Operations
AI in customer support: where it helps, where it hurts
An honest map of what AI is actually good at in a support team, and the three places it makes things measurably worse.
The Adarna Team
Editors
The promise made to support leaders for the last three years has been some version of "AI will reduce ticket volume by 60%." The reality is messier and more useful: AI shifts the work. It does not delete it. The teams that get value from it understand that. The teams that do not, automate away the bits of the job that were already cheap, and leave their humans to handle a denser load of harder cases with no relief.
Where AI consistently helps
- Drafting first replies on routine tickets. The agent edits, the customer gets a faster reply, and the writing standard goes up because the AI starts from a known template.
- Summarising long ticket threads at handover. Saves the next-shift agent ten minutes of scrolling. Almost never wrong in a way that matters.
- Surfacing the relevant KB article inside the agent's workflow. Cuts 'I don't know, let me check' time without removing the human's judgement.
- Routing: categorising incoming tickets to the right team or queue. A well-tuned classifier outperforms keyword rules and is invisible to the customer.
- Bulk anomaly detection: spotting that 30 tickets in the last hour are about the same dashboard widget. Twenty minutes saved on every incident.
Where AI consistently hurts
Three places. Universally.
Refund and billing decisions. Account security. Emotional escalation. In all three, the cost of being wrong is high and the customer is already at the limit of their patience. Humans take longer; customers stay longer.
Refund decisions are the most obvious. An AI confidently writes "I've processed your refund" before checking eligibility, then the human has to undo it without losing the customer. Account security (locked-out users, suspected fraud) is the second. The customer needs to feel heard and certain; the AI can do neither. The third, emotional escalation, is the trickiest: the AI's "I completely understand how frustrating this must be" can feel offensive precisely because the customer knows it does not.
The integration shape that works
The integration pattern that produces gains without producing risk is this: AI produces drafts, suggestions and summaries. The human edits, decides, and presses send. The audit trail records what the AI suggested and what the human actually did. Over six months, the gap between the two is your training signal, both for the AI and for the team.
How we test for it at Adarna
Every candidate in our Stage 2 assessment is graded on, among other things, their judgement about when to use AI and when not to. We do not test whether they can prompt. That is a two-day skill. We test whether they can read an AI-generated draft and spot the policy invention, the over-promise, or the wrong customer name. The candidates who flag those issues consistently are the ones we place at clients running AI-augmented support, and they are the ones who succeed there.
AI in customer support is not a debate about tools. It is a debate about which decisions a human should sign. The companies that get that right will deliver faster, calmer customer experiences for the same headcount. The ones that do not will find that automation made everything worse, faster.