Making an unused AI tool work for customer service
A watch house had bought an AI tool for its customer service, yet nobody used it day to day. We made it useful by starting with organization, before technology.
The company
A Swiss high-end watch house whose customer service team of 15 agents serves a demanding, international clientele.
The problem and its cost
The team handles repair follow-ups, warranty, servicing, authenticity and boutique order requests. The AI tool had been bought with no defined use case, no reliable answer base and no named owner. Agents kept searching for information in old emails, shared files or with colleagues, then rewrote by hand messages that were nearly identical.
This cost appears in no budget, but it is real. It is calculated as follows: 15 × 360 CHF in licences + 15 × 3 h × 46 × 70 CHF of lost time, or about CHF 150,000 per year. The licences are paid for nothing, and three hours per week per agent go into searching and rewording, at the expense of the clients who need the most attention.
The strategy
We decided not to change the tool. The problem lay in what it was asked to do and what it was given to read. We began by clarifying request types, reply standards and escalation rules, because an AI connected to vague content produces vague answers.
We then built the reference base and configured the existing tool on the three or four most frequent request categories. Training and steering came last, once agents had good reasons to trust the tool. This order protects adoption: each step delivers a visible result before the next one starts.
The execution
Team: We mobilised a dedicated contact, profiles from our core team in organization, data and AI, and a sector expert with experience in watch after-sales service.
- →A map of request types, each with an owner, a target response time and an escalation rule.
- →A single answer base, validated with the after-sales workshop, covering repair times, warranty terms and care advice.
- →Reply templates and a tone guide aligned with the house's image, configured in the tool already in place.
- →A weekly dashboard, built in existing tools, tracking response time, template reuse and escalations.
The AI drafts a first reply from the request and the validated base, then summarizes long threads and sorts incoming requests. The agent reviews, adjusts and stays in control of every message sent.
The gain comes first from simple rules: one owner per request type, written standards and a monthly review of the base. Without that discipline, the tool would fall back into disuse within weeks.
The result
Expected annual saving: 65% of CHF 150,000, about CHF 98,000. For an investment of about CHF 30,000, payback comes in 4 months.
The gain rate is 65% rather than 100% because part of the time is irreducible (complex cases, high-touch exchanges with loyal clients), adoption is gradual and the licences remain payable, which gives about CHF 98,000 in annual savings and a payback of roughly 4 months.
- A more consistent tone from one agent to the next, true to the house's image.
- More time for the clients and situations that call for judgement and nuance.
- Faster onboarding for new agents and seasonal reinforcements.
The assumptions
- 15 customer service agents.
- Annual licence of CHF 360 per agent.
- 3 recoverable hours per week per agent, over 46 weeks.
- Loaded hourly cost of CHF 70; investment of about CHF 30,000, one-off.
Scenario type, built from real missions and anonymized.
In the same cocoon
