AI in E-commerce: Undercurrents Surging - Can AI Cure the Price War in E-commerce?
-
If competing on low prices is an involution, then besides becoming even more involutionary, the only way out is technological innovation.
Around the theme of 'low prices,' e-commerce giants have struggled for an entire year, from adjusting business strategies to reorganizing structures, and even replacing top executives, yet none of these measures seem to have shaken Pinduoduo's structural advantage in 'low prices.' As a result, Pinduoduo has become the 'reflection point' for both Jack Ma and Richard Liu.
Sandalwood e-commerce monitoring data shows that from October to early December, the GMV growth rates of the four major platforms (Tmall, JD.com, Douyin, and Kuaishou) saw a recovery due to the lower base from the previous year. Compared to the previous quarter, there was a slight improvement, with year-on-year growth rising from 5% to 8%, with Douyin performing the strongest. The aforementioned institutional analysts noted that in terms of product categories, liquor, contact lenses, outdoor sports, and health supplements performed exceptionally well, all achieving double-digit growth. In contrast, cosmetics and clothing sales were relatively sluggish, underperforming the overall market. CITIC Securities previously estimated that Pinduoduo's GMV growth rate during the Double 11 period would increase by more than 20% year-on-year.
Data source: Sandalwood
Facing the fierce competition from Pinduoduo, Taotian, and Douyin E-commerce, even Baidu, despite its repeated setbacks, has turned its attention to AI. Earlier, Jack Ma mentioned on the internal network, "The AI e-commerce era has just begun, presenting opportunities for everyone." Wu Yongming, who concurrently serves as the CEO of both Taotian and Alibaba Cloud, is seen as the precursor to advancing AI e-commerce. He previously stated that AI technology would become the core means to achieve two goals: price competitiveness and consumer-centric services.
It's not just Taotian; even Baidu has reignited its e-commerce ambitions for the fourth time.
On January 10th, Baidu E-commerce announced its AI e-commerce plan, initially leveraging free AI digital humans, an upgraded search engine, and content monetization through Baijiahao to restart its e-commerce business. Meanwhile, Robin Li's new year OKRs once again include objectives related to e-commerce. In addition to Taotian and Baidu, JD.com and Douyin E-commerce are also eyeing the market aggressively. For instance, at the end of November, Douyin established a new AI department called 'Flow'. Earlier this year, it internally tested a feature called 'Worth a Look,' where users could jump to product pages via tags in short videos—a format seen as the next-generation shopping cart solution.
AI-driven e-commerce is not a new concept; it represents an 'efficiency revolution' at its core. By leveraging AI and large models, platforms, merchants, users, and products are now connected more efficiently. The core logic is to reduce costs and improve efficiency through technological means, offering an alternative solution beyond Pinduoduo's structural 'low-price' strategy.
Late at night, A-Cheng (pseudonym) sat alone in his office, with only a few computer screens glowing in the darkness. "In the past, if the computer performance was poor, the digital human live streaming would crash with just a slight lag. Now, with AI advancements, the computer can run it smoothly even in the background," said A-Cheng, an online operator for a consumer goods company. He noted that technological progress has not only made live streaming smoother but also highlighted cost advantages. Real-person live streaming costs 300-400 yuan per hour, while a cartoon-style digital human costs about 20,000 yuan per year, averaging less than 55 yuan per day.
This doesn't mean digital humans can dominate the field.
Currently, A-Cheng is troubled by two issues: one is that digital human live streaming requires pre-edited scripts, resulting in poor interactivity; the other is that platforms like Taobao often impose various restrictions on digital human live streaming, including limited traffic. "The cost is already lower than real-person streaming. If the traffic were also higher, who would use real-person streaming in the future?" From a cost perspective, the three forms of live streaming e-commerce—celebrity broadcasts, store broadcasts, and digital avatars—have completely different cost structures.
Industry insiders have calculated the requirements for maintaining positive gross margins under different live streaming formats. The comprehensive fees (commission + slotting fees) for top and mid-tier celebrity broadcasts range between 30%-70%, with brand ROI typically controlled between 1%-3.3%. The advantage of celebrity broadcasts is their ability to quickly scale up sales.
