被冷落的 AI ，如何在 SaaS 领域“火”起来
How can AI, which has been left out, "fire" in the SaaS field
本文来自微信公众号“牛透社”（ID：Neuters），作者 倩茹。
This article comes from WeChat public number "Niu Tou She" (ID: Neuters), by Qian Ru.
随着 AI 技术应用的普及化，我们看到不少人工智能也广泛应用于 SaaS 管理软件，为行业打造标准化流程、提升效率方面起到很好作用。
With the popularization of AI technology application, we have seen that many artificial intelligence are also widely used in SaaS management software, which plays a very good role in creating standardized processes and improving efficiency for the industry.
文章整理自崔牛会和 SaaS 产品说联合举办的，关于“AI 在 SaaS 管理软件中的应用“的线上直播。
The article is compiled from the online live broadcast of "AI Application in SaaS Management Software" jointly organized by Cui Niuhui and SaaS Products.
主要内容
Main Contents
第一部分，AI 在 SaaS 管理软件领域的应用
The first part is the application of AI in SaaS management software
中美 SaaS 的对比分析
Comparative Analysis of SaaS between China and America
AI 在企业和行业的应用
Application of AI in Enterprises and Industries
第二部分，AI 在 SaaS 管理软件中的应用和未来趋势
The second part is the application and future trend of AI in SaaS management software
AI 的应用案例
Application Cases of AI
RPA 赛道场景应用和前途
Application and Future of RPA Track Scenario
数据掣肘
Data constraint
产品研发与落地中的瓶颈
Bottleneck in Product Research and Development and Landing
跨部门协作与跨领域协作
Cross-departmental collaboration and cross-domain collaboration
以下为正文，全文有删减。（内容未经相关嘉宾审核）
The following is the text, with deletions. (The content has not been reviewed by relevant guests)
01 AI 在 SaaS 管理软件领域的应用
Application of AI in SaaS Management Software
分享嘉宾：
Sharing Guests:
李一帆 品览 创始人&CEO
Founder & CEO of Li Yifan Pinlan
一、中美 SaaS 软件对比
I. Comparison of SaaS Software between China and the United States
1. 何为 SaaS
1. What is SaaS
首先，SaaS 概念出现在云计算之后，在没有云计算时，更多讲述的是软件，比如桌面软件、企业软件等。但是，在云计算出现之后，才开始有像服务一样的形式定义软件服务，这便是 SaaS。
First of all, SaaS concept appeared after cloud computing. When there is no cloud computing, it is more about software, such as desktop software and enterprise software. However, it was only after the emergence of cloud computing that software services began to be defined in the same form as services, which is SaaS.
最典型的 SaaS 应用，如 Dropbox、Slack、ZOOM、JIRA，基本都是一打开网页，通过 web 就可以访问的程序和服务。
The most typical SaaS applications, such as Dropbox, Slack, ZOOM and JIRA, are basically programs and services that can be accessed through the web once the web page is opened.
那 SaaS 形成的关键因素有哪些呢？我认为主要有三个：
What are the key factors for SaaS formation? I think there are mainly three:
第一，肯定离不开互联网。如果没有互联网，就不存在 SaaS，这是因为 SaaS 里所有的数据和程序都存在于服务器里，消费者与用户只有通过互联网才可以访问。
First, it must be inseparable from the Internet. If there is no Internet, there is no SaaS, because all data and programs in SaaS exist in servers, and consumers and users can only access them through the Internet.
第二，主要是消费级互联网的大发展。
Second, it is mainly the great development of consumer Internet.
