Building an AI customer support agent that doesn't embarrass you
A practical guide to building an AI customer support agent for a startup: scope, knowledge, handoff, guardrails and how to launch it without surprising your customers.
An AI customer support agent is one of the most popular first AI features, and one of the easiest to get wrong in public. When it fails, it fails in front of a customer: a made-up refund policy, a promise you can’t keep, or a polite loop that never reaches a human.
Building one that helps is mostly about decisions you make before writing any prompts. Here’s how we’d approach it.
Pick the questions it should answer
Don’t start with “answer every support question”. Start with your inbox.
- Group recent tickets by topic and find the common, repetitive ones: order status, password resets, how a feature works, billing dates.
- Pick a handful the agent can answer well from information you already have written down.
- List what it must hand off: refunds, legal questions, angry customers, anything involving money or account changes.
A narrow agent that answers some questions well beats a broad one that answers everything badly. See picking your first AI feature for the same idea applied more widely.
Give it the right knowledge
An agent is only as good as what it can look up.
- Use your help center, policies and product docs, retrieved per question, rather than stuffing everything into the prompt. RAG or fine-tuning explains why.
- Clean up the sources first. Outdated articles and contradicting policies will be repeated faithfully.
- Connect live data carefully. If it needs to check order status, give it a narrow, read-only lookup for the signed-in customer, not broad database access.
- Tell it to answer only from what it retrieved and to say when it doesn’t know.
If answers come back vague or wrong, check retrieval before the prompt. Our post on why RAG stops working covers the usual causes.
Make the handoff to a human easy
The fastest way to annoy a customer is to trap them with a bot.
- Always offer a way to reach a person, and make it visible, not hidden after three failed attempts.
- Hand off automatically on the topics you listed, on repeated “that didn’t help”, and on signs of frustration.
- Pass the conversation along so the customer doesn’t have to repeat themselves.
- Tell customers they’re talking to an AI. Hiding it damages trust when they find out.
Set guardrails for what it says and does
- No promises it can’t keep: refunds, discounts, delivery dates or exceptions to policy.
- No actions without confirmation, and nothing irreversible at all in the first version.
- Stay on topic. Decline unrelated requests politely rather than improvising.
- Protect customer data. Only show a customer their own information, and be careful what goes into logs. See keeping customer data safe in LLM features.
Test it before customers do
- Build a test set from real tickets, including awkward ones: vague questions, two questions in one, sarcasm and things it should refuse.
- Write down what a good answer looks like for each, and when it should hand off.
- Run the set on every prompt or knowledge change. We cover how in test your LLM feature before users do.
Launch in stages
- Draft mode: the agent writes replies and your team reviews and sends them. You learn fast with no customer risk.
- Limited launch: answer one or two topics directly, for a share of traffic, with easy handoff.
- Expand topic by topic as the reviewed answers prove reliable.
Watch it after launch
- Read a sample of conversations every week
- Track how often it hands off, and why
- Track answers customers mark as unhelpful
- Turn every bad answer into a new test case
The short checklist
- Start with a short list of common, repetitive questions
- Answer from clean, retrieved sources, and say “I don’t know” otherwise
- Make reaching a human easy and automatic for risky topics
- Never promise, refund or act without a human
- Test on real tickets, launch in draft mode first, and keep reviewing
If you’re planning a support agent and want it built properly the first time, see our fixed-scope offers or book a 30-minute call.