A person vs. an AI agent for answering messages: costs and limits

Why the salary comparison misleads

Customer service is more than writing replies. A person notices tension, negotiates, grants authorised favours, coordinates with colleagues and takes responsibility. An agent processes language and runs instructions inside a system. Comparing a full salary with the cost of software assumes both cover the same function, and that is almost never true.

The real alternative is usually hybrid: the agent handles first contact, answers what is stable, gathers details and spots intent; a person receives prepared cases and focuses on decisions, sales or exceptions. Savings, if any, come from changing how work is distributed, not from deleting the team from the org chart.

Which costs to put on both sides

For the human process, record shifts, training, cover for breaks, tools, supervision, turnover and the cost of late replies. For the agent, include discovery, build, messaging, the model, hosting, monitoring, updates and human review time. Do not assume the second one runs alone.

Also measure quality. A fast but wrong reply destroys trust. An excellent human reply that arrives tomorrow can lose the prospect. A useful indicator combines time to first response, the share of cases resolved with evidence, correct handoffs and opportunities the team could continue.

Tasks worth splitting

AI can cover first availability, frequent questions, gathering requirements, sorting and summarising. The person keeps discounts, conflicts, special requests, sensitive information and decisions with financial impact. If the business does not yet know where that line is, the first weeks should be assisted: the agent proposes and a person approves.

The handoff needs design. “I’ll pass you to an advisor” is not enough if it makes the customer repeat everything. The team should receive who is writing, what they want, what was answered, what is missing and why it was escalated. The customer should also have a clear expectation of when a person will continue.

• AI: speed, repetition, search and summaries
• Person: judgment, empathy, negotiation and responsibility
• System: permissions, logging, alerts and continuity

How to test without risking service

Start with a limited schedule or channel, or with the agent in draft mode. Review conversations every day and classify errors: missing knowledge, wrong instruction, integration, ambiguity or out-of-scope case. Fixing by category avoids isolated patches.

Define a pause condition. If the agent gives unauthorised information, tries to confirm something without checking or fails to escalate a sensitive case, it must be withdrawable at once. A good implementation increases the team’s capacity without hiding its limits; it does not need to pretend it replaced a person.