The WhatsApp agent that answers when the message arrives

The tool list is the product

WhatsApp Business API integration is what lets software send and receive messages on your business number instead of a personal handset. The platform sets the rules. Customers have to opt in, messages you start first must use templates approved in advance, and the window for free form replies closes after the customer's last message. The agent is designed around those rules from the first day.

Most of the effort in building a WhatsApp agent goes into the instructions, and most of the value comes from the list of things it can actually do. An agent that can look up an order, read live availability and create a booking is useful to somebody. The same agent with a better personality and no tools is a search box that types slowly.

Each tool is a contract: what it takes, what it returns, what it is permitted to change and what happens when it fails. Reading tools are cheap to add and low risk. Writing tools change something in the business, and every one of them deserves a limit, a confirmation and a log entry.

The refusal is the part that gets left out. An agent with no tool for a question should hand over rather than improvise. A polite guess about a refund policy is the failure that costs a customer, and it looks exactly like a correct answer until somebody checks it against the policy.

• Read: order status by reference, delivery date, opening hours, stock, price for a listed item
• Read: whether this number belongs to an existing customer, and what they last bought
• Write: create or move a booking, inside a stated range of times
• Write: raise a support ticket, with the conversation attached to it
• Write: send a payment link, for an amount the agent was handed rather than one it produced
• Refuse: anything with no tool behind it, routed to a person instead of answered from memory

How it decides, and how you find out when it decided wrong

When somebody reports that the bot said something strange, the only useful question is which tool it called and what came back. Without a record of that, the available fix is to rewrite the instructions and hope.

Every turn should leave a trace: the message received, what the agent took it to mean, which tool it called with which arguments, what the tool returned and what went back to the customer. That trace turns a vague complaint into a specific fix. It is also what lets somebody read a week of conversations and find the three questions the agent keeps getting wrong.

A wrong answer and a wrong action are not the same size of mistake. Telling somebody the wrong opening hours is embarrassing. Moving the wrong booking is a phone call and an apology. Gates should be proportionate: read tools run freely, write tools confirm the specifics back in plain language before acting, and anything involving money either stops for a person or uses a value the agent was given.

Voice notes, screenshots and two languages in one sentence

A large share of what arrives on WhatsApp is not typed text. Customers send voice notes because talking is faster than typing, photographs of the thing they want, and screenshots of a payment they say they have made. An agent that only reads text will answer a fraction of its inbox and look broken for the rest of it.

Voice notes are transcribed before the agent sees them, which adds a second place for errors to enter. Names, addresses and reference numbers are precisely what transcription gets wrong, so the agent should read the important detail back in text and wait for confirmation before acting on it. A booking made against a misheard date costs more than a request for clarification ever will.

Payment screenshots need a rule of their own. An image showing a completed transfer is a claim rather than a verification. The agent can read the reference and the amount out of it and check them against the payment record, and where it cannot, the correct behaviour is to acknowledge receipt and pass it to a person instead of confirming an order that has not been paid for.

Language is the other half. Customers routinely mix English and Urdu inside one message, sometimes in Urdu script and sometimes typed phonetically with no fixed spelling. An agent built to expect careful English will mishandle a lot of ordinary traffic, and the reply should come back in the language the customer used. That is a comprehension requirement rather than a courtesy.

• Voice notes transcribed, with critical details confirmed in text before any action is taken
• Images read for what they contain, with claims checked against records rather than believed
• Mixed English and Urdu handled inside one message, including spellings that vary
• Replies matched to the language the customer used rather than to a house default

Designing inside the window and the templates

WhatsApp Business API messaging runs under rules that shape the agent more than any design preference does. Customers have to opt in. After a customer's message the business can reply freely for a limited window, and once that window closes anything the business sends has to be a template approved in advance.

The practical consequences are large. A follow-up sequence cannot be improvised, because the message going out well after the conversation has to exist and be approved before the conversation starts. Templates are written up front with variables where the specifics go, so a sequence is designed as a small set of shapes rather than free conversation.

The agent has to know which side of the window it is on before it composes anything at all. Inside the window it can write. Outside it, the only decisions available are which approved template to send and what to put in the variables, and an agent that does not check will produce a message the platform simply refuses to deliver.

One design move follows directly from this. Give the customer a reason to reply, because a reply reopens the window. A short question that is easy to answer at the end of a message does more for a follow-up sequence than a better-written template will.

• Opt-in recorded with its source and date before anything is sent
• A small set of approved templates covering the sequences the business genuinely runs
• Variables designed in from the start, so one template covers many cases
• The window state checked before every outbound message, without exception
• Free text reserved for inside the window, where the conversation is live

Handover, and the shape of a good one

The handover is the part customers judge. Done badly it is a dead end where the agent stops replying and nobody arrives. Done properly the customer is told what is happening and roughly when, and is met by somebody who has already read the conversation.

