Stuck talking to a bot
A customer types 'speak to a person'. The bot replies with three article links. They type 'agent'. The bot asks them to rephrase. They type 'HUMAN PLEASE'. The bot says it did not understand and offers the main menu again.
Eventually they find your phone number elsewhere, ring, and start the whole story again, now annoyed. Or they leave a review about your chatbot. Or they simply go away.
You may have seen this in complaints, or you may not know it is happening at all, because nobody reads the transcripts.
How a bot ends up as a wall
It is rarely deliberate. Chatbots are often bought or measured on how many conversations they 'deflect' from the support team. That number looks good in a report even when the deflected customer simply gave up. When the target is fewer contacts, the easiest design is a bot that makes contact hard.
Handover is also often an afterthought. It needs a staffed queue behind it, rules for out-of-hours, and a way to pass the conversation over. If that plumbing was not built, the bot has nowhere to send people, so it keeps them.
And older, intent-based bots simply fail on anything they were not trained for, which includes many ways of asking for a person.
What the loop costs you
| What happens | What it costs |
|---|---|
| Customer gives up | Lost sale or lost customer, invisible in reports |
| Customer rings instead | Longer, angrier call that starts from scratch |
| Customer posts about it | Public reviews about being stonewalled |
| Deflection figures look good | Management believes the bot is working |
The last row is why this can run for a long time unnoticed. The metric says the bot is succeeding at exactly the moment it is failing.
How SpiderHunts fixes a bot that will not let go
- Read the transcripts. We review a sample of real conversations, find the loops, dead ends and abandoned chats, and show you what customers experienced.
- Make the route to a person always visible and always working: a clear 'talk to a person' option, and recognition of the many ways people ask, from 'agent' to 'this is useless'.
- Detect frustration and failure automatically: repeated questions, the same answer given twice, negative wording, or several turns without progress. Any of these offers a person straight away.
- Build the handover properly: to live chat when staff are available, to a ticket or callback request when they are not, with an honest message about when a person will pick it up in your normal working pattern.
- Pass the full conversation and anything the bot collected, such as the order number, to the person taking over, so the customer never has to repeat themselves.
- Keep the bot to topics it handles well, and route sensitive topics, such as complaints, cancellations and vulnerable customers, to people from the start.
- Replace deflection as the main measure with resolution: did the customer get what they needed, whether from the bot or a person.
We can usually do this within your existing chatbot platform, such as Intercom, Zendesk, Tidio or LivePerson, if it exposes the right controls. If it does not, we explain what the limits are and what the alternatives would be.
A bot that knows when to step aside
Customers who want a person can reach one, on the first ask. The bot still handles the routine questions it is good at, and when it hands over, your team sees the whole conversation. Frustrated customers are caught early rather than after they have written a review.
Handover also becomes a source of improvement. Every conversation the bot passes on is a question it could not answer or a situation it should not handle. Reviewing them shows you which new topics are worth teaching it and which should always go straight to a person.
Your reports show whether customers actually got help, which is a better guide to whether the bot is earning its place.
Is your bot a wall?
- Complaints or reviews mention being unable to reach a person.
- Your chatbot is measured mainly on deflection.
- Typing 'speak to a person' does not reliably reach one.
- Staff taking over from the bot have to ask the customer to start again.
- Nobody regularly reads the bot's conversations.