An AI customer support chatbot helps when it answers from your own written policies, knows exactly when to hand a conversation to a person, and is measured on resolved problems rather than closed chats. This guide covers choosing a tool, preparing the knowledge base, writing the bot's instructions, escalation, tone, metrics and the mistakes that make customers hate support bots.
- A support bot is only as good as the help articles it answers from; fix the knowledge base before you switch the bot on.
- Write explicit escalation rules (refunds, complaints, safety, legal, repeated failure) and pass the full conversation to the human who takes over.
- Measure resolution and satisfaction, not deflection alone: a customer who gives up counts as deflected.
- Your company is responsible for what the bot says, as Air Canada learned in a 2024 tribunal ruling.
- In the EU, the AI Act requires chatbots to make clear that people are talking to an AI, with that duty applying from August 2, 2026.
What makes an AI customer support chatbot helpful?
A helpful AI customer support chatbot answers routine questions correctly from your own documentation, admits when it doesn't know, and hands off to a person fast when the problem needs one. Customers don't hate bots. They hate bots that loop, guess and block the way to a human.
It is a good fit when:
- A large share of your tickets are repeat questions with written answers (shipping times, returns, account access, pricing, how-to).
- Customers write outside your working hours and wait until morning.
- Your help center exists but people don't find the right article.
It is a poor fit when most tickets need account-specific investigation, judgment calls or negotiation, or when you have no written policies at all. In that case the bot has nothing reliable to answer from, and it will either refuse everything or invent answers.
Anthropic's customer support use case guide lists similar signals: high volume of repetitive queries, the need to pull information from a knowledge base quickly, round-the-clock availability, and sudden peaks in volume.
Should you buy a support chatbot or build one?
Most teams should start with the AI agent built into their help desk, because it already has the inbox, the handoff and the reporting. Build your own only when you need deep integration with your product or data that off-the-shelf tools can't reach.
| Option | Examples | Good for | Trade-off |
|---|---|---|---|
| Help desk AI agent | Intercom Fin, Zendesk AI agents, other help desk add-ons | Teams already on that help desk | Per-resolution pricing, less control |
| Chatbot platform | Standalone bot builders trained on your docs | Simple sites, small teams | Another tool to connect to your inbox |
| Custom build on a model API | Claude or another model with your own retrieval | Product-specific actions, full control | You build and maintain escalation, logging, evaluation |
At the time of writing (October 2026), Intercom's pricing page lists Fin at $0.99 per outcome, with help desk seats starting at $29 a month. Usage pricing means costs grow with volume, so set a budget and an alert.
Questions to ask any vendor before you sign:
- What counts as a billable resolution? Some vendors count a conversation where the customer stops replying. Ask for the exact definition.
- Which sources can it read? Help center only, or also PDFs, past tickets and your product database?
- How does handoff work? The human should see the full transcript in the same inbox, not a separate tool.
- Can you restrict topics and actions? You want to forbid refunds, discounts and anything account-changing.
- Where is conversation data stored, and is it used to train models? Get the answer in the contract or data processing terms, not only on a sales call.
- What reporting do you get? At minimum: resolution rate, escalations, ratings and the full text of every conversation.
A custom build gives you tools the bot can call, such as "look up order status" or "generate a quote". The Claude guide above walks through one in Python. If you go that way, read up on retrieval-augmented generation, because grounding answers in your documents is most of the work.
How do you prepare the knowledge base?
Prepare the knowledge base before you choose wording or tone, because the bot can only be as accurate as what it reads. Most bad bot answers trace back to a missing, outdated or contradictory article.
- Export recent conversations. Pull the last two or three months of tickets and chats.
- Group them by question. An AI assistant can cluster them for you; check the clusters by hand.
- Rank by volume. The top 20 questions usually cover a big share of tickets.
- Write or fix one article per question. One question per article, the answer in the first paragraph, exact numbers (days, amounts, limits), and the conditions and exceptions spelled out.
