An AI chatbot can answer a password question, check an order or collect a few details before a person takes over. The value is not that it replaces the agent. The value is that the agent starts with context.
Most failed chatbot projects do not fail at the first answer. They fail at the handoff.
Containment is not resolution
A containment metric counts conversations the bot handled without a human. It says nothing about whether the customer got an answer. A bot can contain a conversation by looping, guessing or refusing to escalate.
Measure resolution instead. Did the customer complete the task? Did they return with the same issue? Did the next agent have to start from scratch?
Where the handoff breaks
The first break is context. The customer explains the issue to the bot, then repeats it to the agent because the transcript is not attached to the ticket. The fix is technical: pass the conversation, the identified customer, the order reference and the bot’s attempted steps into the ticket.
The second break is timing. A bot that keeps asking clarifying questions after the customer has asked for a person feels like a barrier. The escalation rule should be visible and fast, not hidden behind three more prompts.
The third break is authority. The bot may promise a refund, a delivery date or a policy exception that the human cannot honour. Give the bot a clear list of what it may commit to, and route anything else to a person.
Designing escalation
Write the escalation rules before the conversation flows. Escalate on explicit requests, repeated failed attempts, high-value accounts, legal or safety language, payment disputes and any case where confidence is low.
Show the customer what happens next. “I am passing this to a support agent with your order details” is better than a silent transfer.
Keep a human queue for the cases the bot cannot resolve. A bot that creates more work for the same team while reporting high containment is not saving money.
Cost per contact
Compare cost per resolved contact, not cost per message. Include bot platform fees, model usage, integration maintenance and the agent time spent fixing bot mistakes.
A bot that resolves 40 percent of contacts cleanly can be excellent. A bot that handles 80 percent but sends confused customers to an agent twice is worse than no bot.
Metrics that matter
Track resolution rate, repeat contact rate, handoff time, agent rework and customer satisfaction after handoff. Segment by intent; a bot can be strong on order status and weak on billing disputes.
Review transcripts weekly. The failures reveal missing tools, unclear policies and questions the bot should never attempt.
What to automate first
Start with high-volume, low-risk intents that have a known answer and a system of record: order status, appointment changes, store hours, document requests and basic troubleshooting. Add account-specific actions only when the identity check is reliable.
Leave complaints, negotiations and anything that requires judgement with people. The goal is a faster path to the right answer, not a wall between the customer and the company.
Writing the escalation script
The escalation script is a product document, not a support macro. It should name the triggers, the information passed to the agent and the promise made to the customer.
Triggers include an explicit request for a person, two failed attempts at the same intent, a payment dispute, a legal threat or a high-value account. The bot should not wait for a third failure.
The handoff package should include the customer identity, the order or booking reference, the conversation summary, the steps already tried and the customer’s preferred language. The agent should not have to ask for any of it again.
The promise should be honest: a callback within a stated window, a ticket reference or a transfer to a live queue. Do not promise a resolution time the team cannot meet.
FAQ
Will an AI chatbot reduce support headcount?
It can reduce repetitive contacts, but the first gain is usually faster response and better context for agents. Headcount changes follow after resolution and handoff are measured.
What should the bot never do?
It should not promise refunds, exceptions or delivery dates outside a defined policy, and it should not continue a loop after the customer asks for a person.
How do we measure success?
Use resolution rate, repeat contacts, handoff time and post-handoff satisfaction. Containment alone rewards the wrong behaviour.
For the product side, see AI Chatbot and AI chatbot app development cost. If tickets are the system of record, connect the flow to ticketing.
