How an AI Agent Called 680 Car Dealers in the UAE in Three Weeks: Funnel Numbers and Mistakes
From Sep 25 to Oct 15, 2026, our AI voice agent called car dealers and importers in the UAE, offering cars for export. Below is the full funnel in numbers, with no rounding in our favor, plus the mistakes we found and fixed along the way.
Contents
The task
An exporter sells new cars from China wholesale, starting at one container, which is 3–4 cars of the same model. The buyers are dealerships, importers and traders in the UAE. The goal was to find out fast who actually buys these cars, who cares about the price, and who is worth a sales rep's time.
A human sales team can't cover that volume in three weeks. So the AI agent took the first touch: it calls, introduces itself, briefly explains the offer, asks two or three questions and, if there's interest, arranges to send a quote via WhatsApp or email.
The funnel in numbers
Period: Sep 25 – Oct 15, 2026. UAE calls only; test calls and our own numbers excluded.
| Stage | Companies / calls | Share |
|---|---|---|
| Companies called | 680 | 100% |
| Dials | 1,644 | — |
| Connected | 813 calls | 49.5% of dials |
| Conversation over 20 seconds | 314 companies | 46.2% |
| Warm leads (showed interest) | 112 | 16.5% |
| Requested a quote for a batch | 92 | 13.5% |
| Received a written offer | 91 | 13.4% |
| Negotiations with a sales rep | 6 | 0.9% |
How we counted: “warm” means the company reached the Warm, Hot, Offer or Negotiation stage in our CRM at least once. “Offer” means an email or message with a quote that was actually delivered. A conversation over 20 seconds means a person answered, not voicemail.
Where the calls went
Almost half of all dials never reach a person. That's normal for cold calling, but you need to know what “didn't get through” is made of:
- 831 dials never connected: 469 not picked up, 241 unreachable or busy, 121 failed to place (these 121 are part of the 1,644);
- of the 813 connections, many still weren't conversations: 222 hung up in the first seconds, 77 voicemail, 36 phone menus (“press 1...”), 34 rang out, 19 silence on the line, 7 technical drops.
So the honest metric isn't “connected” but a conversation with a person longer than 20 seconds: 314 of 680 companies (46.2%).
A meaningful conversation lasted about 79 seconds on average. That's enough to name the company and the offer, hear “sounds interesting, send it over” or “we don't do that,” and say goodbye.
Why people said no
Rejections are data too. The agent heard three things most often: “wrong number” (not a car dealer, or the wrong department), “we don't import, we buy locally,” and “Chinese brands aren't our segment.” We logged every one of these answers and used them to clean the list: if a company doesn't import, it doesn't need a second call.
Five mistakes we found along the way
1. “I'll pass it to the team” instead of an answer
A client asked in chat how much container shipping costs and how many cars that price covers. The answer was already in our data: the freight rate, how many cars of that model fit in a 40-foot container, and the price per car. The agent replied, “I'll check with the team and get back to you.”
We fixed it with a rule: look for the answer in the quote, catalog and logistics data first, and answer right away. “I'll pass it to the team” only when the answer isn't anywhere. More on this in our breakdown of seven voice agent mistakes.
2. A robot haggling over price
At first, when a client said “too expensive,” the agent asked what price they were paying now. That doesn't work well in B2B: buyers don't want to share numbers with a robot, and the call turns into an interrogation. Now the agent says a sales rep will discuss pricing, asks when the client plans to buy and when it's convenient to talk. If the client names a number, the agent records it but doesn't argue.
3. Not asking about purchase timing
The business owner heard this question on one of his own calls and asked us to add it everywhere: “So when are you planning your next purchase?” The answer instantly splits leads into “buying in the next two months” and “someday.” The first group gets a call from a person; the second gets a reminder in a month.
4. Calling the same number over and over
Because of a tagging bug, a missed call was logged as “call didn't happen,” and the system thought no attempt had been made. Some companies got up to eight calls in a day. We found it, fixed 95 records, set a limit of one dial per day for a cold number, and built a fallback for missed calls: a WhatsApp message and an email instead of a third call. We describe how that works in our article on voicemail and no-answers.
5. Reaching the receptionist instead of the buyer
In the UAE, reception often picks up. The agent tried to get the buyer's name and annoyed people. Now, if the receptionist asks to leave contact details, the agent gives the email once, says we'll message this number on WhatsApp first, and says goodbye. We get the buyer's name later, in the chat.
What three weeks taught us
An AI agent is good at the first touch. It doesn't get tired by the hundredth call, stays equally polite and records every answer. On the peak day the system made 477 dials in the UAE.
A person still closes the deal. Six negotiations out of 680 companies came from a sales rep working with the people the agent found. The agent saves the sales rep weeks of searching for who's worth talking to.
You only see mistakes in the recordings. We found almost every fix above by listening to and reading conversations. That's why we review conversations daily. We wrote about how that works separately: how the agent learns from its own conversations.
FAQ
Where did the list come from?
We built it ourselves from public business sources: maps, industry directories, classifieds portals. Of 2,733 companies found, 2,266 passed verification (82.9%), and in three weeks the agent called 680 of them. How we clean the list is covered in our article on building a B2B calling list.
What language did the agent use?
In the UAE, mostly English, the working language of business there. The agent can switch to Arabic, but we lock the conversation language in advance so it doesn't switch because of a random word.
Can this be repeated in another industry?
Yes, the mechanics are the same: a list, a short script, a question about purchase timing, handing warm leads to a person. The numbers will differ, since they depend on the list and the offer. We won't promise a conversion rate up front.