Precisely because celebrity broadcasts place extremely high demands on gross margins, mature brands that have established themselves well on Taobao find their profit margins limiting their ability to transition to Douyin. "Big brands primarily focus on store broadcasts on Taotian, with a maximum frequency of about 20 live sessions per month," the source mentioned. The situation on Douyin is entirely different—while brands do have store broadcasts there, boosting sales still relies heavily on large-scale influencer collaborations. "From what I know, many new brands on Douyin can exceed 20,000 influencer-hosted sessions in a single month, which provides unparalleled exposure."
Perhaps noticing the trend of established brands' "Douyin fatigue," Taobao has been aggressively strengthening brand store broadcasts over the past year. According to estimates from two brands, the cost of human-hosted store broadcasts is at least 400 RMB per hour per person. With multiple hosts, supporting roles, stage management, set design, and equipment amortization, the costs can be even higher. "The issue with store broadcasts is that the costs are rigid. Compared to influencer collaborations, it's more like a heavy-asset investment. If sales don't pick up, it's equivalent to losing money."
To address the rigid cost issue of human-hosted store broadcasts while ensuring round-the-clock coverage, many brands, especially in fast-moving consumer goods and health products, have introduced digital humans, adopting a hybrid model of human hosts + AI streamers. A relatively well-produced digital human costs around 100,000 RMB per year, while the "standard" cartoon-style digital humans offered by platforms/service providers can go as low as 20,000 RMB, with potential for further cost reductions. Perhaps to capture user attention, major players like Baidu and Taotian, as well as smaller companies like Silicon Minds, have adopted free strategies to attract users to digital humans over the past year. Digital humans are essentially low-cost "online shopping assistants" and, in a sense, serve as alternatives to live streaming by real hosts.
In contrast, JD.com has taken a more conservative approach, positioning digital humans as supplements to its live streaming business. During Double 11, the company focused on procurement-focused live streams, further lowering the priority of digital humans. Earlier, Shang Xin, Vice President of JD Group and Head of JD Retail Technology R&D and Data Center, explicitly stated that digital humans would primarily serve as fill-ins for real hosts in the short term.
Previously, JD's technology department believed that combining digital humans with large models would be more suitable for customer service scenarios, possibly due to JD's strong cost-cutting needs. Moreover, the operational status of digital humans varies across different industries. Sectors such as food & beverages, personal care, and health supplements exhibit higher awareness levels, which may be closely related to the intensity of industry competition and overall online gross margins.
Data source: Sandalwood
According to Sandalwood's e-commerce monitoring data, food & beverages, personal care, and health supplements have all shown varying degrees of rebound since the first week of October. However, growth trends indicate that the acceleration during promotional periods falls short of non-promotional time nodes, suggesting that daily operational investments have a more significant impact on growth rates. JD Shangxin stated that the focus of AI technology integration with e-commerce lies in four key areas: intelligent shopping guidance and customer service, intelligent quality inspection, R&D efficiency improvement, and operations maintenance.
Shangxin's statement reflects the common approach of comprehensive e-commerce platforms in AI transformation.
In December last year, Taotian consolidated its more than 20 AI teams into four groups: Alimama, C-end consumers, B-end merchants, and industry-specific applications. Specific implementation scenarios include marketing represented by digital humans and search recommendation optimization, as well as consumer-facing services like Taobao's Q&A and intelligent customer service. The third category is intelligent business tools for merchants, such as generating images and video detail pages using multimodal technology. If combined with the technical services and machine deployment provided by Alibaba Cloud for Taotian, the logic is no different from that of JD Retail.
Intelligent customer service is currently the most mature application scenario for AI in e-commerce.
Chen Zhe, VP of Technology at Zhichi Technology, told Photon Planet that intelligent customer service mainly comes in two forms. One is a product based on knowledge extraction from a knowledge base, utilizing AIGC capabilities for generation. "The advantage of this solution is that it can reduce manual maintenance costs and improve the response rate to questions. The drawback is that because the robot learns automatically and uses large models to answer, the responses may not be as precise." Shortcomings hinder implementation.
