2024-07-31
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
arts | Wang Lin
Cover source | Tuchong Creative
When it comes to the most successful application of big models, the first thing that comes to your mind may be conversational robots.Kimi, Doubao, Wenxin Yiyan, etc. These star products with millions of monthly active users are still not profitable, and are not even ready to answer the question "Which is bigger, 9.11 or 9.8".
They need strong technical support, but it is still a long way from realization.AIGC(Artificial intelligence generated content) The future development is not optimistic enough. The views of Zhu Xiaohu, managing partner of Jinshajiang Venture Capital, have been pushed to the hot search. In his opinion, the first wave of AIGC opportunities in China will burst out in enterprise services, which will quickly grow and improve productivity. More importantly, after experiencing the last round of AI craze, more and more people have become market believers and tightened their wallets.
In the To B field, there is almost no scenario where the effect of AI outbound calls can be more clearly seen. The big model seems to bring an efficiency revolution to the service industry - while an employee with a monthly salary of 2,000 yuan can only make hundreds of calls a day, it can double or even 10 times the performance, and it is cheaper.
But consumers are not buying it - very few people can calmly answer an AI outbound call.
The other party may sell you a single-family villa in Qinhuangdao, or recommend a programming course for 7-8 year old children to you who have just graduated from college, or arrange a low-interest loan, or an insurance and financial management product...
At a time when most AI applications are suffering from high customer acquisition costs, huge technical investments, and difficult monetization scenarios, AI outbound calls powered by large models have almost invaded everyone's daily life, especially during major promotional events such as 618 and Double 11.
But the reality is not as good as imagined.
A phone call costs only 1 cent
One of the biggest annoyances for modern Internet users is being harassed by sales calls from unknown sources. Moreover, it seems that you can never block all the calls manually. You want to lose your temper, but the other party continues to talk calmly.
These calls are not made by real people. After all, humans have emotions. The calls you received were as calm as AI. In fact, they are AI. The system formulates the script, scans through a series of phone numbers, and calls one by one to find potential customers. This is a common operation for companies nowadays.
The key is that AI is cheaper than humans. Tech Planet consulted several companies that have launched AI outbound calling services. They said that there are generally two modes in the industry. One is charging by time, 1 cent per 6 seconds, and the other is charging by 6 seconds if it is less than 6 seconds. "The fee is divided into two parts, one is the operator's underlying call line fee, and the other is the AI robot fee." Some leading domestic intelligent customer service companies also said that they charge by traffic, and a call costs only a few cents.
The charging rules are also related to the industry. An industry insider said that the financial industry charges 0.15 yuan per minute, while the non-financial industry charges 0.13 yuan per minute. The reason why the financial industry is more expensive is that there are too many complaints, which leads to increased costs for AI outbound call companies.
Large models have always been considered a tool for productivity innovation.
According to Tech Planet, the average level of the AI outbound call industry is that a robot can make 800-1200 calls a day. This is equivalent to the workload of 3-5 people. If calculated by the cost, assuming one call per minute, 1200 calls will cost at most 180 yuan. Obviously, it is much cheaper than humans.
AI outbound calls are typical SAAS services. Some companies may need to customize the script, conversation process, knowledge base, robot training, etc. The fees for these links range from thousands to tens of thousands. But in fact, in many cases, their systems do not charge separately, but only charge according to the conversation duration.
Most AI outbound calling systems require a connection before they can deduct fees, but they cannot guarantee the connection rate. In order to ensure the effect, many companies will promise customers that they will call the numbers on the intended list multiple times. A salesperson from an AI outbound calling company said that the connection rate is more related to the accuracy of the data provided by the company. At present, the connection rate of AI is similar to that of manual work.
How many calls an AI robot can make in a day depends on the line resources. A salesperson from an AI outbound call robot company told Tech Planet that, like the urban transportation system, AI outbound calls also have peak hours, generally from 9 to 11:30 in the morning and from 2 to 5 in the afternoon.
In order to ensure work efficiency, companies often prepare multiple lines. They will match them according to the region. If they call a user in Beijing, they will use a line that shows the Beijing area to prevent users from seeing unfamiliar calls and directly refusing to answer.
But the first reaction of busy workers when they receive AI outbound calls is often irritation, and then they will block the number. An AI outbound industry insider told Tech Planet that if you don’t want to be harassed, you can directly complain about a certain number segment. If a number segment is complained, you need to immediately activate another number segment, that is, line resources.
If there are too many complaints, a number segment will be marked as a nuisance call by the operator. At this time, AI outbound calling companies often remove the mark every once in a while.
An AI outbound call industry insider showed Tech Planet their backend system. After a call is made to a consumer, they will classify them according to their intentions. Category A represents clear consumer intentions, Category B is in the middle, and Category C is a direct rejection. These data will eventually be precipitated into documents and returned to the customer. Customers with clear consumer intentions can be guided by subsequent text messages, follow the company's WeChat or official account, etc.
