Most customer support chatbots fail for the same reason: they were pointed at the wrong tickets. Routing every inbound message through an AI does not cut costs, it shifts complaints from the inbox to social media. The teams that get value from a customer support chatbot draw a hard line between two categories of work: repetitive, low-context tasks where automation lowers handle time without lowering CSAT, and high-context, emotionally loaded tasks where a bot’s confidence becomes a liability. Knowing the line is the whole job.
What to automate
Password resets, order tracking, return initiation, business-hours checks, and “where is my refund” queries are textbook chatbot territory. The intent is narrow, the data lives in a system the bot can reach, and the customer wants an answer faster than a human can type one. According to Zendesk’s CX Trends 2026 report, high-performing support teams now resolve between 40% and 60% of tier-1 queries through an automated channel without a CSAT drop, and the median first-response time on those tickets falls below 30 seconds.
Knowledge base lookups are another safe bet. A retrieval-augmented chatbot pointed at your help center, product docs, and policy pages handles “how do I cancel my subscription” or “is overnight shipping available to Spain” without escalation. The bot does not need to be creative. It needs to retrieve the right paragraph and quote it accurately.
Triage is the third clear win. Even when the bot cannot resolve a ticket, it can collect the order ID, the device, the screenshot, and the urgency before a human agent picks up. Agents start the conversation with full context, average handle time drops, and CSAT often rises because the customer did not have to repeat themselves.
What not to automate
Complaints, churn risk, financial disputes, and anything involving a medical, legal, or safety question should bypass the bot. The cost of a confident wrong answer in those categories is measured in escalations, refunds, regulatory exposure, or worse. A customer disputing a charge does not want a chatbot to summarize your refund policy. They want a human who can override it.
The other category to keep off the bot is anything that requires reading tone. A customer typing in all caps about a missed delivery does not need a friendly suggestion to “check our tracking page.” They need an agent who can apologize, escalate, and offer a credit inside the same conversation. AI tone detection has improved, but it still confuses frustration with confusion often enough that the safest pattern is to route emotionally loaded conversations to a human within the first two turns.
How to think about the line
The simplest test is the cost of being wrong. If a wrong answer costs the customer 30 seconds and a clarification, automate it. If a wrong answer costs the customer money, time, or trust, route it to a human. Build the bot around that asymmetry, not around a deflection-rate target.
The other test is data access. If the chatbot has structured access to the answer (an order database, a policy document, a clinical formulary), it can be accurate. If the answer requires judgment on top of partial information, it cannot.
Where the platform choice fits
SaaS support tools like Intercom, Zendesk, and Ada cover most of the “what to automate” list out of the box. They are the right call for teams under 30,000 monthly conversations with non-sensitive data and a tight launch deadline.
Teams that need tighter control over the AI behavior, data residency, or integration depth tend to outgrow SaaS within a year. That is where open-source platforms come in. Chatguru, an open-source chatbot platform from Netguru, is one example: it ships as a production-ready RAG stack you self-host, so customer data never leaves your infrastructure and every intent, prompt, and escalation rule stays editable. The trade-off is operational maturity. Self-hosted means someone on the team owns embeddings, vector stores, and prompt versioning. For brands in regulated industries (healthcare, insurance, finance) or with custom support flows that SaaS cannot model, that trade-off pays off.
The platform is rarely the constraint. The harder work is the policy decision underneath: which tickets get automation, which tickets get a human, and how you measure the difference. Get that right and the chatbot becomes a multiplier. Get it wrong and the chatbot becomes the reason customers churn.