"We will add a bot and free up the support team" sounds sensible until the first customer gets "Sorry, I did not understand that" three times and leaves for a competitor. A bot does not compete with a person: it takes the top layer of enquiries where the answer is identical every time, and hands over the rest promptly. Below: which requests a bot handles better, where it annoys people, how to build the handoff, and what to measure.

What a bot does better than an agent

• It answers at three in the morning. Half of retail enquiries arrive outside working hours. A bot will not close the sale overnight, but it captures the contact and answers the simple question — so the customer never reaches a competitor.
• Repetition does not wear it down. Opening hours, address, delivery terms, whether size M is in stock. An agent answering that for the fifteenth time today answers it worse; a bot at 18:40 performs as at 09:00.
• It keeps first response time flat. The gap between a reply in 5 seconds and one in 40 minutes moves conversion more than most edits to your landing page.
• It does dull lookups without slipping. Order status, tracking number, bank details, a duplicate receipt — humans make more mistakes on exactly this kind of task.
• It qualifies before the conversation. Three questions about city, volume and deadline mean the agent gets context instead of "hello". That is 3–5 minutes saved on every dialogue.

Where a bot loses money

1. 1Complaints and refunds. The person is already annoyed, and a cheerful automated reply at that moment costs you a public review and the customer permanently.
2. 2Non-standard terms. Wholesale, a negotiated discount, an awkward installation. A bot either says no or promises what the company will not deliver.
3. 3High-value purchases. The bigger the order, the more the buyer wants a human voice. A bot between a customer and a $3,000 order is a saving made at the funnel’s most expensive point.
4. 4Long descriptive messages. The customer writes a paragraph containing three questions, the bot answers the easiest one, and the customer concludes nobody read it.

The hybrid split: who handles what

The model that works is not "bot instead of human" or "bot in front of human" — it is routing by request type.

Handing the conversation to a person

This is where most implementations break. The bot gives up, calls an agent, and the customer explains everything again because the agent cannot see the earlier messages. The other version: the bot escalates out of hours and nobody arrives.

1. 1

   Detect the moment

   Two failed attempts in a row, the words "agent" or "human", emotional language, a large basket — triggers, not a reason to show the menu again.

2. 2

   Say it plainly

   "Passing you to a colleague, they will reply within 10 minutes" — and that number has to be true. Promising "right now" at 23:00 is worse than an honest "at 09:00 tomorrow".

3. 3

   Carry the context, then go quiet

   The agent sees the last 10 messages and anything the bot promised. After handover the bot stays out of the conversation until the agent gives it back.

4. 4

   Keep a fallback

   If no agent replies within the stated window, the bot takes contact details and commits to a callback. Silence in the chat is the worst possible ending.

At configuration level, escalation rules look roughly like this — a fragment from our projects, trimmed to the essentials:

{
  "escalate_when": [
    { "rule": "fallback_count >= 2", "note": "bot missed the intent twice in a row" },
    { "rule": "intent in ['refund', 'complaint', 'legal']", "note": "money and emotion go straight to a human" },
    { "rule": "text matches /agent|human|manager/i", "note": "explicit customer request" },
    { "rule": "cart_value > 500", "note": "a large basket deserves a live conversation" },
    { "rule": "session_seconds > 180", "note": "customer stuck in menus for over 3 minutes" }
  ],
  "on_escalate": {
    "notify": "telegram:support_chat",
    "attach": "last_10_messages",
    "sla_seconds": 600,
    "if_no_agent": "collect_contact_and_promise_callback"
  }
}

What to measure so you do not fool yourself

The usual way to prove a bot works is to show the number of conversations handled. That number means nothing: a bot that said "I did not understand" a thousand times also handled a thousand.

A note on language-model bots: they solve phrasing recognition but add the risk of confident invention. Source answers only from your own knowledge base and hard-block improvisation on pricing, deadlines and guarantees; the guardrails are covered in our piece on AI chatbots for business.

Frequently asked questions about bots and live support

In short

• A bot takes repeated questions with unchanging answers; complaints, non-standard terms and expensive purchases stay with people.
• A "talk to a person" route must exist at every step — without it a bot loses more than it saves.
• Handover without conversation history performs worse than an ordinary contact form.
• Measure containment rate, time to a human reply and enquiries generated from chat — never the raw conversation count.