提到这一点，就不得不说云计算。其实，云计算最早是因为 Amazon 开始做这样的服务。Amazon 把电商服务过程中空余的虚机拿出来做销售，卖给普通的开发者，让普通开发者用很便宜的价格买到服务器。因此，这些服务器最初就是从消费级的电商而来。而且，不管是消费级互联网，还是现在看到的企业微信，包括一些拥有着 To C 基因和影子的应用，都与此相关。所以，消费级互联网也是非常关键的驱动技术。
At this point, we have to say cloud computing. In fact, cloud computing was first started by Amazon. Amazon takes out the spare virtual machines in the e-commerce service process for sale and sells them to ordinary developers so that ordinary developers can buy servers at very low prices. Therefore, these servers originally came from consumer e-commerce. Moreover, both consumer-grade Internet and WeChat, which is now seen, including some applications with To C gene and shadow, are related to this. Therefore, consumer Internet is also a very critical driving technology.
第三，则是云计算本技术的发展。无论是从最开始的销售机器，还是到在机器上去销售一些平台型的应用和数据，或者在网上销售软件应用等，这都是比较重要的发展历程。
Third, it is the development of cloud computing technology. No matter from the beginning of selling machines, or to selling some platform-based applications and data on machines, or to selling software applications on the Internet, this is a relatively important development process.
最后，用一句话总结，SaaS 就是通过云端来使用的软件级服务。
Finally, to sum up in one sentence, SaaS is a software-level service used through the cloud.
2. 中美 SaaS 软件对比分析
2. Comparative Analysis of SaaS Software between China and America
如图，这是 BVP 研究机构做的关于云计算领域的整个情况，分为 SaaS、PaaS 和 IaaS。今天主要分享企业管理软件 SaaS。具体可从两个维度分析：
As shown in the figure, this is the whole situation of cloud computing done by BVP Research Institute, which is divided into SaaS, PaaS and IaaS. Today, I mainly share the enterprise management software SaaS. It can be analyzed from two dimensions:
维度1 按企业职能划分
Dimension 1 by Enterprise Function
举例来说，比如 Marketing 里面的 MailChimp。MailChimp 是湾区（旧金山湾区的）做 E-mail DEM 的一家公司，主要提供的SaaS服务是通过买MailChimp的账号，便可以群发邮件给客户，或者是感兴趣的对象，内容则可以是对公司服务的各种描述。
For example, MailChimp in Marketing. MailChimp is an E-mail DEM company in the Bay Area (San Francisco Bay Area). Its main SaaS service is to send mass emails to customers or interested objects by buying MailChimp's account number, and the content can be various descriptions of the company's services.
那什么是 Service呢？主要指的是售后、客服、收集反馈。
What is Service? It mainly refers to after-sales service, customer service and feedback collection.
再通过例子来看，比如 Uservoice。消费级互联网产品拥有很多的用户，但对于如何手机用户反馈，更多的产品以前只是选择留一个 support，也就是非常简单的留言窗口，做不到很好的互动性。因此，Uservoice 就是提供一个即插即用的用户论坛，反馈的插件可以给到企业级用户，服务于消费级互联网产品的 App 中去。
Let's look at an example, such as Uservoice. Consumer-grade Internet products have a lot of users, but for how to feedback from mobile phone users, more products used to only choose to leave a support, that is, a very simple message window, which could not be very interactive. Therefore, Uservoice is to provide a plug-and-play user forum, and the feedback plug-in can be given to enterprise users and serve the App of consumer Internet products.
再比如，在 HR 领域的 ZENEFITS。之所以 ZENEFITS 能在两三年时间，成为独角兽企业，最主要的原因是创始人切入的场景非常正确。ZENEFITS 选择帮助企业搭建一个给员工提供保险购买福利的 SaaS 软件。
Another example is ZENEFITS in HR field. The main reason why ZENEFITS can become a unicorn enterprise in two or three years is that the founder cut into the scene very correctly. ZENEFITS chose to help enterprises build a SaaS software that provides employees with insurance and purchase benefits.
当然，随着整个互联网的繁荣，也产生了很多协作的软件，或者是在 BI 和分析领域也有很多流行的 SaaS 软件，当然还有其他的。
Of course, with the prosperity of the entire Internet, there are also many collaborative software, or there are also many popular SaaS software in BI and analysis fields, and of course there are others.
