A service agent earns its place on the narrow band of contacts that are repetitive, answerable from published content, and low-consequence when wrong. It does that band well. Outside it, the agent becomes a delay customers have to get through before they reach a person, and the deflection number goes up while satisfaction goes down. Getting the band right is the whole design problem.
What it genuinely does well
Three shapes of contact, and they are worth naming precisely because everything else is where the trouble starts.
Questions with a published answer. Delivery windows, returns policy, how to reset a password, whether a plan includes a feature. If the answer exists in knowledge and is current, the agent will find it and phrase it for the specific question asked. This is the strongest case and it is not glamorous.
Status lookups against a record. Where is my order, what is the state of my case, when does my contract renew. The agent reads a record through an action and reports it. The value here is not the language model. It is that the customer did not have to log in, navigate, and find the right tab.
Simple, reversible changes. Update a delivery address, change a contact preference, reschedule an appointment into an open slot. Reversible is the operative word. If getting it wrong means an apology rather than a refund, the agent can hold it.
Where it frustrates people
Customers do not arrive with a clean intent. They arrive annoyed, halfway through a story, describing a symptom rather than a request. An opening message of “this is the third time I’ve had to contact you” carries no routable content at all, and the agent’s attempt to classify it produces something confidently beside the point.
The failure patterns are consistent:
- The loop. The agent asks a clarifying question, the customer rephrases, the agent asks something adjacent. Three turns in, nothing has moved. This is the single biggest driver of complaints about agents, and it is a design failure rather than a model failure.
- Answering a slightly different question. Grounded retrieval pulls the closest article rather than the right one, and the answer is plausible enough that the customer follows it and comes back angrier.
- Refusing without a route. “I can’t help with that” is acceptable exactly once, and only when it is immediately followed by a handoff that works.
Deflection is not the number you want
Deflection counts conversations that ended without a human. It counts the customer who got their answer and the customer who gave up and went to social media identically. If the only metric on the dashboard is deflection, the agent will be tuned to be harder to escape, which is the opposite of the goal.
Better pairs: resolution rate against a sampled and human-read set of transcripts, and repeat-contact rate within seven days. A conversation that deflected and then produced a case the next morning did not deflect. It deferred.
Handoff is part of the product
The agent must be able to hand over at any point, and the handover must carry what was said. A customer who explains the problem twice has had a worse experience than one who never met the agent. That means routing to the right queue with the transcript attached, and it means an escape hatch the customer can trigger themselves in plain language without guessing a magic word.
A scope that tends to work
Take your top five contact reasons by volume, check which have a current knowledge article, build only those, and escalate everything else on the first turn. It feels thin. It also produces an agent that is right almost every time it speaks, which is the reputation you want before widening it.
Frequently asked
Should the agent say it is an AI? Yes, plainly, in the opening message. Some jurisdictions require disclosure and all of them make you look worse when a customer works it out mid-conversation.
Can it handle a frustrated customer? It can adopt a calmer register if you instruct it to. It cannot make a judgement call about a goodwill gesture, and it should not try. Detect frustration signals and route.
What about voice? Voice channels remove the customer’s ability to skim and reread, which makes every misunderstanding more expensive. Check current documentation for what channels are supported before you plan for one.
How much knowledge cleanup does this need? More than the project plan says. An agent surfaces stale articles that nobody had opened in two years, and it surfaces them to customers.