The triggers belong in writing rather than in the agent's discretion, because an agent asked to judge whether it is out of its depth will usually decide that it is not.

What the person receives matters as much as when. A notification saying a customer needs help is a worse handover than a short summary of what the customer wants, what has been established and what the agent could not do, with the conversation attached underneath.

Out of hours needs a stated answer rather than silence. If a handover fires at two in the morning, the customer should be told when somebody will be there, and the item should be at the top of the queue when they arrive. A handover that disappears into an empty office is the failure to design against.

• The customer asks for a person, in any wording, including plain frustration
• A complaint, a refund request or any dispute about money
• Anything above a stated value threshold
• Two consecutive turns where the agent did not understand
• Anything touching health, safety or somebody's legal position
• Any request with no tool behind it

When an agent should not answer for you

If nobody is available to take a handover, the agent becomes a wall rather than a door. The escalation path is a staffing commitment, and a business that cannot make it should keep the agent narrow enough never to need one.

If the business cannot answer its own common questions consistently, the agent will not either. Where a policy lives in three people's heads and those three disagree, writing it down is the project and the agent is what comes afterwards.

If the number is a personal handset shared between a family and a business, moving it onto the Business API changes how it behaves and somebody will mind. That is a conversation to have before the migration rather than during it.

And an agent that has to check with a person every second turn is worse than a list of saved replies. Where the incoming questions are mostly unique, the honest recommendation is to improve how quickly messages reach the right human and leave the answering to them.

What is already running, what moves the number, and who holds the number itself

Messaging of this shape is running rather than described. Center for Sight in New York books appointments largely hands-free across calling agents, SMS and patient messaging. Culligan Pakistan, delivered through East River, took 140 leads in a single month behind an AI powered CRM that cut response times and automated the lead replies. Big Texas Land Buyers runs more than 500 calls a day on voice agents with automatic CRM categorisation, which is the same tool discipline pointed at a different channel. Twenty five engagements across six countries, worked from Karachi and billed month to month.

No price appears on this page. What moves the number on a WhatsApp agent is the tool list rather than the conversation. Read tools are cheap and low risk. Write tools each need a limit, a confirmation and a log entry, and they are where the work actually goes. After that it is how many approved templates a sequence needs, whether your order and booking systems will let anything write to them, and who maintains the agent when the platform changes a rule. A supplier quoting on message volume alone has not looked at your write tools.

The audit takes the first week and it goes on that tool list, because an agent's usefulness is decided there and nowhere else. One agent with two or three read tools and a single write tool is live inside a fortnight, which is enough to answer order status and create a booking. Further write tools are added one at a time afterwards, each with its own confirmation and its own trace, because a wrong answer and a wrong action are different sizes of mistake and only one of them ends in an apology.

The number, the templates, the opt in list and the conversation history are yours and sit in accounts under your own logins, so you own the system and can move it to another supplier or run it yourself without asking anybody. Running it yourself becomes realistic the moment the tool contracts are written down, and a business with an engineer who enjoys this should take the agent in-house at that point. That is the natural end of the arrangement rather than a failure of it. Ibrahim owns build and workflows at Wobble and would rather hand over the keys than be the person somebody has to ring to change a sentence.

Common questions

What is a WhatsApp AI agent?

Software that reads incoming WhatsApp messages, works out what the customer wants, uses a defined set of tools to do something about it, and replies. It runs on the WhatsApp Business API rather than the consumer app, which is what allows automated replies, routing and a shared record. The useful part is the tool list.

Can a WhatsApp agent handle voice notes and images?

Yes, with a rule attached. Voice notes are transcribed before the agent reads them, and transcription reliably mishears names, addresses and reference numbers, so anything important should be confirmed back in text before it is acted on. A screenshot of a payment is a claim rather than a verification and belongs checked against the payment record.

How does the WhatsApp messaging window affect the agent?

Inside the window, which opens when the customer messages, the agent can write freely. Outside it, everything the business sends has to be a template approved in advance, so follow-up sequences are designed and approved before they run rather than composed on the day. The agent checks which side it is on before composing anything.

When should a WhatsApp agent hand over to a person?

On an explicit request in any wording, on a complaint or refund or dispute about money, above a stated value threshold, after two consecutive turns of not understanding, on anything touching health or safety, and on any request it has no tool for. The person should receive a summary plus the full conversation.

Can the agent call the customer instead of messaging?

WhatsApp supports business-initiated calling, though not everywhere: Meta excludes the United States, Canada, Egypt, Vietnam and Nigeria. Availability is a platform rule that can change, so confirm it for your own market before designing around it. A design where a missed call triggers an immediate message achieves most of the same outcome without depending on it.

Will customers know they are dealing with an agent?

They should be told, and telling them costs nothing. The outcome to avoid is a customer discovering it in the middle of a complaint and feeling handled. Disclosure also makes the handover read as a feature, because somebody who knows they were talking to software expects a person for the difficult part.