How many requests can an AI customer support agent really handle?

"How many requests can it handle?" is the wrong question, but it is the one everyone asks first. The technical answer is: practically unlimited in parallel, because there is no queue of people. The useful answer is different, and it concerns what share of requests the assistant can close well on its own — that is where the real saving lives.

Technical capacity is not the bottleneck

An assistant connected to WhatsApp, email and website chat handles dozens of simultaneous conversations without degrading, replies in seconds at any hour, and has no unmanageable Monday-morning peak. A company with 800 contacts a month and one with 8,000 use the same architecture: what changes is model consumption, not the structure.

The percentage that actually matters

In the projects I run, a well-built assistant connected to company data closes 50-75% of incoming requests on its own. That is not magic: in almost every company most contacts are a handful of repeated questions — order status, availability, pricing, opening hours, documents, returns, first-line technical support. The rest — negotiations, serious complaints, unusual cases, angry customers — must reach a person, and that path must be designed properly.

Escalation is the most important part of the project

An assistant that does not know when to stop does more damage than no assistant at all. The three minimum rules I always set: hand over immediately if the customer asks, if the tone signals frustration, or if the request touches money, contracts or sensitive data. And the handover must carry the whole conversation, so the customer never tells the same story twice.

Cost per contact

Model cost for a typical support conversation is a few cents. Across 1,000 conversations a month that is tens of euros of consumption, plus hosting and maintenance: in an SMB the total recurring cost usually sits between 50 and 250 € per month. The real cost is the initial project — connecting data, writing rules, testing — which for a serious assistant wired into business systems runs 2,500-8,000 €.

The four numbers to judge it by after one month.

• Self-resolution rate: how many conversations close without human involvement.
• First response time: before and after, over the same hours.
• Correct escalations: how often it handed over when it should have, and how often it failed to.
• Satisfaction: one question at the end of the conversation, not a survey.

When it is not worth it

If you get fewer than 150-200 contacts a month and the questions are nearly all different, the return is marginal: invest in saved replies and a better contact form instead. AI support pays off when repetition is high, not when volume is low.

If you want a realistic estimate for your own inbox — what share could close itself and what it would cost — the last 100 messages you received and a free 15-minute call are enough.