Voice and chat ordering connected to your ERP: what it really takes to build it
Equipo Tecnea
Tecnea
In wholesale distribution almost everyone has the same bottleneck: orders come in by phone, WhatsApp and email, and somebody types them into the ERP by hand. It is slow, it goes wrong at peak times, and it depends on that person being there.
The idea of the customer ordering by speaking or writing, with the order landing by itself, has sounded like a promise for years. Today it is viable, but not in the way most people think. This is what it actually takes.
The easy part is understanding the message
Transcribing a WhatsApp voice note or interpreting "give me the usual and 20 more boxes of the new reference" is, with today's technology, the solved part. That is not where projects fail.
They fail on the next three.
1. The ERP is always in charge
This is the rule that decides whether the system is useful or a dangerous toy: the AI can never quote a price, a stock level or a lead time from memory.
Language models generate plausible text. If you ask for a price and it does not have it in front of it, it will invent one perfectly naturally, and it will be well written. In an order that is not an amusing error: it is a customer who has been confirmed a price that does not exist.
The only correct approach is for the assistant to query the ERP at that moment — that customer's price list, agreed discounts, real availability in their warehouse, credit limit — and for whatever it cannot query, it does not say. Technically this is solved by forcing the system to request the data from a specific function instead of answering freely. It is the difference between a pretty demo and something you can put in front of a customer.
2. Identify the customer before listening to the order
"The usual" only means something if you know who is speaking. And in B2B the same customer has their own price list, their own payment terms and sometimes their own product codes in their own naming.
That forces you to resolve, before anything else, how the person ordering is identified: by the phone number they write from, by a portal login, by a code. And what happens when someone new from a known customer calls. It looks like an administrative detail and it is the first design decision of the project.
3. From the customer's language to your product code
Here is the real work. The customer says "the small bags, the usual ones", "the 5-litre oil" or the trade name from a catalogue three years old. Your ERP has a code with an abbreviated description, written for something else entirely.
Matching one to the other requires three things nobody gives you for free:
- A catalogue with synonyms and ways of saying things, built from each customer's real order history, not from the item master.
- A clear rule for ambiguity. If two codes are possible, the system asks. It does not choose.
- A confirmation before saving. The customer sees what was understood, with real prices, and confirms. That is what turns a misunderstanding into a ten-second correction instead of a return.
What NOT to automate on day one
In order of common sense: start with repeat orders, which are most of the volume and the most contained case. Leave for later new product set-ups, price negotiations, urgent orders with special transport, and anything that today already requires somebody to pick up the phone.
And leave a person reviewing the queue for the first few months. Not out of distrust in the system, but because it is the only way to discover your business's edge cases, which are always more numerous than you think.
What you gain, honestly
The saving is not in the technology, it is in administrative time: orders stop being typed and the sales team stops acting as an intermediary between a voice note and a screen. The side effect usually matters more than the saving: the customer can order at eleven at night, which is when many professionals do their office work, and their order is recorded by morning.
What you do not gain: a system like this does not sell more by itself. It can suggest what is usually ordered together or flag that something needs restocking, but that is a separate feature, and it is better not to buy it in the same package until the basics work.
If you want the detail of how we approach it, with a fixed price: WhatsApp and voice ordering with AI, integrated with your ERP (in Spanish).
This article is informational. Tecnea builds this type of portal, so we have a direct commercial interest in the topic; we say so up front. The criteria above work for evaluating any proposal on the market, including ours — and the question about quoting prices "from memory" is the one that separates them fastest.
¿Te ha resultado útil este artículo?
Publicamos análisis sobre IA y tecnología empresarial. Sin spam — solo cuando escribamos algo que valga la pena leer.
Did you like this article?
Tell us what you'd like to automate in your company and we'll tell you, with no strings attached, where to start.