- Remove contradictions. If your returns page says 30 days and an old FAQ says 14, the bot will quote either.
- Mark internal-only information. Discount codes for staff, internal escalation contacts and fraud rules must not be in the bot's sources.
A useful article for a bot looks like this:
## Can I return an item after 30 days?No. You can return unused items within 30 days of delivery for a full refund.After 30 days we can't accept returns, except for faulty items, which arecovered for 12 months (see "Reporting a faulty item").- Start a return: Account > Orders > Return item- Refunds reach your card within 5 to 10 business days after we receive the item- Sale items marked "final sale" can't be returnedHow do you write the chatbot's instructions?
The instructions (the system prompt) set the bot's role, its source of truth, what it may and may not do, and how to escalate. Keep them specific and short enough that every rule matters. The Claude guide recommends adding several examples of good interactions, including ones where the bot applies a guardrail.
You are the support assistant for Northwind Outdoor, an online store forcamping gear. You answer customers in the chat widget on our website.Sources: answer only from the help articles provided below. If the answerisn't in them, say you don't have that information and offer to connectthe customer with the team. Never guess dates, prices, stock or policies.Link the article you used at the end of your answer.Style: friendly, plain English, short paragraphs, no more than 120 wordsunless the customer asks for detail. Match the customer's language.You can: explain policies, help find orders with the order lookup tool,explain how to start a return.You cannot: issue refunds, change orders, offer discounts, or makeexceptions to policy. For these, hand off to a human.Hand off immediately when the customer: asks for a person, is angry orupset, mentions injury or safety, mentions a lawyer or chargeback, or whenyou have failed to answer the same question twice.When you hand off, say a teammate will reply by email within one businessday and summarize the issue for them.<help_articles>...</help_articles>Test the prompt against your top 20 questions and a set of awkward cases: off-topic questions, attempts to get a discount, a customer pasting another customer's order number, and someone trying to make the bot ignore its instructions. Our prompt engineering guide covers the structure in more depth.
When should an AI support bot hand off to a human?
Hand off whenever the conversation needs authority, empathy or investigation the bot doesn't have, and whenever the customer asks. The handoff should carry the full conversation and a short summary so the customer never repeats themselves.
| Trigger | Why | What the bot does |
|---|---|---|
| Customer asks for a person | Blocking it destroys trust | Hand off at once, no persuasion |
| Refund, exception or compensation | Needs authority | Collect order number and reason, then hand off |
| Anger, distress, complaint | Needs empathy | Acknowledge briefly, hand off |
| Safety, injury, legal, chargeback | Risk to the company | Hand off and flag as priority |
| Two failed answers on one issue | Bot is out of its depth | Apologize, hand off with summary |
| Account security, identity checks | Fraud risk | Never handle in the bot |
Anthropic's guide suggests tracking escalation accuracy, the share of conversations that should have reached a human and did, and aiming for 95% or higher. Review missed escalations weekly; each one is either a new rule or a knowledge base gap.
How do you get the tone right?
Aim for the tone of your best human agent: short, specific, warm without being chatty. Bots go wrong in two directions: corporate filler that says nothing, or exaggerated cheerfulness that grates when the customer has a problem.
| Weak reply | Better reply |
|---|---|
| "I'm so sorry to hear about your experience! We truly value you as a customer and are committed to making this right!" | "Sorry your tent pole arrived bent. Faulty items are covered for 12 months, so we'll replace it. Can you share your order number?" |
| "Great question! Our return policy is designed to give you flexibility..." | "You can return unused items within 30 days of delivery. Start here: Account > Orders > Return item." |
Put three or four real examples like the better replies into the prompt. Examples shape tone far more reliably than adjectives like "friendly" or "professional".
How do you measure whether a support chatbot works?
