any information verifiable in a system.
- !With validationGoodwill gestures, discounts, deadline commitments: the assistant prepares, an adviser confirms.
- ×NeverComplaints, disputes, sensitive personal data, contractual promises.
A useful assistant says “I don’t know” as well as it answers.
Handing over to a human is not a failure
The moment the assistant passes the conversation on is the most sensitive point of the whole setup. Badly designed, the customer repeats their problem from the beginning and the automation has cost them time. Well designed, the adviser receives the history, the context and a suggested reply: they gain a minute instead of losing three.
We always trigger the handover on three signals: a complaint, a request involving money, or a second rephrasing by the customer. Everything else can take an automated answer.
Measure something other than the automation rate
The percentage of conversations handled without a human is the easiest number to display and the least useful to track on its own. We always pair it with the takeover rate, post-conversation satisfaction and first response time. An assistant that automates 80 % of exchanges but generates 30 % complaints costs more than it brings in.
Where to start
Take your last two hundred inbound messages and sort them into three piles: answer already written somewhere, answer to be built, answer that commits the company. The first pile is your starting scope. It is almost always enough to justify the project.