Customer Service Chatbots: What to Automate, What to Escalate, What to Never Answer
Most chatbot advice is about setup. Almost none of it is about restraint. A support chatbot earns its place by closing the boring questions completely and getting out of the way fast on everything else. Here is how we would draw that line.
The short version
- Sort your questions before you configure anything. Three buckets, one rule each.
- Escalation is the feature, not the fallback. Five signals should trigger it.
- A bot that answers everything is answering things wrongly.
- Disclosure: EzyConn is our product. The advice below works on any tool.
What a support chatbot is actually for
Deflection is the word the industry uses, and it sounds worse than it is. It does not mean pushing customers away. It means a customer got a complete, correct answer without needing a person, and left satisfied. If they got an answer and then messaged again, or gave up, that is not deflection. That is a failure with better metrics.
The reason this matters for small teams is timing. Most enquiries do not arrive at 11am on a Tuesday when someone is free. In our scan of 4,278 small business websites only 7.5% had any chat at all, and among those that did, more than half were a plain WhatsApp button pointing at somebody's personal phone. The competition for a support chatbot is not another chatbot. It is a contact form nobody reads until Monday.
Sort your questions into three buckets first
Before touching any settings, read your last hundred conversations and put each one in a bucket. This takes an hour and saves a month of tuning.
| Bucket | Looks like | Rule |
|---|---|---|
| Repeat questions | Opening hours, delivery times, returns policy, do you take card | Bot closes it |
| Account questions | Where is my order, can I change my booking, what did I pay | Bot gathers, human confirms |
| Judgement calls | Complaints, refunds outside policy, anything unusual | Human, immediately |
Only the first bucket should be closed by the chatbot on its own. The second is where most of the value hides: the bot collects the order number, the booking reference and the actual problem, so the human who picks it up starts with everything they need instead of a message saying "hi, are you there?". The third bucket should reach a person as fast as the software allows.
Escalation is the feature, not the fallback
Teams get this backwards. They tune the answers for weeks and treat handover as the thing that happens when tuning fails. It is the other way round. Customers forgive a bot that says it does not know. They do not forgive one that traps them.
| Signal | What should happen |
|---|---|
| The customer asks for a person | Hand over at once. Never argue, never offer another article. |
| The tone turns angry or upset | Stop answering, apologise once, pass it to a human with the transcript. |
| The same question returns reworded | Treat the second attempt as a failure and escalate. |
| The answer needs verified account data | Collect the order or booking reference, then hand over. |
| Money or legal consequences appear | Escalate without attempting an answer. |
The second signal on that list is the one people forget to configure. If a customer rephrases the same question, the first answer did not land. Answering again in a slightly different way is how a chat turns into an argument. Escalate on the reword and you convert an irritated customer into a rescued one.
Tone: write it like your best agent on a calm day
Four rules cover almost everything. Be short: two or three sentences beats a wall of policy text. Be direct: answer the question in the first line, then add the detail. Say you do not know when you do not know, because a hedge reads as a dodge. And never be cheerful about bad news, because "Great question!" above a refund refusal is how you end up screenshotted.
One more, on honesty. Tell people they are talking to an assistant. Not in a banner nobody reads, in the first message. Customers are fine with a bot that says so and then helps. They are not fine with discovering it three messages in, and in several markets being upfront is heading towards a legal expectation rather than a courtesy.
What it should never answer
The test is simple. If a confident wrong answer would cost you money, trust or a legal problem, the chatbot does not get to try.
| Never answer | Why |
|---|---|
| Medical, legal or financial advice | A confident wrong answer here causes real harm |
| Discounts you have not published | Every invented discount becomes a promise you must honour |
| Delivery or refund promises outside policy | It creates an expectation your team then has to break |
| Anything needing identity verification | The bot cannot prove who it is talking to |
| An open complaint or dispute | The customer has already asked for a person, in effect |
Blocking a topic does not mean going silent. Silence reads as a broken widget. The right behaviour is one honest sentence and a handover: tell the customer a colleague will pick this up, say roughly when, and pass the transcript across so they never repeat themselves.
Where the human actually picks it up
An escalation is only as good as the place it lands. If it lands in a dashboard nobody has open, your average response time is however long it takes someone to remember the tab exists.
This is why we built EzyConn so agents can reply from Slack, rather than being notified and sent somewhere else. The escalated conversation arrives in the app your team already has open, with the full history attached. You can see how the handover works on our live chat feature page.
What this costs, and why the model matters more than the price
Billing shapes behaviour. If you pay per seat, you think twice about adding weekend cover. If you pay per AI resolution, every successful deflection raises the invoice, which is a strange thing to optimise against. If your AI runs on a credit balance, it can stop working in the third week of the month.
We chose flat plans with a published conversation cap, so you can see the ceiling before you buy. GPT-4o and Claude are included on every plan, free one included.
| Plan | Price | AI conversations/mo | Seats |
|---|---|---|---|
| Free | $0 forever | 100 | 2 |
| Lite | $19/mo | 200 | 2+ |
| Starter | $29/mo | 250 | 2+ |
| Basic | $59/mo | 1,000 | Up to 10 |
| Pro | $89/mo | 5,000 | Up to 10 |
| Business | $189/mo | 10,000 | Up to 10 |
Annual billing takes 20% off. Full detail sits on the pricing page, and if you are weighing us against a specific tool, the comparison pages list what each rival does better than us as well as what it costs.
Frequently asked questions
What is a customer service chatbot?
It is software that reads a customer question in your chat widget and answers it from your own content, such as your help pages, policies and product information. A modern one is not a decision tree with buttons. It reads the question in plain language and writes a plain-language answer. Its job is to close the easy questions completely and hand the rest to a person with the conversation already attached, so nobody has to ask the customer to repeat themselves.
What deflection rate should a customer service chatbot achieve?
Sort your last hundred conversations into repeat questions, account-specific questions and judgement calls first. The repeat questions are the only ones a chatbot should be closing on its own, and in most small teams they are somewhere between a third and two-thirds of the total. Aim to deflect that bucket well rather than to hit a headline percentage. A bot that answers 90% of everything is almost always answering things it should have escalated.
When should a customer service chatbot hand over to a human?
On five signals: the customer asks for a person, the customer sounds angry or upset, the same question comes back a second time in different words, the answer would need account data the bot cannot verify, or money and legal consequences are involved. Handover should be one step, carry the full transcript, and never restart the conversation. If a customer has to explain the problem twice, the handover has failed even if the routing worked.
What should a customer service chatbot never answer?
Anything where a confident wrong answer is expensive. That means medical, legal and financial advice, one-off pricing or discounts you have not published, promises about delivery dates or refunds outside your written policy, anything requiring identity verification, and anything about an ongoing complaint or dispute. The correct behaviour is not silence. It is a short, honest sentence saying a colleague will pick this up, followed by an actual handover.
How much does a customer service chatbot cost?
It depends entirely on the billing model, which matters more than the sticker price. Per-seat plans get more expensive as you add cover. Per-resolution billing charges you more the better the AI performs. Credit balances stop the AI mid-month when they run out. EzyConn, which is our product, uses flat plans with an included conversation cap: free forever at 500 messages a month, then $19, $29, $59, $89 and $189 a month, with 20% off annually.
Put the boring questions on autopilot
500 messages a month, 2 seats, and escalations that land in Microsoft Teams, Slack or your phone. Free forever, no card needed.
Start freeDisclosure: EzyConn is our own product, so treat the pricing section as a pitch and the rest as method that works on any tool. Adoption figures come from our own August 2026 scan of 4,278 small business websites across seven countries.