Customer service representative — tasks, one by one

Answering repeat questions

High volume, low variance, answer already exists in writing, and a wrong answer is cheap to correct. This is the single clearest fit for language models in any service job.

Deflecting the easy half does not halve the work; it removes the easy half and leaves a queue where every remaining contact is harder than average. Handle time per contact goes up even as headcount falls.

Routing and triage

Classification with a fixed label set and abundant historical training data. The failure mode is misrouting, which is annoying but recoverable — a low bar for automation.

Misrouting is cheap only when a person is waiting at the other end to catch it. As more downstream steps also automate, the cost of a routing error stops being recoverable and starts being a customer who gives up.

Handling a customer who is already angry

The customer's goal is partly to be acknowledged by a person with authority to fix it. Automation that handles this well still tends to fail the moment the customer asks whether they are talking to a human.

This protects the role, not the headcount. If automation absorbs 70% of contacts, the remaining escalation work can be done by far fewer people.

Deciding an exception

Spending the company's money against policy is an accountability decision. Companies are slow to delegate it to a system, largely for liability and abuse reasons rather than capability ones.

The constraint here is corporate liability policy, not capability — and policy is exactly the kind of thing that changes once the savings are demonstrated elsewhere. Treat this as a delay, not a moat.

Reviewing what the bot said

Every deployed support bot creates a new need: someone who reads its transcripts and knows what good looks like. Experienced agents are the obvious people to do it.

One reviewer can audit the output of many bots, which is precisely why this task does not replace the headcount that the bots displaced. It is a smaller, more senior job.