Caption: Screenshot of the backend system of an AI outbound calling company.
However, a construction materials company that has used AI outbound call robots said that it is impossible for robots to completely replace humans, because machines can never have flexible communication skills like humans. The essence of robots is to help people do preliminary screening. The connection rate in non-financial industries is between 40%-60%, while that in the financial industry may be only 25%-40%, or even lower. To maximize efficiency, companies will generally arrange manual follow-up for interested customers immediately.
But this is only an ideal situation. In reality, the customer conversion brought by AI outbound calls is not ideal. More than one merchant, Tech Planet, said that many cheap ones on the market are OEMs made with source code, and they do not have R&D capabilities at all. Many of their voice technologies are connected to iFlytek, which is not flexible enough.
"We use robots from leading companies in the industry, but their lines are very poor, and most of them are harassing calls. The connection rate is much lower than 60%, and many calls are hung up," said a merchant. "It's okay for after-sales service. In the early stage of customer acquisition, many people use the same batch of data. These people are bombarded repeatedly, and even real people will hang up in seconds, let alone robots."
In fact, many consumers thought they were real people when they first received an AI outbound call. More than one AI outbound call robot staff said that their voices were recorded by real people, not synthesized by a large model. However, if you want to be more human-like, you need more financial and technical investment. For example, make several sets of scripts. The more complete the scripts, the more diverse the calls will be.
An AI outbound call robot industry practitioner said that in the beginning, few companies could make money. Only by making a lot of calls could they make money. Making tens of millions of AI outbound calls a day would make money, such as when takeouts arrive, express deliveries arrive, and e-commerce events come. But the reality is that many companies are already very good if they can make hundreds of thousands of calls a day.
"Generally, companies at this level tend to choose self-research, and internal research and development can also be exported, such as Meituan, JD.com, etc. In fact, only the government and financial industries can make money, and there is no point in other industries," the person added.
In April this year, Zhu Xiaohu, managing partner of GSR Ventures, said in an interview that he is also a believer in AGI (artificial general intelligence), but he is not a devout believer in technology. He believes in applications that can be commercialized immediately.
Due to the huge population base and demand, there are indeed many scenarios in China where efficiency can be improved, such as AI interviews, marketing, etc. Zhu Xiaohu said that after two or three months of training with LLaMA, at least the top 30 human level can be achieved, and 50% of manual sales can be eliminated immediately.
Using open source models for training for two or three months before commercialization does not provide a strong moat, which has led to a mixed bag of AI outbound call services. Many of the AI outbound call robots on the market are startups, which may not have sufficient financial resources to develop large models.
This also led to the fact that in the last round of AI boom, many startups whose main business was AI outbound calling received financing, but not many in this round. Wisdom Tooth Technology, a leader in the field of intelligent customer service that has served many large companies such as OPPO, Xiaohongshu, and Didi Chuxing, received its last round of financing in 2022.
Ultimately, the effectiveness of large models still depends on computing power, algorithms, and data. These resources are mostly in the hands of leading enterprises or large companies.
A researcher from a large company told Tech Planet that there are three core technologies for AI outbound calls. The first is to understand what the other party is saying, then make a correct response, and finally reply to the other party with voice.
This involves three technologies: ASR speech recognition technology, NLP natural language processing, and TTS speech synthesis technology. Many people in the market use manual recording because they have not mastered TTS speech synthesis technology.
"In the past, it was a vocoder + acoustic model, but now the big model can do it all by itself, which has completely changed the technical paradigm. The interaction can be very realistic, and it can also simulate the process of real people breaking sentences, breathing and thinking, and can also understand dialects," he added.
For example, when an AI outbound call robot recommends coupons to consumers, the other party asks "what are there", which hides what the coupon can buy. When the other party says that it is inconvenient for him to travel, he can tell the consumer where the coupon can be used in the city closest to him.
In China, companies such as JD.com have already started to accumulate knowledge. For example, JD.com’s self-developed Yanxi big model is not only used for its own products such as JD Fresh and JD Self-operated, but also exported. In January this year, when Yili released its new carbonated milk drink “Milk PET”, it used Yanxi to customize the voice of the celebrity spokesperson “Times Youth League”.
Industry insiders said that some special industries, such as those involving financial data, require private deployment, and the price may be as high as hundreds of thousands, generally starting at 500,000 and costing millions. However, they often do not look for small and scattered companies, but choose larger and more powerful companies. After all, no one wants their money to go down the drain.
Big fish eat small fish, and small companies will eventually be swallowed up because of their lack of strength.
However, when technology becomes more mature, consumers can be reached more conveniently, and it becomes increasingly difficult to distinguish whether the other party is a robot or AI, this may not be a good thing.
How to better serve businesses while not excessively harassing consumers is an eternal proposition in the industry.
Welcome to Tech Planet🌟
第一时间接收文章更新👇
Collect and organize the latest news information and make it available for free for everyone to read
