AI customer support on WhatsApp, without inventing prices
In many small businesses sales aren't lost on price. They're lost by answering late. And whoever answers does it between customers.
There is a pattern in any business selling over WhatsApp: nearly every enquiry is the same three questions — do you have this, how much is it, do you ship it — they get answered between customers, and the ones arriving in the evening or at the weekend wait until the next day. By then the customer has already asked somewhere else.
The obvious answer is to automate. The answer that works is to automate without the system making anything up, which is a considerably more specific problem.
The real risk isn't tone, it's price
A generic chatbot wired to WhatsApp learns to sound friendly in an afternoon. The problem shows up when someone asks about a specific part. If the system answers with a plausible price instead of the real one, you end up in one of two situations: either you charge something different from what you said, or you honour a price you never meant to give. Both are worse than not having answered.
An invented price isn't an AI error. It's a commercial commitment your company never made and now has to honour.
So the right technical question is not "which model do I use?" but "where does the number going out over chat come from?". If the answer is "the real catalogue, looked up at that moment", the system is usable. If it is "the model", it isn't.
The four foundations
- The official WhatsApp Business API from Meta. Not a personal number wired up to an automation tool: that drops, gets blocked and offers no delivery guarantees.
- The real catalogue connected. What it answers is what you charge, looked up against your system rather than approximated.
- Your rules written outside the model: what it can confirm on its own, what it must never say, and when it alerts a person.
- Testing against your difficult conversations before letting it talk to a customer. You already know the easy ones go fine.
The third point is what separates an experiment from something you can leave running. If the rules live inside the prompt, the model interprets them — and sooner or later someone talks it out of one. If they live outside, the model never even sees what isn't its business.
What happens when it doesn't know
An honest system has to be able to say "I don't know this" and hand the conversation to a person without the customer feeling a jolt. It sounds minor and it is what builds the most trust: a customer forgives not getting an instant answer; they don't forgive being promised a part that doesn't exist.
Drawing that line properly — what gets confirmed automatically and what gets escalated — is a business decision, not a technical one. It is worth making before connecting anything.
What changed at a real company
A vehicle dismantler selling spare parts received practically all its enquiries over WhatsApp. Every reply meant looking up the part and its price by hand, and sales were lost by answering late, not on price. After building it on the official API with their real catalogue connected, they went to handling 4× more enquiries than before, answering in seconds around the clock, with prices looked up against their own catalogue. The full case is in case studies, and the product is Talqo.
The most telling metric isn't the 4×. Three weeks after the first delivery, the client extended the scope. When someone decides to put more of their operation into a system, they have genuinely tested it.
How much work this actually is
Less than it looks, if you start where it hurts. You don't need to cover every case on day one: you need to cover the three questions that make up most of the volume properly, and escalate the rest to a person. That narrowing is what gets it running in weeks rather than quarters.
And measure one thing at first: how many enquiries are resolved without anyone stepping in. If that number rises and complaints don't, you're on track.
Frequently asked questions
Can you use AI on WhatsApp without getting the number blocked?
Yes, using Meta's official WhatsApp Business API. Automating a personal number with unofficial tools is what causes blocks and outages, and it offers no delivery guarantees. The official API is the only stable route for a business.
How do you stop the AI inventing a price?
You don't ask the model to remember prices: you connect the real catalogue and look the figure up at the moment of replying. The model writes the message, but the number comes from your system. If the data isn't available, the system says so or hands the conversation to a person.
What happens when a customer asks something the AI doesn't know?
It should acknowledge it and escalate to a person without the customer feeling a break. What gets confirmed automatically and what gets escalated is a business decision worth making before connecting anything: it is the line that stops the company being committed to something it can't deliver.
How many enquiries can it handle?
At a vehicle dismantler selling spare parts, they went to handling 4× more enquiries than before without adding headcount, answering in seconds out of hours too. The multiplier depends on how much of the volume is repetitive questions, which on WhatsApp support is usually most of it.
How long does it take to get running?
Weeks, not quarters, if you start with the questions that concentrate most of the volume and escalate the rest to a person. Trying to cover every case from day one is what turns this into a long project.