It is nine at night and someone messages your company’s WhatsApp: “Do you machine 6061 aluminum? I have a drawing, about five thousand parts a year.” Your sales rep sees it at eleven the next morning. By then the buyer has sent the same drawing to three other suppliers, and one of them answered before breakfast.
That is the real promise of a sales chatbot at an industrial company. It is not that it sells on its own. A manufacturer does not buy from a robot: it buys from a person who understands the drawing, knows what capacity the plant has and stands behind the price. What the assistant can do is make sure nobody is left waiting, and that when your rep opens the conversation, everything needed to quote is already there.
First, count who writes to you
Before building anything, read the last hundred messages that came in through WhatsApp and the website form. At a company in Juárez the mix is usually surprising: people looking for a job, current customers asking about their order, vendors pitching something, messages from nobody in particular, and somewhere in there, a few new buyers.
We will not give you percentages, because they change from one company to the next. Count them yourself. That count tells you the assistant’s first job: sorting. The job seeker gets the careers link, the current customer gets the status of their order or the name of whoever handles it, and the rep only sees buyers.
Often that alone pays for the project, because today someone on your team does that sorting by hand, after hours, from their phone.
What it does answer
Everything that is already public or should be, always from a document you approve:
- Which processes and materials you work with, and which you do not.
- Which certifications you currently hold, not the ones in progress.
- Whether you take low volumes or prototypes.
- Where you are, your hours, and whether you deliver in Juárez, El Paso or both.
- The status of an existing order, if the assistant is connected to your ERP or order system. It is the same information as the date you promised, and it only helps if that date was calculated properly.
The document is the key. An assistant that answers with what the model “knows” will make things up with great confidence. One that answers only from your document, and says “I don’t know that, let me connect you with someone” when the answer is not there, gets far less wrong and is easy to correct: every question it could not answer is a line missing from the document.
What it never answers
Prices, delivery dates, available capacity, and anything you do not actually comply with. Not approximate, not “starting at,” not “usually.”
The reason is not technical, it is commercial. A price is a commitment and a date is a promise, and what your assistant says, your company says. There is already a ruling on this: the airline that argued its chatbot was “responsible for its own actions” lost and had to pay. We covered it in the agent prepares, the person signs, and the rule from that article applies here in full.
What the assistant can promise is what you always deliver: “Laura will confirm that today before five.” That is the promise to protect, because if the assistant makes it and nobody keeps it, the chatbot has made things worse.
Its real job: getting the request in complete
Most of an industrial rep’s time on a new request does not go into quoting. It goes into asking for what was missing. And every back-and-forth costs a day.
A well-built assistant asks, one thing at a time and only for what is missing:
- The drawing, with its revision number.
- The material and any treatment or finish.
- The volume: how much per year and how much per lot.
- The timing: first delivery, samples or part approval.
- The quality requirements: material certificate, inspection, whatever standard the customer requires.
- Where it ships and who pays freight.
- Who is asking: name, company, email.
With that, the request lands in your CRM as an opportunity with everything attached and an owner: a named person who receives it and has a deadline to answer. Not “sales.” Someone.
This also fixes a problem almost nobody sees until it happens. If requests arrive today on each rep’s personal WhatsApp, the conversation lives on their phone, and the day they leave they take the pipeline. A company number, with an assistant that logs everything in the CRM, keeps the conversation where it belongs.
Speed matters, and it has been measured
There is a study that gets cited a lot and is worth reading. Researchers audited more than two thousand US companies by sending them a request and timing the reply. Companies that contacted the prospect within the first hour were nearly seven times as likely to qualify the lead as those that waited even one hour longer. And almost one in four companies never answered at all.
The study covers online sales in general, not manufacturing, and it is from 2011. But the mechanism has not changed: a buyer who sends a drawing to four suppliers pays attention to the first one that answers with substance. The assistant does not close the sale. It wins the first hour.
WhatsApp has its own rules
If the assistant is going to live on WhatsApp, which is where people on the border write to you, there are three Meta rules worth knowing before you design it:
The 24-hour window. When someone messages you, a 24-hour window opens, and it resets with each of their messages. Inside the window you can reply with free text. Outside it, you can only send templates approved by Meta. The “did you get a chance to review the quote?” follow-up three days later is a template, and it has to be registered ahead of time.
It is charged per message. Templates are charged per delivered message. And since October 1, 2026, Meta charges again for replies inside the window after the first thousand per month per number. For almost any industrial company that is pennies, but it has to be counted, like the rest of what an agent costs.
An assistant for your own business is allowed. Meta prohibits using WhatsApp to offer general-purpose AI assistants of the “ask me anything” kind. An assistant that serves and sells for your company does not fall under that ban. What its terms do prohibit is using the conversations to train AI models, except to fine-tune one for your exclusive use.
The prospect’s data counts too
When the assistant asks for a name, email and phone number, you are collecting personal data. If you operate in Mexico, the federal personal data protection law that took effect in March 2025 requires you to tell people beforehand, in your privacy notice, what data you process and why. In practice: the assistant links to your notice before asking for the first piece of data.
And a detail that gets missed: the law limits the use of the data to the purposes in that notice. If the prospect gave you their email to receive a quote and you later want to send them your newsletter, that purpose has to be in the notice, or you need their consent again.
The same goes for the model behind it: your prospect’s drawing is their customer’s confidential information. Before connecting anything, review what you can send to a model and what you cannot.
What to measure
Four numbers, every week:
- How long it takes a person to respond after a new request arrives. That is the number the assistant should bring down.
- How many requests reach the rep complete, without the rep having to ask anything.
- Which questions it could not answer. That list becomes next month’s document.
- How many times it said something it should not have. Someone reads a sample of conversations every week. If a price or a date shows up, it gets fixed that same day.
Do not measure “conversations handled.” That number goes up by itself and says nothing about whether you sold more.
When you do not need it
If you get only a few requests a month and someone answers them the same day, you do not need it. A good form that asks for the same things as the list above, and one responsible person, do the job without spending on anything.
It also does not help if almost everything you sell comes in through the purchasing portals of three large customers. Their buyers are not going to use your chatbot. What matters there is the CRM, so that quotes and their history belong to the company.
Where we come in
We build assistants like this as part of AI and automation: connected to your WhatsApp Business, your CRM and, if needed, your order system, with an answer document you approve and the rule that prices and dates come from a person. If you want to see how an assistant we built asks questions, try Chira: it does not sell anything, but it asks step by step and hands you a document at the end. And if you already know this would help, book a call.
References
- The Short Life of Online Sales Leads, Harvard Business Review — the 2011 study that measured how long companies take to answer a request and how much the odds of qualifying the prospect change after the first hour.
- Moffatt v. Air Canada, 2024 BCCRT 149 — the British Columbia Civil Resolution Tribunal decision that held the airline responsible for what its chatbot said.
- Pricing on the WhatsApp Business Platform, Meta — the 24-hour window, per-message pricing since July 2025, and charges for service replies since October 1, 2026.
- Meta Terms for WhatsApp Business Platform — section 4.7, “AI Providers,” on general-purpose assistants and on using platform data to train models.
- Ley Federal de Protección de Datos Personales en Posesión de los Particulares — Mexico’s private-sector data protection law, current text from the Chamber of Deputies. Article 11 limits the use of data to the purposes in the notice, and articles 14 and 15 set what the privacy notice must disclose.