Most support inboxes are full of the same questions. How do I reset my password? Where is my invoice? Do you integrate with this tool? Answering them is necessary, repetitive and slow, which makes support one of the first places teams want to use AI.
It's also a place where AI can do real damage. A bot that invents a refund policy, promises a feature you don't have or leaks another customer's details costs more than it saves. The way to get the benefit without the risk is to treat AI as a triage layer with strict limits, not as a replacement for your support team.
Start with triage, not answers
The safest first step doesn't reply to customers at all. Use a model to read each incoming message and label it: topic, urgency, language, sentiment and whether it mentions billing, cancellation, a bug or a security issue.
That alone helps. Urgent and angry messages jump the queue, messages go to the right person, and you start learning which questions really are repetitive. After a few weeks of labels you'll know exactly which topics are worth automating.
Answer only from your own content
When you do start answering, the model should answer from your help center, policies and product docs, not from its general knowledge. This approach is usually called retrieval-augmented generation, or RAG.
- Collect the source material: help articles, FAQs, pricing and refund policies, and past answers your team is happy with.
- Split it into short passages and index them so the system can search by meaning, not just keywords.
- For each question, retrieve the few most relevant passages and give only those to the model.
- Instruct the model to answer only from the provided passages, and to say it doesn't know when they don't cover the question.
- Include a link to the source article in every answer, so customers and your team can check it.
Retrieval is only as good as the content behind it. If your docs are outdated or contradict each other, the answers will be too. Fixing the help center is often the most valuable part of the project.
Decide when a person takes over
A good handoff is the most important feature of an AI support system. Write the rules down before you build anything.
- The customer asks for a person, in any wording.
- The topic is billing disputes, refunds, cancellations, legal matters, security or account access.
- The retrieved passages don't clearly answer the question.
- The customer is frustrated, or has already written more than once about the same issue.
- The answer would require looking up or changing something in their account.
When the handoff happens, pass the whole conversation and the AI's labels to the human agent so the customer never has to repeat themselves. And tell the customer plainly that a person will reply, with a realistic time.
Guardrails that keep you out of trouble
- Never let the model make commitments: no refunds, discounts, deadlines or promises about future features.
- Keep account data out of the prompt unless it's essential, and never expose one customer's data to another.
- Treat customer messages as untrusted input. Someone will try "ignore your instructions", so the model shouldn't have tools or permissions that would matter if it obeyed.
- Label AI replies clearly, so customers know they're talking to an assistant.
- Set limits on message length and the number of AI replies per conversation before a person steps in.
Roll it out in stages
Moving straight to fully automatic replies is how embarrassing screenshots happen. A staged rollout lets you see quality before customers do.
- Labels only: classify and route messages, no replies.
- Draft mode: the AI writes a suggested reply that an agent edits and sends. Track how often drafts are sent unchanged.
- Auto-reply for a few topics: only the questions where drafts were consistently accepted, with handoff rules active.
- Widen gradually, one topic at a time, based on the numbers.
Measure what matters
Look past the share of tickets closed by AI. That number goes up just as easily when the bot is unhelpful and customers give up.
- How often agents send AI drafts without edits.
- How often customers write back after an AI answer, or ask for a person.
- Customer satisfaction on AI-handled conversations compared with human-handled ones.
- A weekly review of a sample of AI conversations, read by a person.
Keep a test set of real questions with good answers, and re-run it every time you change the prompt, the model or the docs.
Deeraf builds support triage like this as a Build Sprint, starting with labeling and drafts, and can keep it tuned through On Call as your product and docs change.