At 11:40 on a Tuesday night, someone in Leeds is trying to work out whether their order shipped, and the answer is sitting in a warehouse system nobody on the support team can check until 9am. That gap, the eleven hours where a question exists but no human is around to answer it, is where an AI customer support chatbot earns its money.
It’s also where a lot of companies get burned. Deploy a bot badly and you’ve built a machine that confidently tells customers the wrong returns window for eight hours a day. The difference between those two outcomes isn’t the model. It’s what you point it at, and what you let it touch.
Where a support chatbot actually pulls its weight
Roughly 60 to 80 percent of the tickets landing in a typical ecommerce or SaaS inbox are variations on a handful of questions. Where’s my order. How do I reset my password. What’s your refund policy. Do you ship to Ireland. Can I change the plan I bought yesterday.
None of those need judgement. They need an accurate answer, quickly, at whatever hour the customer happens to ask. This is the work a well-configured chatbot should own outright:
- Order and ticket status lookups wired into your actual fulfilment or helpdesk system, not a canned “we’ll get back to you”
- Policy questions answered from a single source of truth you keep updated
- Password resets, plan upgrades, address changes that are really just form submissions behind a conversation
- Pre-sales questions about pricing, compatibility or stock levels, asked at 2am by someone with a credit card out
- Triage, gathering the five details a human agent would need so the handoff starts halfway done
The measurable win here is real. Companies that get this right typically see 30 to 50 percent of inbound conversations resolved without a human touching them, and the average first-response time drops from hours to seconds. If you’re weighing up whether the whole thing is worth the effort, our breakdown of what an AI chatbot for your website genuinely earns its keep on is a useful reality check before you commit budget.
Where it falls apart
The failures are rarely dramatic. They’re quiet and repetitive, which makes them worse.
Multi-part problems
A customer writes: “I ordered two items, one arrived damaged, I want a refund on that one but keep the other, and also the discount code didn’t apply.” That’s three tickets in one message, two of which need account access. A bot that handles the first clause and ignores the rest creates more work than it saved, because now an agent has to unpick a half-finished conversation.
Confident wrong answers
Ask a language model about your returns policy without grounding it in your actual documentation and it will invent something plausible. Customers screenshot those answers. Then you’re honouring a 90-day return window you never offered, or arguing with someone on Twitter about it.
Anything emotional
Someone whose flight was cancelled or whose delivery never arrived is not looking for a knowledge base article. They want to feel heard, and a cheerful “I understand your frustration! Here are some helpful links!” makes it worse. Anger detection is improving, but you should assume the bot cannot handle this and route it away fast.
The handoff is the whole ballgame
Most complaints about support bots aren’t about the bot. They’re about being trapped. Being asked to re-type your order number for a human who can already see the transcript. Being told “let me transfer you” and then disconnected.
Get these three things right and the ceiling on customer patience rises considerably:
- The agent receives the full transcript plus any structured data the bot collected, so nobody repeats themselves
- Escalation triggers are explicit: third repeat of the same question, mention of a refund above a set value, words like “lawyer” or “chargeback”, clear frustration signals
- The bot tells the customer it’s handing over, instead of pretending it can still help
A clean handoff also protects your team. Agents who spend their day cleaning up bot messes burn out faster than agents doing straight support, and that shows up in your attrition numbers long before it shows up in CSAT.
The numbers that actually tell you something
Containment rate is the metric every vendor puts on the slide, and it’s the easiest one to game. A bot that says “was this helpful?” and counts a click as a resolution will report beautiful numbers and solve nothing.
Track these instead:
- Repeat contact rate within seven days on bot-resolved conversations. If people come back with the same issue, it wasn’t resolved.
- CSAT split by bot-only versus bot-assisted. The gap tells you more than the average.
- Escalation precision. What share of conversations the bot handed over actually needed a human? Too high and it’s wasting agent time. Too low and it’s trapping people.
- Top ten unanswered questions, reviewed weekly. This is your content roadmap, handed to you for free.
Choosing something you won’t regret in six months
The market is crowded and the demos all look the same. What separates a tool you’ll keep from one you’ll be ripping out by spring usually comes down to boring things: whether it connects to your helpdesk without a middleware project, whether you can edit its answers without a support ticket to the vendor, and whether the pricing survives a busy month.
Our guide to picking an AI chatbot builder you won’t want to replace in six months covers the questions worth asking before you sign anything. The short version: test it on your ten messiest real tickets, not the demo script.
The cost maths, briefly
A support agent costs somewhere between £28,000 and £38,000 a year once you’ve added employer costs, plus the recruitment and training churn that comes with a queue nobody enjoys. A chatbot platform typically runs £200 to £2,000 a month depending on volume and integrations.
That gap is why the business case looks easy. It’s also why people over-buy. You don’t need the enterprise tier with voice and eleven languages if 80 percent of your traffic is UK-based and asking about delivery times. If you’re a small team working out where AI genuinely saves hours rather than just sounding impressive, our piece on which AI tools actually earn their place in a small-business stack is worth twenty minutes.
Roll it out in shadow mode first
The single most useful thing you can do is run the bot in shadow mode for two weeks. Let it draft answers to live conversations, but have a human review and send them. You’ll learn three things fast: which questions it nails, which ones it hallucinates, and which ones it shouldn’t be touching at all.
After that, open it to 10 percent of traffic. Not 100. Watch the transcripts daily for the first week, then weekly. Keep a running list of every answer you had to correct, and feed those corrections back into the knowledge base the same week. Most of the accuracy problems people blame on the technology are really stale documentation, and stale documentation is fixable in an afternoon.
Then widen it. Add one new capability at a time, whether that’s order lookups, then returns, then plan changes. Each addition gets its own two-week review. It’s slower than switching everything on at once, and it’s the difference between a support bot your customers stop noticing and one they complain about to their friends.