维度2 从垂直角度去划分
Dimension 2 is divided vertically
先举第一例子，是美国的一家教育科技公司 EdmoDo，主要提供的是教育领域的内容管理和家校互动的功能。通俗来说，就是家长、老师还有孩子，可以在 App 上看内容和互动留言，但因为是教育行业的，会面临严格的审查，还得适应美国的教育体系。再比如，Othre 里的 Shopify，类似国内的有赞模式，用户通过 Shopify 可以开一个自己的网店或者店铺。
First of all, EdmoDo, an educational technology company in the United States, mainly provides the functions of content management and home-school interaction in the field of education. Generally speaking, parents, teachers and children can read content and interactive messages on App, but because they are in the education industry, they will face strict censorship and have to adapt to the education system in the United States. Another example is Shopify in Othre, which is similar to the domestic model of praise. Users can open their own online store or store through Shopify.
对美国 SaaS 软件的总结
Summary of SaaS Software in the United States
从行业和业务两个角度，通过一横一纵维度去观察 SaaS 领域的惯例软件可以发现，基本是利用了移动互联网和云计算技术，从而帮助客户完成了信息化、数据化。尤其是数据化，可以让用户行为、用户交易、企业内部数据、管理流程数据等，都留存在系统中。我主要有两点发现：
From the perspective of industry and business, observing the conventional software in SaaS field from one horizontal and one vertical dimension, we can find that mobile Internet and cloud computing technology are basically used to help customers complete informatization and digitization. In particular, digitization can allow user behaviors, user transactions, enterprise internal data, management process data, etc. to be retained in the system. I have two main findings:
观点1 美国 SaaS 领域中的 APP ，功能都是按照职能和垂直两个维度设计，拥有非常复杂且丰富的功能。但产品设计的重点在于能否化繁为简，尤其像 Box、Dropbox、Mixpanel 之类的软件，尽管需要存储各类数据，做各种各样的数据分析，但是产品却很好使用，很容易上手。
Viewpoint 1 APP in SaaS field in the United States is designed according to two dimensions of function and vertical, and has very complex and rich functions. However, the key point of product design is whether it can be simplified, especially for software such as Box, Dropbox and Mixpanel. Although it needs to store all kinds of data and do all kinds of data analysis, the product is very easy to use and use.
观点2上图标红的 SaaS 企业能做起来，是因为有一波很重要的用户。这些重要用户来源于消费互联网企业，一是这类企业基本上来自硅谷和世界 IT 公司，二是这类企业增长速度很快，团队发展很快，通过一些管理工具能够很好地管理团队，从而帮助企业减少一些不必要的职能岗位设置，还能更高效完成工作。
SaaS enterprises with red icons on Viewpoint 2 can do it because there are a wave of very important users. These important users come from consumer Internet enterprises. First, these enterprises basically come from Silicon Valley and world IT companies. Second, these enterprises are growing rapidly and their teams are developing rapidly. Through some management tools, they can manage their teams well, thus helping enterprises to reduce some unnecessary functional positions and complete their work more efficiently.
中国 SaaS 现状
Present Situation of SaaS in China
通过观察可以看到，美国在云计算领域，有 SaaS、PaaS、IaaS。在中国，大多数的企业大家都很熟悉，基本上如上图所示。根据我的观察，在职能划分和技术本质上，与美国相差不大，但是有一个缺点和一个优点。
Through observation, we can see that the United States has SaaS, PaaS and IaaS in the field of cloud computing. In China, most enterprises are familiar to everyone, basically as shown in the above figure. According to my observation, in terms of functional division and technology, it is not much different from that of the United States, but it has one disadvantage and one advantage.
