A customer messages your chatbot at half past ten on a Saturday night about a delivery that never arrived. The bot asks the right questions, pulls up the order, and confirms it dispatched three days ago. Then the conversation stalls, because the actual problem, a parcel that went missing somewhere between the warehouse and the customer’s doorstep, isn’t something any chatbot can solve on its own. What happens in the next thirty seconds usually decides whether that customer stays or starts looking elsewhere.
Most conversations about chatbots focus on what the bot should say. Just as important is knowing when it should stop talking altogether and hand the problem to a person, and how it does that handoff well. Get this wrong in either direction, holding on too long or escalating too eagerly, and the bot ends up doing more harm than good, regardless of how well it handles the questions it was actually built to answer.
Not Every Problem Belongs With a Bot
Chatbots are built to handle the predictable: the questions that repeat, the requests that follow a known pattern, the enquiries a business has answered a thousand times before. A missing parcel, a billing dispute, or a customer who is simply upset rarely fits that mould. These situations usually need judgement, a willingness to bend a policy, or an apology that sounds like it comes from a person rather than a script, none of which a chatbot can convincingly produce.
The instinct for many businesses is to keep pushing the bot to handle more, on the theory that automation should always be growing. That instinct isn’t wrong so much as incomplete. Even organisations furthest along with AI in customer service are automating a meaningful share of contact, not all of it. McKinsey’s insight on agentic AI in customer experience points to some businesses automating as much as 70 percent of customer contact, which still leaves a substantial slice that depends on a person. Getting chatbot ticket escalation right is about making sure the right conversations land with a person quickly, before a customer has spent several rounds watching a bot try and fail to help.
The Signals a Bot Should Watch For
A chatbot doesn’t need to understand exactly why a customer is upset to know that a human should take over. It just needs a short list of signals to watch for, built into the conversation flow instead of left to chance.
- The customer repeats the same question twice, or rephrases it because the first answer missed the point.
- The request involves money changing hands outside a standard process, such as a refund, compensation, or a price exception.
- The message contains frustration cues: capital letters, exclamation marks, or words like “unacceptable” or “cancelling.”
- The bot’s confidence in matching a customer’s intent falls below the threshold the business has set, instead of it guessing anyway.
- The customer asks for a person directly, at any point in the conversation, not only after the bot has already failed once.
What Data Should Travel With the Handoff
A messy handoff often causes more frustration than the original problem. Customers who have already explained their issue to a bot understandably don’t want to repeat all of it to a person, so the full conversation history, the order or account reference, and whatever the bot has already tried should travel with the ticket automatically.
This is also where data handling deserves a second look, because a chatbot transcript can contain far more personal information than a business realises: names, contact details, sometimes payment references or health information depending on the industry. The PDPC’s guidelines on personal data in AI systems are a useful starting point for deciding what gets logged, stored, and passed along automatically, and what should be stripped out or handled with extra care before a ticket is created. A support ticket should carry enough context for a human agent to help immediately. It doesn’t need to carry everything the bot has ever collected about that customer.
Designing the Handoff Itself
The moment of escalation deserves almost as much design attention as the chatbot’s opening message. A good handoff tells the customer plainly what happens next: that a person has been notified, roughly how long a reply will take, and where to find the conversation if they need to step away and come back. A poor one leaves the customer wondering whether anyone read the message at all, which undoes whatever goodwill the bot built up beforehand.
Where Chatbot and CRM Integration Fits In
Escalation rarely happens in isolation. The same underlying question, whether a conversation should move to a person now or continue with the bot a little longer, comes up constantly in sales as well as support, and the systems a business already has in place shape how well that handoff works in practice. Businesses that have already put work into connecting your chatbot and CRM for lead handling tend to find support escalation easier to get right too, because the ticketing tool and the customer record are drawing from the same clean source of information instead of two disconnected exports.
The reverse is also true. A business that has never connected its various customer-facing tools usually discovers the gap at the worst possible moment, when a frustrated customer has to explain their history to a human agent who has no visibility into what the bot already knows. Fixing that gap usually does more for the customer experience than any single improvement to the bot’s wording.
None of this requires an elaborate rebuild. Most Singapore businesses can make real progress with a few small fixes: give the bot a clear list of moments to step back, make sure a handoff carries the right context without carrying too much, close the loop with a message that tells the customer what happens next, and keep reviewing transcripts to catch the escalations that got missed. Get those pieces right, and escalation stops looking like a failure of the chatbot and starts looking like exactly what good design predicted it would do. If you would like a second opinion on how your own chatbot decides when to step back, feel free to reach out to us and we can walk through it together.