AI agents in customer service: where they work and where they get expensive
An agent that answers correctly 85% of the time sounds excellent until you find out what was in the other 15%.
The real jump from decision-tree bots
The button-based bot of three years ago was a form in disguise. An agent running a language model connected to your catalogue and CRM is a different category: it understands a badly typed question, cross-references live inventory and answers in the brand voice. On high-volume repetitive queries (opening hours, order status, availability, pricing) the reduction in load on the human team is genuine and measurable.
That is where the good news ends and serious design begins.
Where it works without argument
- High-volume informational queries. Anything with a single answer verifiable against a database.
- Lead qualification. Capturing budget, timeline and need before handing to a human. This lifts sales team productivity more than any email automation.
- After-hours coverage. In LATAM a high share of enquiries arrive at night and on weekends. Answering in that gap is pure competitive advantage.
Where it gets expensive
Three red zones. The first is binding commitments: special pricing, delivery windows, warranty terms. If the agent invents a discount, that discount is yours to honour. Block free generation on those fields and replace it with a lookup against a source of truth, or with straight escalation.
The second is complaints. An upset customer who realises they are talking to a machine escalates immediately. The rule we use: any complaint signal triggers transfer to a human on the first turn, with no attempt to resolve.
The third is personal data. An agent that requests and stores ID numbers, addresses or payment details without controls is a privacy incident waiting to happen. Define what it may ask for and what must never be written into the conversation log.
How to measure it properly
The vanity metric is percentage of conversations resolved without a human. The useful metric is correct escalation rate: of the conversations that should have gone to a person, how many actually did. An agent at 60% resolution with flawless escalation is worth more than one at 85% that clings to conversations it cannot handle.
And one operational rule that prevents disasters: manually review a weekly sample of closed conversations. The day you stop reading them is the day the agent starts degrading without anyone noticing.