Customer service AI: listen to every call before you automate

When a company wants to bring artificial intelligence into its customer service, the first idea is almost always the same: a conversational assistant that answers in place of the advisers. That is usually the wrong first step. Not because the tool is bad, but because you end up automating a flow of requests whose composition and causes you do not understand. Here is the reverse order, the one we apply: measure everything, understand why customers are calling, then automate what can be automated, and only that.

What you don’t know about your own customer service

In most contact centres, quality is assessed by listening to a handful of randomly chosen calls. In practice, that amounts to less than two per cent of the flow. The remaining ninety-eight per cent is never listened to, never classified, never counted.

The consequences are concrete. You know how many calls come in, but not why. You know the first-contact resolution rate, but not what keeps it from rising. And when an irritant appears, a portal that stops working, a procedure that is poorly explained, an ambiguous letter, it takes weeks for it to travel from the advisers up to someone who can act on it.

Putting a conversational assistant on top of this flow amounts to automating blind. It will answer the simple questions correctly, which were never the problem, and fail on the ones that generate repeat calls and dissatisfaction, because nobody has identified them.

Analyse the whole flow, not a sample

What has changed with language models is the cost of listening. Transcribing and classifying a call now costs a few cents. It has therefore become reasonable to analyse the whole of the exchanges rather than a survey.

We did this for the customer relations centre of a large public transport group: forty-five thousand conversations over two quarters, around two thousand hours of audio. And for a quality measurement platform, where we went from under two per cent to the whole of the calls being audited. An SME’s volumes are more modest, which makes the exercise simpler, not less useful: a few hundred calls or emails a month are enough to bring the dominant patterns to the surface.

Three rules of method

Anonymise before processing. Recordings are renamed with a technical identifier and stripped of personal data before ever passing through a model. This is a condition, not an option, and it is settled at the same time as the choice of supplier and hosting region. We set this out in the guide on generative AI and confidential data.

Classify against a closed reference list, never in free text. Each exchange is given a two-level reason code, a customer segment, a resolution status, a satisfaction level, a repeat-contact flag and a friction point. Fixed categories, comparable from one call to the next and capable of being aggregated. A free-text summary of each call is worthless: you cannot add up forty-five thousand paragraphs.

Cross severity with volume. A rare but badly experienced issue counts for as much as a frequent, mild one. The final ranking of problems combines the two, and each group is then checked against a targeted read of real transcripts. Every conclusion carries an explicit confidence level, so that management knows what it is relying on.

What the analysis generally reveals

The most common result is a reversal of the diagnosis. You thought you had an adviser training problem, and you discover a systems problem.

In the public transport group’s audit, fifteen groups of problems traced back to five root causes: case statuses invisible to the customer, who therefore calls to find out where their request stands, fragmented online portals, weak authentication. None of these causes can be fixed by improving what the advisers say. All of them are fixed upstream.

That is an important lesson before automating anything. A significant share of calls do not need a faster answer, they need not to exist. A customer who can see the state of their case online does not call to ask about it, and no conversational assistant will ever match a call that never happens.

What you can automate next

Once the flow is known, the question of automation rests on solid ground. The audit qualified twenty-three per cent of the flow as automatable, and it is the precise list of those reasons that has value, not the percentage.

Three uses come up repeatedly, in order of ease.

Feedback to advisers. Once every call is classified and scored on a few axes, greeting, empathy, resolution, closing, each adviser can receive individual feedback comparing them against the team’s averages. This is the first benefit, because it improves quality without changing anything on the customer journey.

Assistance for advisers. An agent that finds similar cases already resolved and proposes an answer, with its sources and a confidence score. The adviser decides. For the technical support desk of a software publisher, we grounded these suggestions in the product’s actual code, which took search accuracy from around sixty to ninety-five per cent. Without this grounding, the agent makes things up, and an agent that makes things up is worse than useless.

Customer self-service. This is the use everyone wants to start with, and it is the one that must come last. It only works if it answers strictly from up-to-date documentation, and if it knows its limits. The prior analysis, incidentally, provides valuable material for building it: the customers’ real vocabulary, the words they actually use to describe their problem, which are almost never the words in the internal documentation.

What not to automate

A good audit also produces a list of exclusions, and it matters as much as the rest. Some requests touch on difficult personal situations, on disputes, on matters where a mistake has serious consequences for the person concerned. These flows must stay handled by a human, and saying so explicitly stops an automation project from absorbing them out of convenience.

The general rule is the same as on all our projects: the machine prepares, sorts and proposes, the human keeps the decision whenever it commits the company or closely affects a person.

Where to start in an SME

You do not need a hundred-adviser contact centre to apply this approach. Three steps are enough.

1. Gather a month of exchanges. Call recordings if you have them, otherwise the support inbox or tickets. Volume matters less than completeness over the period.
2. Classify them against a simple reference list. Around ten reasons, a resolution status, a repeat-contact flag. It is a few days’ work with a properly framed language model.
3. Decide from the numbers. What gets fixed upstream, what gets tooled up for advisers, what can go to self-service, and what must not move.

This is exactly what the customer support lever covers, and it is a good candidate for a diagnostic: in eight days, you know why your customers are contacting you, which of those reasons are the wrong ones, and where generative artificial intelligence will genuinely save time.