缺点：中国 SaaS 企业，在管理理念和管理咨询上不如美国企业。这是因为中国企业文化太新，企业管理所有的根基都来自于西方世界，而美国等国家则具有完善的人力管理、财务管理理念，像 Salesforce、SAP 的 ERP 理念来自于现代管理学之父彼得德鲁克和营销领域的菲利普科特勒大师，因此，西方国家在这一领域就会很强，这是中国企业需要学习的。
Disadvantages: Chinese SaaS enterprises are not as good as American enterprises in management concepts and management consulting. This is because the Chinese corporate culture is too new, All the foundations of enterprise management come from the western world, while the United States and other countries have perfect concepts of human resources management and financial management. ERP concepts such as Salesforce and SAP come from Peter Drucker, the father of modern management, and Philip Kotler, the master of marketing. Therefore, western countries will be strong in this field, which Chinese enterprises need to learn.
优点：中国的消费互联网发展迅速，在市场规模、消费速度上，都是西方国家不能相比的。就是在这样的环境下，催生出中国的 SaaS。比如，在国内的电商领域有有赞以及其他的企业，但在美国板块只有 Shopify，这是因为国内本身具有的强社交关系链。
Advantages: China's consumer Internet is developing rapidly, which cannot be compared with western countries in terms of market size and consumption speed. It was in this environment that SaaS in China was born. For example, there are Zan and other enterprises in the domestic e-commerce field, but only Shopify in the U.S. Sector, because of the strong social relationship chain in the country itself.
以此来看，SaaS 是更广的一个概念，SaaS 是将软件当作服务。而管理软件是根据职能或者行业进行划分，从而对公司进行有效管理，其核心是输出管理理念。
From this point of view, SaaS is a broader concept. SaaS regards software as a service. The management software is divided according to functions or industries, thus effectively managing the company. Its core is to export management concepts.
二、AI 的应用
II. Application of AI
1. AI 的概念
1. The concept of AI
AI 发展的三个阶段
Three Stages of AI Development
第一阶段，1956年夏季，以麦卡赛、明斯基、罗切斯特和申农等为首的一批有远见卓识的年轻科学家在一起聚会，共同研究和探讨用机器模拟智能的一系列有关问题，并首次提出了“人工智能”这一术语。这个阶段，基本上是各种理论性的论文和科学研究。
In the first stage, in the summer of 1956, a group of visionary young scientists, led by McCarthy, Minsky, Rochester and Shennong, gathered together to study and discuss a series of related issues concerning the simulation of intelligence by machines, and put forward the term "artificial intelligence" for the first time. This stage is basically a variety of theoretical papers and scientific research.
第二阶段，到了80年代，随着计算机的出现，产生了机器学习，能够做一些数据处理。
In the second stage, in the 1980s, with the emergence of computers, machine learning came into being and was able to do some data processing.
第三阶段，主要指的是2010年之后，由于 GPU 技术的大力发展，有了机器的深度学习，这才有了现在 AI 大爆发。
The third stage mainly refers to the AI explosion after 2010 due to the vigorous development of GPU technology and the in-depth learning of machines.
AI 的四个门类
Four Categories of AI
门类1：机器学习和深度学习
Category 1: Machine Learning and Deep Learning
虽然 NLP、CV、Forecasting 都属于机器学习的一部分，但这里特指的是分析结构化的数据，从而找到规律进行预测，通常应用在反欺诈、股价预测、温度预测方面。
Although NLP, CV and Forecasting are all part of machine learning, this refers specifically to the analysis of structured data to find rules for prediction, which is usually applied to anti-fraud, stock price prediction and temperature prediction.
门类2：自然语言处理
Category 2: Natural Language Processing
这就是让机器能够像人一样去理解语言和文字，能够从中挖掘信息进行决策。
This is to enable machines to understand languages and characters like human beings and to mine information from them for decision-making.
门类3：CV（Computer Vision）
Category 3: CV (Computer Vision)
这是让机器通过相机，能够像人一样去处理看到的数据，机器看到的通常的是图片、视频等，从而进行推断分析。
This is to enable the machine to process the data it sees like a human through a camera. What the machine sees is usually pictures, videos, etc., thus making inference and analysis.
