In this article
An AI sales agent on WhatsApp does five jobs per conversation: answer from your documents, ask the qualification questions you defined, write the answers into the contact's profile, move the deal to the right pipeline stage, and hand over to a person at the point you set. It does not run unsupervised by default. Every capability that changes data is a tool you switch on, and the risky ones (payment links, template sends) wait for a human approval until you change that.
This guide shows one conversation turn by turn with the tool calls underneath, then the setup decisions that decide whether it sells or embarrasses you. For a general introduction to AI chatbots versus rule-based bots, read AI WhatsApp chatbot first; this post is about operating one.
What does an AI sales agent actually do in a WhatsApp chat?
Example (hypothetical): an online IELTS academy runs click-to-WhatsApp ads. Tool calls are in brackets; the customer never sees them.
Customer: Hi, I saw your ad. How much is the IELTS course?
AI: Hi! The 8-week evening batch is AED 1,400. To point you to the right batch: what band score do you need, and by when?
Customer: Band 7 by December. It's for a visa.
[save_customer_details: target_score = 7, deadline = December, purpose = visa]
[add_contact_label: "visa-deadline"]
[move_pipeline_stage: "Qualified"]
AI: Band 7 by December is realistic if you start the batch on 4 November. We do a 15-minute level check first so the tutor knows your starting point. Shall I book it for tomorrow evening?
Customer: Can you do 20% less? A friend paid less.
AI: I can't change fees myself. What I can do is ask my colleague Sara whether instalments are possible. Want me to?
Customer: Yes please.
[handoff_to_human: reason = discount and instalment request, summary = "Band 7 by Dec for visa, wants 4 Nov batch, asking for 20% off", department = Admissions]
AI: Done. Sara from admissions will reply here today. Nothing else you need to do.
Look at what happened. The agent quoted a price from its knowledge base and did not negotiate. It saved three qualification answers as structured data instead of leaving them in the chat. It moved the card without a human clicking. It stopped at the discount, the exact point a person should take over, and passed a summary so Sara does not re-ask.
Had the customer said "yes, I'll take it", the next call would be [create_customer_payment_link], which is a high-risk tool that needs approval by default. The customer sees "sending your seat link now", a manager approves the pending action, and the link goes out.
Which tools can a WhatsApp AI sales agent call?
Only these are real, from WhatsGrow AI's tool list. Tools are off until you enable them in Settings → AI (two exceptions: save_customer_details switches on by itself once you have defined custom fields, and send_training_image once you have taught it files). Each has a risk level that sets whether it needs approval by default.
| Tool | What it does | Default approval |
|---|---|---|
| save_customer_details | Writes qualification answers to custom fields you defined | No |
| update_contact_field | Updates name, email, labels or custom fields | No |
| add_contact_label / remove_contact_label | Tags the contact | No |
| move_pipeline_stage | Moves the lead to a stage by exact name | No |
| create_followup_task | Schedules an automatic follow-up message to this customer | No |
| create_task / assign_task / complete_task | Internal tasks for your team | No |
| assign_to_agent / assign_to_department | Routes the conversation | No |
| handoff_to_human | Passes the chat with a reason and summary | No |
| add_note | Internal note on the conversation | No |
| send_training_image | Sends an image or document you taught it | No |
| create_and_send_invoice | Creates and sends an invoice or receipt | No |
| create_customer_payment_link | Creates a payment link with your enabled gateway and sends it | Yes |
| send_template_message | Sends an approved template | Yes |
| fetch_external_data | Reads an external URL, such as an order status endpoint | No |
Industry packs add their own tools (for example eCommerce order lookup or appointment booking). Qualification fields work only for custom fields you have created, with the exact keys; that is a feature, since the agent cannot invent columns. Create the fields first, under custom fields.
What should the agent qualify, and where does the answer go?
Ask only what changes who handles the lead or what you offer them. Four to six fields beat twelve. Example set for a service business:
| Field (custom field) | Why it matters | Example answer |
|---|---|---|
| Need | Decides the product or team | "IELTS Academic" |
| Budget or plan size | Filters out mismatches early | "Up to 1,500" |
| Timeline | Sets urgency and stage | "Band 7 by December" |
| Decision maker | Avoids chasing the wrong person | "Self" or "Parent pays" |
| Location or branch | Routes to the right person | "Dubai" |
In your instructions, tell the agent to ask one question at a time, to skip anything already answered, and to save each answer as soon as it arrives. Then mirror the fields in the pipeline: a lead with need, budget and timeline filled in belongs in "Qualified"; one with only a name does not.
