Customer service and AI: what you can automate and what should stay human

The best candidates are questions that come up again and again and have a stable answer: opening hours, address, areas served, usual lead times, documents to provide, the steps of a procedure. An AI that has this information can answer consistently, at any hour.

Next come intake tasks: collecting the name, number and reason for the request, offering an appointment slot, directing the person to the right contact, summarising the exchange for the team. This groundwork saves advisers time on every case they pick up.

An AI must not make decisions that commit the business: refunds, goodwill gestures, exceptions, legal or medical answers. Nor should it improvise: if the information isn’t in its knowledge base, the right response is to say so and offer a callback or a transfer.

Highly emotional situations (a very unhappy customer, a bereavement, an emergency) must be handed over to a person as soon as they are detected. Finally, an AI must never pass itself off as a person: a caller who finds out partway through may lose trust in the business.

On the phone, the exchange is spoken and in real time: the AI must understand spontaneous speech, reply in short sentences and cope with interruptions. Waiting weighs particularly heavily here, because the caller is stuck on the line while they hold.

Website chat is written and near-instant: it lends itself to answers with links and to exchanges while the visitor browses. Email is asynchronous: there, AI is mainly used to sort messages, suggest draft replies and fill in missing information. A single knowledge base can feed all three channels, provided the form of the answers is adapted to each.

Start with a scope where the gain is obvious and the risk low: out-of-hours calls, overflow when the team is busy, or practical questions. Then write the knowledge base from customers’ real questions, as they appear in the emails and messages you receive.

Define the escalation rules before going live: which topics go straight to a person, to which number, at what times. During the first few weeks, review the transcripts to spot unanswered questions and fill the gaps in the knowledge base; only widen the scope once this first level is stable.

Automating doesn’t exempt you from customer service rules. In France, the DGCCRF (the French consumer protection authority) points out that businesses that have chosen the phone as a way for consumer customers to contact them must provide a number that is not premium-rate for contract follow-up and complaints handling; a breach can lead to an administrative fine of up to €15,000 for a legal entity.

If calls are recorded, the caller must be informed at the start of the conversation, together with the purpose and their right to object, as described by the CNIL, France’s data protection authority, in its guidance sheet on call recording (published in 2009, flagged as outdated and under revision).

The phone is often an overloaded channel. Alyas lets you start on this channel: the business forwards its calls to the agent’s line from its operator. The agent answers 24/7, says it is an AI and that the call is recorded, answers from the information provided (website, text, PDF), captures contacts, books appointments via Google Calendar or Cal.com and transfers to a configured number; a chat widget uses the same information on the website. Data hosted in the EU, 10-day free trial.