7 AI Voice Agent Mistakes in B2B Sales and How We Fixed Them

Every example below comes from real calls and chats in our car export campaign: 1,644 dials in the UAE and 1,114 in Venezuela. Names and companies removed, lines translated.

1. “I'll pass it to the team” when the answer is already there

Before. A client asked: “Container shipping costs this much. Is that for three cars or one?” The agent said: “Let me check with the team and get back to you.” Yet the quote the client already had contained everything: how many cars of that model fit in a container and what shipping comes to per car.

Why it hurts. A buyer is comparing several suppliers. The one who answers on the spot looks like they have things under control. “We'll get back to you” sounds like someone who doesn't know their own price.

After. The “answer from our data first” rule: the agent looks for the answer in the quote it sent, the catalog and logistics data, and answers right away, noting how long the figure is valid. It can hand a question to a person only if the answer isn't anywhere.

— That's per car. Three cars of this model fit in a 40-foot container, so shipping per car is a third of the rate. The quote is valid through the end of the month.

2. A robot haggling over price

Before. When a client said “too expensive,” the agent politely asked: what price are you paying now, from whom, on what terms?

Why it hurts. A buyer owes an unknown supplier no numbers, least of all a robot. The call turned into an interrogation, and a warm client went cold.

After. The agent doesn't dig for the client's price. If the client names a number, the agent repeats it once, records it and says a sales rep will discuss pricing. Then it asks about purchase timing and a good time for a call. From there, a person takes over.

3. No question about purchase timing

Before. The agent asked about models and quantity but skipped what matters most to sales: when.

After. On first contact, after models and quantity, one question: “When are you planning your next purchase?” “This month” becomes a task for a sales rep with a call time in the client's time zone. “In the spring” becomes a reminder.

4. The agent switches language over a random word

Before. The other person would drop a couple of words in another language, and the agent might switch mid-call, even though the person spoke perfectly good English. On top of that, the voice changed along with the language: one person called, another one continued.

After. The conversation language is chosen before the call, based on the country and the contact's language, and then locked. One voice for the whole call. If the person directly asks for another language, the agent switches, but in the same voice.

5. The agent doesn't hear “no”

Before. Two kinds of mistake. The client says “no” over our pitch, and the agent asks to finish and starts over. The client turns down a call with a sales rep, and the agent offers it again. In one call, four times in a row.

After. A short “no” is a no, without asking why. A “no” over the pitch makes the agent stop. If the client turns down a sales rep call, we don't offer it again: “The quote still stands, we'll reply on WhatsApp,” and goodbye.

6. A minimum order that scares people off

Before. In the intro the agent said “from 20 cars.” For a small dealer, 20 cars of one model is a huge batch. Some ended the call right there; others, with a request for 3–5 cars, couldn't tell whether we'd even quote them.

After. “From one container,” which is 3–4 cars of the same model. Essentially the same logistics requirement, but it sounds like a normal wholesale batch, not a condition for big chains.

7. Promises the agent can't keep

Before. A client asked about right-hand drive. The agent answered with a line from a template: “The team will confirm right-hand drive along with the quote.” It sounded like a promise, and that version might not have been available at all.

After. The agent records the request but doesn't confirm what it doesn't know: “Noted: right-hand drive. I'll check with the team.” A separate rule: don't misdate past contacts (“we spoke last week”) when the call was yesterday. Small things like that erode trust fastest.

The common thread

Almost none of these mistakes come from the model being “dumb.” They come from the script having no clear rule for that situation. So we log every case: what the client said, why the agent got it wrong, what rule we added and how to check that it works. How that loop works is covered in our article on agent self-learning.

FAQ

Why not let the agent negotiate on its own?

You can, but in B2B with large deals it's a false economy. The buyer wants to talk to a person who owns the terms. The agent keeps what it does best: finding who's interested and setting up the conversation.

How does the agent know an answer is “in the data”?

Before the call and during the chat, it has the quote sent to the client, the model catalog and the logistics facts. If the question is about any of that, it answers right away. The agent never makes up new prices, discounts or terms.

How many rules in a script is normal?

As many as you need, and no more. A long script costs more on every call and gets followed worse. We cap the script size, and when we add a rule, we remove an outdated one.