门类4：Forecasting&Optimization
Category 4: Forecasting & Optimization
这属于比较专的一个AI领域，主要是对平台进行优化。应用的领域，比如让滴滴能及时调度几千万的司机，让电商平台可以管理发货速度等。
This belongs to a relatively specialized AI field, mainly optimizing the platform. Application fields, such as enabling Didi to dispatch tens of millions of drivers in time and enabling e-commerce platforms to manage delivery speed, etc.
数据的重要性
Importance of data
正如上图所示，明斯基定的目标非常宏大。可现在的 AI 技术即使到了第三个阶段，但离真 AI 还是非常遥远，幸运的是拥有庞大的数据。
As shown in the above figure, Minsky's goal is very ambitious. However, even if the current AI technology has reached the third stage, it is still very far from real AI. Fortunately, it has huge data.
那么，在 AI 领域的数据是什么呢？其一，数据是AI算法的养料，通过经验可以告知我们一件事情发生的规律，还可以帮助人将世界具象化，排除一些不确定的条件。
So, what is the data in AI? First, data is the nourishment of AI algorithm. Experience can tell us the law of one thing's occurrence, and it can also help people to visualize the world and eliminate some uncertain conditions.
举例来看，如图4所示，左图是微软非常出名的聊天机器人小冰，人类可以跟它聊天气、聊人文地理、聊哲学。但是，对于它不知道的游戏世界，人说再多它也听不懂。
For example, as shown in Figure 4, the left picture shows Microsoft's very famous chat robot Xiao Bing, with which human beings can talk about weather, human geography and philosophy. However, for the game world that it does not know, it cannot understand no matter how many people say.
右图则是 Google 非常出名的项目 Waymo，这是一个智能驾驶的车，曾获得大额融资，且确实离无人出租车越来越近。但如果将其放置于中国的道路上，对它来讲就是很复杂的条件，就需要更多的经验和数据才能来处理。
The picture on the right shows Google's very famous project Waymo, which is an intelligent driving car that has received large amount of financing and is indeed getting closer and closer to unmanned taxis. However, if it is placed on China's road, it is a very complicated condition for it and requires more experience and data to process it.
从这两个例子，都足以看出数据对于 AI 的重要性。
From these two examples, it is enough to see the importance of data to AI.
2. 企业和行业中的 AI 应用
2. AI Application in Enterprises and Industries
AI in SaaS 属于应用层
AI in SaaS belongs to the application layer
如图所示，最下面是数据源和 API，以及一些开源的框架。这类型企业基本上做最基础的芯片，或者是做算力的，或者是做基础数据的收集。
As shown in the figure, the bottom are data sources and API, as well as some open source frameworks. This type of enterprise basically does the most basic chip, or does the calculation power, or does the basic data collection.
往上一层是很多开源的企业，但开源并不代表就是免费。这类型企业通常都做得很大，营收规模也很大。这个空间主要是给开发者和 AI 行业去提供工具，所以这层是工具层。
On the next level are many open source enterprises, but open source does not mean free. This type of enterprise usually does a lot of work and has a large revenue scale. This space is mainly to provide tools for developers and AI industry, so this layer is the tool layer.
再往上，左边是偏技术本身的，包括云计算、大数据处理和数据库相应的服务。中间是垂直 AI 算法站，包括做数据分析、BI、CV、NLP 语言处理和语音等等。
Further up, on the left is the technology itself, including cloud computing, big data processing and database-related services. In the middle is the vertical AI algorithm station, including data analysis, BI, CV, NLP language processing and speech, etc.
那么，企业软件位于什么地方呢？SaaS 类的 AI或 AI in SaaS 其实属于应用层。
So, where is the enterprise software located? The SaaS class AI or AI in SaaS actually belongs to the application layer.