"When Wenxin Yiyan or Spark produces an incorrect answer, users might laugh it off, but in serious business practice, a single error can mean immeasurable losses." This comparison by Cao Peng, Chairman of JD.com's Technology Committee and President of JD Cloud Division, reflects the caution many companies exercise when applying such technologies.
Another integration method involves allowing customers to upload documents, which are then automatically generated by large models to build a knowledge base. Typically, only reviewer approval is needed before going live. From a business model perspective, the extent to which AI customer service can reduce costs and increase efficiency determines the profit margin for related service providers. According to Zhichi Technology's calculations based on dozens of retail clients, after integrating large models, the accuracy rate of responses improved from 40% to 70%, with a cost rate of only 20%.
Chen Zhe revealed that retail brands are more willing to purchase such services. "Before 2023, even with promotion, we might not have convinced clients. Since 2023, many more clients have come to us proactively."
Offline retail prefers third-party AI customer service products for multiple reasons, including more noticeable cost reductions and the presence of offline shopping guides and private domain services. Additionally, unlike online retail, offline retail is not dominated by platform economies. In fact, current e-commerce platforms are not friendly towards third-party service providers, as most platforms do not open their API interfaces to the public. According to an industry insider, the sensitivity of private domains and data makes platforms reluctant to open up to third parties.
Platforms not opening their interfaces does not mean that in-house development by major companies is effective. On the contrary, under the pressure of KPIs such as revenue and conversion, e-commerce platforms have shown little enthusiasm for investing in the development of universal features like AI customer service.
Some service providers have pointed out that effective AI customer service depends on two factors: the richness of the pre-set response pool and the timeliness of dynamic adjustments. "Taotian is currently only in a semi-AI state, where AI identifies users' general needs, while more in-depth questions are transferred to human agents," they said. Beyond digital humans and intelligent customer service, AI and large models hold broader application prospects in marketing, though they face short-term limitations due to technical bottlenecks and organizational constraints.
"After the organizational restructuring, Taotian's technical team has yet to take any action," an insider revealed, noting that Taotian needs to bridge the gap between technology and operations to achieve AI-driven e-commerce. "Internally, new features are evaluated based on either user engagement time or call rates, but there are no technical metrics to assess actual conversion rates or how conversions are achieved."
Additionally, Taotian's recommendation algorithms still have significant room for improvement. A simple example is that after a user purchases a product, similar items continue to be recommended—indicating that historical orders are not yet factored into the recommendation algorithm. This is likely constrained by computational resources. Earlier, when JD.com experimented with large models in e-commerce, it attempted to incorporate users' historical orders into shopping guidance. The vast amount of data and exponentially increasing computations made this feature costly and inefficient.
If incorporating historical orders presents such difficulties, solving multimodal problems to help merchants completely address product material and content operation issues is even more challenging.
There are countless paths for AI in e-commerce, but the core always revolves around "cost reduction and efficiency improvement." Only when costs and efficiency can help merchants release more gross profit space can comprehensive e-commerce platforms have a chance to compete with Pinduoduo's "low-price" strategy. "We've previously calculated that under the same conditions, the basic fees + marketing fees across platforms are actually quite similar. This means merchants' decisions on where to operate depend more on the business environment," said an industry insider.
Beyond "low prices," Pinduoduo has another advantage: low operational barriers for merchants. "The biggest issue with comprehensive e-commerce platforms used to be the complexity of price stacking—what can be stacked, what can be stacked in parallel, whether to offer membership prices, and even having to solve linear equations," the insider added.
Over the years, as pricing strategies have become increasingly layered, operational mishaps during major promotions have frequently occurred, leading to merchants being taken advantage of. Examples include the Guo Xiaoyun Taobao flagship store in 2019, the Changhong Kitchen Jingdong flagship store in 2020, the Genki Forest Tmall flagship store in 2021, and the Minitutu Tmall flagship store in 2023. Almost every year, there are cases where merchants suffer significant losses due to pricing errors. The problem facing cats and dogs may not be difficult. To achieve low prices, it is still necessary to solve merchants' operational issues. Wang Xing once learned a principle from a senior: there is only one way for retail to succeed—smart people doing hard work. However, the so-called AI e-commerce today seems like reverse engineering.