How to Scope a WhatsApp Chatbot or AI Voice Agent.

"We want a WhatsApp bot" and "we want an AI that answers our phones" are good starting points, but they aren't a scope. Two businesses can say exactly the same sentence and need projects that differ by months of work. One needs a bot that answers ten common questions. The other needs one that checks live stock, takes a booking, sends a payment link and hands angry customers to a manager.

This checklist covers the questions that decide the size, risk and cost of a conversational AI project. Answer them first, even roughly, and every quote you receive becomes easier to compare. You'll also spot early whether a simpler tool would do.

1. Name the one job the bot has to do

Start with a single sentence that describes success: "Answer delivery and return questions on WhatsApp without a person, 24/7." Or: "Pick up every call after 6 pm, book appointments into our calendar, and text the caller a confirmation." If you can't write that sentence, the project isn't ready to scope yet.

Then list what the bot will not do in version one. Refunds, complaints and anything involving money or medical, legal or account changes usually belong in the "hand to a person" column at first. A clear "not yet" list is the cheapest scope control there is.

2. Pick the channel for your customers, not for the trend

• WhatsApp works best where customers already message businesses there, which includes much of the Middle East, India and parts of Europe and Latin America. It handles text, images, documents, buttons and lists well.
• Voice works best where people still pick up the phone: clinics, home services, real estate, restaurants, and older or less tech-comfortable customers. It's also the answer to missed calls after hours.
• SMS works on any phone and suits reminders and short two-way replies, especially in the US. Business senders usually need to be registered first.

Many projects end up combining two channels, such as a voice agent that books the appointment and a WhatsApp or SMS message that confirms it. Decide which one is primary for version one.

3. List what it needs to know, and where that lives

An AI bot is only as accurate as the content behind it. Write down every source it will answer from: FAQs, policies, price lists, service menus, product catalogues, opening hours. For each, note whether it's a tidy document, a messy spreadsheet or "in someone's head". The last category is common, and turning it into written answers is real work that belongs in the plan.

Also note how often each source changes. A catalogue that updates daily needs a live connection. A returns policy that changes twice a year can simply be loaded and refreshed.

4. Separate answering from doing

This is the line that changes project size most. A bot that only answers questions from your content is a contained build. A bot that does things, like creating bookings, checking order status, updating your CRM or sending payment links, is an AI agent, and each action is an integration to build, secure and test.

For every action, write down the system it touches (calendar, CRM, helpdesk, store, payment provider), whether that system has an API, and whether a person should confirm the action before it happens. Sensitive steps should start with human approval. You can relax that once the bot has earned trust.

5. Design the handoff before the happy path

Every bot meets questions it can't answer and customers who want a person. Decide up front:

• What triggers a handoff: the customer asks for a person, the bot isn't confident, a keyword such as "complaint" or "cancel", or a high-value customer.
• Where the conversation goes: a shared WhatsApp inbox, a live call transfer, a callback request, or a ticket in your helpdesk.
• What happens outside working hours: a promise of a reply time, a callback slot, or an emergency number.
• What the person receives so the customer never has to repeat themselves, usually a short summary of the conversation so far.

A bot with a clumsy handoff frustrates people more than no bot at all, so this part deserves as much attention as the answers.

6. Languages, tone and voice

List every language the bot must handle and whether customers mix them in one message, as English with Arabic or Hindi often is. Modern models cope well, but each language still needs testing with real examples.

For voice agents, also decide on the voice itself: a stock voice, an original brand voice, or a clone of a real person's voice, which should only ever be done with that person's documented consent. Decide how the agent introduces itself too. Telling callers they're speaking with an AI assistant is good practice and, in some places, a legal requirement.

7. Accounts, numbers and rules for your market

Before anything goes live, a few pieces have to exist, and they should be in your name:

• WhatsApp: a number on the WhatsApp Business Platform (not just the free Business app), a verified business, and approved message templates for anything you send first. Customers need to have opted in to hear from you. Meta charges for some conversation types, with rates that vary by country.
• Voice: a phone number with a telephony provider, or forwarding from your existing number, and a check of local calling and recording-consent rules.
• SMS: sender registration where required, such as A2P 10DLC in the US or DLT in India.
• AI providers: API accounts with OpenAI, Anthropic, Google, ElevenLabs or whichever you use, billed to you so usage costs stay visible.

Owning these accounts means you keep your number, your data and your bot if you ever change development partners.

8. Decide how you'll judge it after launch

Pick two or three measures you can actually check: share of conversations resolved without a person, missed calls answered, bookings made, average first-reply time. Ask for a way to read real conversations, because transcripts reveal problems no dashboard will.

Plan for the first few weeks after launch as part of the project, not an afterthought. That's when you find the questions nobody predicted, tighten answers and adjust handoff rules. Someone needs to own that, on your side or ours.

The one-page version

If you only have ten minutes, write down: the one job, the channel, the content sources, the actions it takes, the handoff rule, the languages, and who owns the accounts. That single page is enough for any competent team to give you a meaningful written scope, and to tell you honestly if a simpler tool would do the job. For WhatsApp specifically, our WhatsApp chatbot development page covers how we approach builds.

Not sure which of your workflows is worth automating first? Our free AI automation audit looks at one workflow and tells you in writing which steps suit a bot, which suit a person, and what would need to be in place first.