往常很多人会说做 SaaS 类 AI 公司没有技术，这显然是错的，只是其所处行业就是在应用层，可帮助企业解决具体问题，更多关注的是应用场景。
In the past, many people would say that SaaS AI companies have no technology, which is obviously wrong, but their industry is in the application layer, which can help enterprises solve specific problems and pay more attention to application scenarios.
图中还有一些垂直类型的企业，在我看来，垂直行业类的 AI 企业有很强的应用场景，出现的也很早。
There are also some vertical enterprises in the picture. In my opinion, AI enterprises in vertical industries have strong application scenarios and appear very early.
它们主要是在指纹识别、人脸识别领域的公司，这类公司在上世纪90年代之前就出现了，但只是在做技术的早期积累和一些早期客户的服务。之后，因为云计算和开源的出现，才有了越来越多的 AI 平台，也就出现了更多的 AI 公司，面向更细分和更小众的行业人群。
They are mainly companies in the fields of fingerprint recognition and face recognition. Such companies appeared before the 1990s, but they are only doing early accumulation of technology and services for some early customers. Since then, due to the emergence of cloud computing and open source, there have been more and more AI platforms and more AI companies, targeting more subdivided and smaller industry groups.
我认为，向垂直细分类发展的趋势是对的。因为人工智能就如同社会，需要各种分工，应该每个行业都有一些特别专业的人，未来 AI 一定是会这样慢慢去超这一趋势发展的。
In my opinion, the trend towards vertical subdivision is right. Because artificial intelligence, like society, requires a variety of division of labor, and there should be some special professionals in each industry, AI will certainly slowly surpass this trend in the future.
AI 和 SaaS 结合的发展趋势
Development Trend of AI and SaaS Combination
根据 Gartner Hype Cycle 2019年在 AI 领域的数据显示，GPU 基本已经到了非常成熟的阶段，当然也包括语音识别。
According to Gartner Hype Cycle's 2019 AI data, GPU has basically reached a very mature stage, including speech recognition.
从图中还可以看出，自动驾驶其实处于一个很低迷的时期，尽管前几年炒得很火热，但目前还是比较困难，相信曙光就在眼前。
It can also be seen from the figure that self-driving is actually in a very depressed period. Although the speculation was very hot in the past few years, it is still relatively difficult at present. I believe the dawn is just around the corner.
再看聊天机器人，AutoML 等公司都处于最高处，都在自动建模。还有在上升阶段的 Neuromorphic Hardware，是一个新型的芯片，它是用神经拟态在做计算，不需要走传统 x86 的框架，就可以直接提供更高阶的 AI 模型的建模能力。
If you look at chat robots, AutoML and other companies are at the highest level and are modeling automatically. There is also Neuromorphic Hardware, which is in the rising stage. It is a new type of chip. It uses neural mimicry to do calculations. It can directly provide higher-order AI model modeling capability without taking the traditional x86 framework.
最后，再看云计算的发展趋势。云计算包括混合云、私有云、容器部署管理等，还有现在比较新的，如区块链、无服务器云计算、边缘计算，这里都有很多新的需要学习的技术。
Finally, look at the development trend of cloud computing. Cloud computing includes hybrid cloud, private cloud, container deployment management, etc. There are also relatively new technologies, such as block chain, serverless cloud computing and edge computing. There are many new technologies to learn.
随着技术的发展，一定还会有一波新的浪潮。要赢在当下，但是也要抓住未来，当你在非常困难的时候，不要太在意当下的困难，要努力把公司推到上升期，未来也许就在眼前。
With the development of technology, there will certainly be a new wave. Win in the present, but also seize the future. When you are in very difficult times, don't pay too much attention to the current difficulties. Try to push the company to a rising period. The future may be just around the corner.