The AI page also checks your instructions against your setup. If you mention a stage name that does not exist, or ask the agent to move leads while that tool is off, it flags it, which catches the most common silent failure: instructions that promise an action the agent cannot take.
When should the AI hand a WhatsApp chat to a human?
Write handoff rules as triggers and actions, not as a general "escalate if needed":
| Trigger | Action |
|---|---|
| Customer asks for a person, or is angry | handoff_to_human immediately, summary included |
| Discount, custom terms, refunds or legal questions | handoff_to_human to the right department |
| Price or policy not in the knowledge base | Say you will check; hand off; never guess |
| Ready to pay | Payment link goes through approval, then human confirms |
| Two failed attempts to understand the customer | Hand off |
| Customer stops replying after a question | create_followup_task, 20 hours out |
On follow-ups, WhatsApp only lets you send free-form messages within 24 hours of the customer's last message. The follow-up tool's own guidance is to keep the delay under 23 hours so it can auto-send; anything at 24 hours or beyond notifies your team instead of sending. If the customer writes again first, the scheduled follow-up cancels itself.
Per-channel modes help while you build trust: "Drafts for me to send" has the agent write the reply and leave sending to a person, and "Replies automatically" sends it. Start in drafts, read a week of them, then switch to automatic. You can also limit the agent to your working hours so it stays silent outside them.
How does an AI sales agent fail, and how do you prevent it?
| Failure | What you see | Guardrail |
|---|---|---|
| Invented price or policy | Customer quotes a number you never set | Put prices in the knowledge base; instruct "if it is not in the documents, say you will check and hand off" |
| Wrong stage name | Lead does not move | Use exact stage names; heed the setup check |
| Over-asking | Customers drop out mid-qualification | One question per message, max five fields |
| Wrong language or tone | Answers in English to Arabic or Spanish speakers | Instruct it to reply in the customer's language; test in each language |
| Promise it cannot keep | "We'll call you in 5 minutes" | Ban time promises; use handoff and tasks |
| Payment sent too early | Link before the customer agreed | Keep payment-link approval on |
| Loops | Same question repeated | Review weekly; add "do not repeat a question already answered" |
Two outside guardrails. First, Meta's business terms bar general-purpose AI assistants on the WhatsApp Business Platform from 15 January 2026, but Meta told TechCrunch it does not affect businesses using AI to serve their own customers (TechCrunch, October 2025). Keep your agent scoped to your business. Second, since 1 October 2026 each reply beyond 1,000 per number per month is a billable service message; see what that costs. Shorter agent answers are cheaper.
Before switching to automatic replies, test with the simulator in Settings → AI and with a colleague's phone. Try the unfair questions: a competitor's price, a refund demand, a message in the wrong language.
Which KPIs show whether the AI sales agent works?
Measure these weekly from your inbox and pipeline, using your own data:
- First response time. Should drop to seconds; report the median, not the average.
- Qualification completion. Share of new chats that end with all required fields filled.
- Handoff rate and reason. If most handoffs are "price not found", your knowledge base has a gap. If most are "asked for a person", the opening needs work.
- Approval rejections. Payment links or templates managers decline show where the agent oversteps.
- Stage conversion. Qualified to booked, booked to paid, by source. Compare AI-handled chats with human-handled ones from the same ad.
- Cost per conversation. Meta's service-message charges plus your plan; use the ROI calculator for the blended figure.
Do not compare against vendor "close rate" claims. Your baseline is last month's conversion with humans only, on the same traffic.
Setup order that works: write the knowledge base, define fields and stages, turn on four tools (save details, move stage, create task, hand off), run in drafts for a week, then go automatic. See WhatsApp AI agent for the feature overview and WhatsApp automation for flows that run before and after the agent.
Sources (checked 13 Oct 2026)
- WhatsGrow AI tool definitions and approval defaults (product behaviour, current release)
- TechCrunch: WhatsApp changes its terms to bar general-purpose chatbots, effective 15 January 2026
- Meta: Pricing on the WhatsApp Business Platform (service messages charged from 1 October 2026, 1,000 free per number)