AI Employee on WhatsApp: Setup Guide

WhatsApp is where the work already happens. If customers, suppliers, and your own team are all in one thread, an ai employee on WhatsApp makes more sense than a separate help desk. It keeps the conversation in the place people already check, which means faster replies and fewer dropped leads.

This is also where the phrase "AGI employee" gets used loosely. The useful version is much simpler: one system owns a channel, follows rules, and finishes work end to end. That is the category Perla sits in. It is the same category we call an AI employee.

If you want the broader definition first, read What is an AI Employee?. If you want the structural difference between a prompt box and a real operator, read AI Employee vs. AI Assistant. Then come back here and wire the channel.

1. Pick one job, not five

Do not start with "handle WhatsApp." Start with one repeatable job that already lives in the thread.

Good first jobs:

• Answer price, hours, and availability questions
• Qualify inbound leads before a salesperson steps in
• Triage support messages and collect the missing details
• Confirm booking requests and capture the date, time, and location

Bad first jobs:

• Handle every message in every language
• Negotiate refunds
• Make policy decisions with no written rule
• Jump between sales, support, and operations in the same week

WhatsApp works best when the job is narrow and the trigger is obvious. A customer asks for a quote. A supplier asks for payment status. A lead asks for the next available slot. The thread starts with a clear shape, and that is enough.

If you need a reference point for the kind of work Perla already does in this channel, see Perla for Customer Support. The pattern is the same even when the job changes.

2. Write the handoff rules before you connect anything

This is the part most teams skip. They connect the inbox first and think about rules later. That is backwards.

Write a short rule sheet with three buckets:

1. Reply
2. Ask one follow-up
3. Hand off

Keep the rules blunt. For example:

• Reply if the question is about price, hours, stock, or basic service info
• Ask one follow-up if the request is missing date, size, delivery address, or contact name
• Hand off if the thread mentions refund, chargeback, legal, angry, broken, urgent, or VIP
• Hand off if the answer is not in the source of truth
• Hand off if the thread goes past three back-and-forths without resolution

That rule set is enough to make the first version useful.

It also keeps the system aligned with how an AI employee should behave. The goal is not to sound smart. The goal is to stay inside the line you drew.

3. Give it one source of truth

An AI employee on WhatsApp should not guess from memory. It should read from one place and cite that place internally in the workflow.

Use one primary source and keep it boring:

• Google Docs for policy and canned replies
• Google Sheets for pricing, availability, or job status
• Notion for internal process notes
• Slack for human handoff
• Gmail when the work needs a real outbound email after the WhatsApp thread ends

If the answer lives in five places, the channel will drift. If the answer lives in one doc or one sheet, the system can stay consistent.

For the setup details on the channel itself, read Meta's official docs:

• WhatsApp Business Platform
• Cloud API overview

Those pages matter because they show how the channel is meant to work. Build around the official path, not around guesses.

And yes, this is where Google Workspace earns its keep. WhatsApp handles the conversation. Google Docs and Sheets hold the rules.

4. Teach three message patterns

Most WhatsApp work comes down to three message patterns. Teach those first.

Pattern 1: inbound and simple

Customer: "How much is the booth for Saturday?"

Good reply:

"Thanks. I’m checking the Saturday rate now. Can you send the city and the event start time?"

That reply does two things. It answers quickly and collects the one missing detail.

Pattern 2: inbound and known

Customer: "Can you book 2pm on 14 September?"

Good reply:

"Yes. I can hold 2pm on 14 September. Please confirm the venue name and contact number."

That is a clean handoff from conversation to action.

Pattern 3: inbound and sensitive

Customer: "The order arrived broken."

Good reply:

"I’m flagging this for a person now. Please send a photo of the item and the order number."

This is where an AI employee earns trust. It does not bluff. It collects the minimum details, then stops.

Use examples from your real inbox, not invented ones. Your own messages will teach the system faster than any generic prompt ever will.

5. Start in review mode, then let it send

Do not switch on auto-send on day one.

Run the first batch in review mode. A good first target is 20 to 30 real threads. That is enough to see where the rules fail without turning the whole inbox into a lab.

Review three things:

• Did it answer the right questions?
• Did it ask for the right missing detail?
• Did it hand off the right sensitive threads?

Then tighten the rules and move to auto-send for the safe buckets only. Keep sensitive buckets on review until the failure rate is close to zero.

This is the same operating pattern we use anywhere an AI employee touches real work. First it learns the shape. Then it earns more autonomy.

If you want a broader explanation of why that matters, the companion read is What can an AI employee do?. If you want the procurement angle, read AI Employee Pricing and Cost.

6. Measure what matters in the first week

Do not measure "AI quality" as a vague feeling. Measure the inbox.

Track these numbers:

• First response time
• Handoff rate
• Incorrect reply rate
• Time saved per thread
• Threads resolved without a human

You do not need a giant dashboard. A simple Google Sheet is enough for week one. Put the thread ID, the intent, the action taken, and whether a human had to fix it.

That sheet becomes your training loop. It shows which replies are reliable and which ones need a rewrite.

It also keeps the team honest. If the AI employee only looks good when nobody checks the threads, the setup is not ready.

What to do if it breaks

Three failure modes show up early.

The scope is too wide

If the system is trying to do support, sales, bookings, and billing on day one, cut it back. Pick one job. Ship that.

The source of truth is messy

If the answers are spread across Slack, Notion, Google Docs, and old WhatsApp chats, clean that up before you add more automation. One source beats four half-truths.

The handoff rule is weak

If angry, refund, broken, or legal messages still get answered automatically, stop auto-send and fix the rule. That is not a tuning issue. That is a boundary issue.

When the system is stable, expand it one job at a time. Add a second thread type only after the first one is boring. That is how an AI employee becomes useful instead of noisy.

If you want the bigger picture again, start with What is an AI Employee?, then compare it with AI Employee vs. AI Assistant. If WhatsApp is where your business actually runs, this is the channel to automate first.

Frequently asked questions

What should an AI employee do on WhatsApp first?

Start with one job that repeats every day: reply to common questions, qualify leads, or route support requests. The best first use is the job with clear rules and a clear handoff path, not the job with the most exceptions.

Do I need WhatsApp Business Platform for this?

If you want a system that can work with real message flows at scale, yes, that is the right place to start. For policy and setup details, read Meta's WhatsApp Business Platform docs and the Cloud API overview before you wire anything into live conversations.

Is this the same thing as an AGI employee?

No, but it is the same category people mean when they say AGI employee in marketing. The useful test is simple: does the system own a channel, follow rules, and finish work end-to-end without a human typing every step.

What if the AI employee answers something wrong?

That is why you start in review mode and set a hard escalation rule. If confidence is low, the thread is sensitive, or the customer is angry, the AI employee should stop and hand off to a human.

How much setup does WhatsApp need?

Less than most teams think. One source of truth, one handoff rule set, one review queue, and a short list of message patterns are enough for a real first deployment.