In short
- The answer library is the project; the integration is a week
- Demos are rigged by definition — bring your own questions
- Ask who pays Meta's messaging charges, and how
- Handover to a human is a design decision, not a feature
Hiring an AI chatbot developer is unusually easy to get wrong, because the demo always works. A bot answering questions the vendor chose, about a business the vendor prepared for, will look superb. Yours will face questions nobody anticipated, about edge cases nobody documented.
These are the questions that separate the two situations.
1. What happens when it does not know?
Ask this first. The correct answer is that the bot says so plainly and hands over to a person, and that this behaviour was designed and tested rather than left to chance.
A wrong answer sounds like a description of the model — how capable it is, how large, how recent. Capability is not the issue. Every capable model will confidently invent your refund policy if nothing stops it.
2. Bring your own questions to the demo
Write down ten things your customers actually ask, including two awkward ones, and ask to see the demo answer those instead of the vendor's script. If they decline, or want them in advance to prepare, you have learned what the demo was.
Pay particular attention to the awkward two. Confident nonsense on a hard question is far more damaging than a graceful refusal.
3. Who writes the answers, and when?
This is where the project actually lives. A chatbot grounded in a thin FAQ page will improvise the rest. Someone has to collect the real questions, write correct answers in your voice, and decide the boundaries — and that work takes longer than the integration.
If the proposal has content as a minor line item, or assumes you will supply the answers without saying how, the quote is for the easy half. Our view on why this is a writing job is on the AI chatbots page.
4. Who pays Meta, and how much?
WhatsApp automation carries Meta's own per-message charges, billed by Meta rather than by your developer. Any quote that does not separate build cost from running cost is concealing something you will meet later.
Ask specifically which message categories your use case will generate, because the cost difference between answering inbound questions and sending marketing is substantial. The rules behind that are in AI chatbot development in India.
5. What does day two look like on Instagram?
Instagram gives a business twenty-four hours to reply after a customer makes contact, extendable to seven days when a human responds rather than an automation. A developer who has not designed for the moment that window closes has not built one of these before.
6. How will we read the transcripts?
The month after launch is when you discover what the bot gets wrong, and that is only visible in the conversations. Ask how transcripts are reviewed, who fixes the answers, and whether that work is inside the quote or billed later.
A vendor who treats launch as the end of the engagement is selling you a deployment, not a working assistant.
What a weak proposal looks like
| Signal | What it suggests |
|---|---|
| Content is a minor line item | The hard half has not been scoped |
| No request for your enquiry history | The answers will be generic |
| Handover described as “phase two” | The bot will have nowhere to escalate |
| A single price with no Meta charges mentioned | Running costs will arrive as a surprise |
| No mention of what the bot will refuse to answer | Boundaries have not been considered |
| Named model, unnamed writer | They are selling technology, not an outcome |
When you do not need one
If you receive a handful of enquiries a day, a person answering them is both cheaper and better. If your questions are genuinely varied and judgement-heavy, automation will frustrate people. And if your website has no clear answers on it, fix that first — a bot trained on a confusing site produces confusion faster.
The honest test: if you cannot write down the ten questions you are automating, you are not ready to automate them.