Measure whether customers' problems got solved, not only whether they stopped talking to the bot. Track a small set of metrics weekly and read actual conversations alongside them.
| Metric | What it tells you | Target suggested in Anthropic's guide |
|---|---|---|
| Deflection rate | Share handled without a human | 70 to 80%, depending on complexity |
| Escalation accuracy | Right conversations reached a human | 95% or higher |
| Response accuracy | Facts match your sources | 100% on company and product information |
| CSAT | Customer rating after the chat | 4 out of 5 or higher |
| Repeat contact | Same customer back within 7 days on the same issue | Lower is better (not in the guide) |
Deflection rewards bots that wear customers down. A customer who gives up, or emails you an hour later, looks deflected. Pair deflection with repeat contact and CSAT, and sample conversations the bot "resolved".
A weekly review routine takes about an hour: read 30 random bot conversations, every escalation, and every conversation with a low rating. For each failure, decide whether the fix is an article, a prompt rule or a new handoff trigger.
What are the biggest mistakes with AI support bots?
The costliest mistakes are letting the bot invent policy, hiding the human option, and assuming the company isn't responsible for what the bot says.
- Invented answers. In Moffatt v. Air Canada (2024), a British Columbia tribunal held Air Canada responsible after its chatbot described a bereavement refund process that contradicted the airline's own policy, and rejected the argument that the chatbot was responsible for its own actions. Ground every answer in your sources and tell the bot to say "I don't know". Our guide to AI hallucinations explains why models guess and how to reduce it.
- No way out. Loops that refuse to connect a person cause more anger than no bot at all.
- No disclosure. Under Article 50 of the EU AI Act, AI systems that interact directly with people must make it clear they are AI unless that is obvious, and that obligation applies from August 2, 2026. A line such as "I'm Northwind's AI assistant" covers it and sets expectations.
- Stale knowledge. Policies change and articles don't. Assign an owner who updates the knowledge base whenever a policy changes.
- Leaking data. Never let the bot reveal order details without verifying the customer, and keep internal notes out of its sources.
- Launching everywhere at once. Start on one channel, learn, then expand.
Your launch checklist
Use this before the bot talks to its first customer:
- Top 20 questions each have one clear, current help article
- Internal-only information removed from the bot's sources
- Instructions include sources, limits, tone examples and handoff rules
- Handoff passes the transcript and a summary to a person
- The bot introduces itself as an AI assistant
- Tested against awkward cases, including attempts to override its instructions
- Budget alert set for usage-based pricing
- Weekly review of 30 conversations, every escalation and every low rating
If you are adding a bot to your site for the first time, our guide on how to add an AI chatbot to your website compares widget options and costs. And if you build a custom support bot with a coding agent, Orkio gives the agent an always-on box so it can keep working on the code while you are away.
Frequently asked questions
How much does an AI customer support chatbot cost?
Help desk vendors usually charge per resolution or per conversation on top of agent seats. At the time of writing, Intercom lists its Fin AI agent at $0.99 per outcome. A custom build on a model API costs the tokens per conversation plus the engineering time to build and maintain it.
Can an AI chatbot replace human support agents?
It can take over a large share of repetitive questions, but not the conversations that need judgment, empathy or authority, such as refunds outside policy or complaints. Most teams use the bot to answer routine questions and to collect details before a human takes over the rest.
What is a good deflection rate for a support chatbot?
Anthropic's customer support guide suggests aiming for 70 to 80 percent, depending on how complex your inquiries are. Treat any deflection number with care, because customers who give up or contact you through another channel also look deflected.
How do I stop a support chatbot from making things up?
Give it your help articles and policies as its only source, instruct it to say it does not know when the answer is not there, and have it link the article it used. Review a sample of conversations every week and fix the gaps it reveals in the knowledge base.
Do I have to tell customers they are talking to a bot?
In the EU, yes: Article 50 of the AI Act requires systems that interact directly with people to make that clear unless it is obvious. Outside the EU it is good practice anyway, because customers who feel tricked trust the whole company less.
