How to build your first AI agent for WhatsApp commerce
A step-by-step guide to building your first AI commerce agent: role, knowledge base, skills, guardrails, connecting WhatsApp, testing, and launching small.
An AI commerce agent is built, not coded — you configure it in plain language. But “configure it” hides the part that decides whether customers trust it: what you set up, and in what order. This is the step-by-step for a first agent that’s reliably useful, using AI Studio as the build surface and WhatsApp as the channel. For the bigger picture of what these agents are and do, start with the pillar: AI agents for commerce.
Before you start: what you need
Three things, and none of them is technical:
- A channel — a WhatsApp Business number (existing is fine), a website to host a chat widget, or both.
- Your real content — the FAQ you answer every day, your shipping and returns policy, and your product information. The agent is only as good as what you give it.
- One clear first job — the single thing you want it to do well before anything else. Resist “do everything.” Narrow and reliable wins.
Step 1 — Give the agent a role and a voice
Write, in plain sentences, who the agent is and how it speaks. Name the brand, the tone (warm and concise, not chirpy), the languages it serves, and the boundaries of its job. A good role is specific: “You are the assistant for [brand], a fashion label. You help customers find items, check orders, and answer sizing and returns questions. You are friendly and brief. You never discuss anything outside [brand].” Specific roles produce consistent agents; vague roles drift.
Step 2 — Feed it a knowledge base
This is where reliability comes from. Pull your existing FAQ, shipping and returns policy, sizing guidance, and product details into the agent’s knowledge base. You can upload documents or point it at your site to crawl. Two rules: keep it current (a wrong return window is worse than no answer), and keep it specific to questions customers actually ask. The agent answers from this content, so the better it is, the less the agent improvises.
Step 3 — Turn on the right skills (start with one)
Skills are the jobs the agent can do. The temptation is to switch them all on. Don’t — start with the one your first job needs, prove it, then add the next. Match the skill to the goal:
| Your first goal | Switch on | Add next |
|---|---|---|
| Cut “where’s my order?” tickets | Order Tracking | Data Collection |
| Turn questions into sales | Product Recommendation | Create Order |
| Capture leads from ads | Data Collection | Follow-up |
| Recover quiet conversations | Follow-up | Product Recommendation |
A single skill done well builds the trust that makes customers lean on the agent for the next one.
Step 4 — Set guardrails and a handoff rule
Decide what the agent must never improvise: prices it can’t invent, discounts it can’t offer, policy exceptions it can’t grant, claims it can’t make. Write those as explicit rules. Then set one firm handoff rule — the trigger that passes a conversation to a human, with the summary, order state, and a suggested next step attached. Common triggers: an upset customer, a refund over a set amount, a question outside the agent’s job. The handoff is what makes automation safe; a customer should never feel trapped with a bot.
Step 5 — Connect a channel and test like a difficult customer
Connect WhatsApp — through the Cloud API, or your existing number via Coexistence (how API access works). Before you let real customers in, test it the way a hard customer would: ask it something off-topic, phrase a question three awkward ways, request something against policy, and try to confuse it. You’re looking for two things — wrong answers (fix the knowledge base) and missed handoffs (fix the escalation rule). Need realistic messages to test with? Borrow from the WhatsApp message templates.
Step 6 — Launch small, then measure and tune
Don’t flip it on for every conversation at once. Start with one channel or one segment, and watch three numbers: resolution rate (closed without a human), CSAT (collected right after each conversation), and conversion (replies that become carts or leads). Tune in tight loops — adjust the knowledge base and rules, watch the effect, repeat. Widen scope only when the numbers hold. Everything the agent learns lands on the bitCRM record, so quality compounds: the next conversation starts smarter than the last.
Common mistakes to avoid
- Boiling the ocean. Six skills on day one means six things to debug at once. Start with one.
- A thin knowledge base. If the agent doesn’t know your return window, it’ll guess. Feed it before you launch it.
- No handoff rule. An agent with no exit is how customers end up trapped and angry. Always give it a person to pass to.
- Optimizing for “sounds human.” Customers don’t care if it sounds human; they care if it solves their problem. Measure outcomes, not personality.
- Launching to everyone at once. Small launches surface problems cheaply. Scale after the numbers prove it.
Ready to build? Open AI Studio and start with one job — or read the pillar, AI agents for commerce, for how the whole system fits together.
Frequently asked questions
How long does it take to build a first AI agent?
You can stand up a working agent in an afternoon: role, a knowledge base from your existing FAQ and policies, one or two skills, a handoff rule, and a connected channel. Tuning it to the point where you trust it on most conversations takes a week or two of watching real chats and tightening the knowledge base and escalation rules. Start narrow and the first version is quick.
What should my AI agent do first?
Pick the single highest-volume, lowest-risk job — usually answering 'where's my order?' and basic product questions. Nail that one before adding order creation or proactive follow-ups. A narrow agent that's reliably right beats a broad one that's occasionally wrong, because trust is what gets customers to lean on it.
Do I need a separate WhatsApp number for the agent?
No. You can run the agent on your existing business number — either through the WhatsApp Cloud API, or by keeping your number on the WhatsApp Business app and adding automation through Coexistence (COEX). You keep your number and your chat history. See the WhatsApp Business API guide for how access works.
How do I stop the agent from saying something wrong?
Three controls. Ground it in a knowledge base so it answers from your approved content, not invention. Write explicit guardrails for the things it must never improvise — prices, promises, policy exceptions. And set a handoff rule so that when it's unsure or out of scope, it passes to a human instead of guessing. Test it adversarially before launch to find the gaps.
How do I know when to let the agent handle more?
Watch resolution rate and CSAT together. When the agent is closing its current job's conversations without humans and CSAT is holding, add the next skill — and only one at a time, so you can see its effect. If CSAT dips when you widen scope, pull back and tighten the knowledge base before trying again.
- AI agents
AI agent guardrails: keeping a commerce agent on-brand
How to set guardrails that keep an AI commerce agent on-brand and accurate: grounding in your own content, a 'never improvise' list, and testing before launch.
- AI agents
Designing clean AI-to-human escalation for commerce
How to design AI-to-human escalation customers never feel: when to hand off, the context to carry, where to route it, and how to measure and tune the seam.