02 AI 在 SaaS 管理软件中的应用和未来趋势
Application and Future Trend of AI in SaaS Management Software
主持人
Moderator
l 孙思远 SaaS 产品说特邀嘉宾
L Sun Siyuan SaaS Product Says Special Guest
对话嘉宾：
Dialogue Guest:
l 林松 纷享销客总裁兼 CTO
L Lin Songfen, President and CTO of Sales
l 陈运文 达观数据创始人兼 CEO
L Chen Yunwen, Founder and CEO of Daguan Data
l 王守 爱因互动创始人兼 CEO
L Wang Shouaiyin Interactive Founder and CEO
l 陈宏志 用友数据智能产品总监
L Chen Hongzhi, Director of UFIDA Data Intelligence Products
话题一 AI 在不同企业中的应用场景和应用案例
Topic 1 Application Scenarios and Cases of AI in Different Enterprises
李一帆：品览是做物品识别的，属于计算机视觉领域的一种。现在主要服务三类企业：零售快消、地产行业、物流行业。
Li Yifan: Product viewing is used to identify objects and belongs to the field of computer vision. At present, it mainly serves three types of enterprises: retail fast-disappearing, real estate industry and logistics industry.
我们的理念是通过物品识别，帮助企业在物品设计、生产、物流和零售四个环节采集数据，再将数据交给流程自动化的 SaaS 软件完成分析。
Our idea is to help enterprises collect data in the four links of article design, production, logistics and retail through article identification, and then hand over the data to SaaS software with process automation for analysis.
林松：纷享销客属于 SaaS 领域中的 CRM，核心是跟企业的营销、销售相关。我们也正在尝试把 AI 与产品本身相结合，从而给客户提供更高效的使用方式，帮助客户更好地去洞察对应的经营数据。
Lin Song: Enjoying customers belongs to CRM in SaaS field. The core is related to the marketing and sales of enterprises. We are also trying to combine AI with the product itself, so as to provide customers with more efficient usage methods and help customers better gain insight into the corresponding business data.
我们正在做且已经有初步成果的案例
Cases that we are doing and have achieved initial results.
案例1 关于销售
Case 1 About Sales
在销售领域，企业最关心的就是如何获客。典型的销售过程是，需要通过获得很多销售线索，再让销售跟进对应的销售线索，并对销售线索进行对应的跟进和转化，逐步转换成商业机会，最后推进到订单和成交的过程。
In the field of sales, enterprises are most concerned about how to obtain customers. The typical sales process is that it is necessary to obtain a lot of sales leads, then let the sales follow up the corresponding sales leads, and carry out corresponding follow-up and transformation on the sales leads, gradually convert them into business opportunities, and finally push forward to the process of orders and transactions.
而在获取线索上，对于很多企业而言，并不是跟一些企业反复有生意往来，只能不断获得新客户，从而提升交易量，所以获客就是这个领域里的刚需。
As for obtaining clues, for many enterprises, they do not have repeated business dealings with some enterprises, but can only continuously obtain new customers, thus increasing the transaction volume. Therefore, obtaining customers is the urgent need in this field.
在我们平台上，可以提供一个“线索推荐”的产品。简单来看，企业通过在平台上使用产品，就可以对已经成交的客户有更多了解，当平台上的成交客户量到了一定级别，就可以通过对应的模式和算法，对企业的成交客户画像进行相应分析。在有了具体分析数据之后，企业可以跟整个库里的客户数据进行匹配，通过这种方式推荐给企业对应的客户。
On our platform, we can provide a "clue recommendation" product. In a nutshell, by using products on the platform, enterprises can know more about the customers who have already made transactions. When the number of customers who have made transactions on the platform reaches a certain level, they can make corresponding analysis on the portrait of the customers who have made transactions through corresponding modes and algorithms. After having specific analysis data, the enterprise can match the customer data in the entire library and recommend it to the corresponding customers of the enterprise in this way.
但是，对于很多行业而言，不仅仅是客户推荐，重要的是能够获取跟客户相关的资源和人脉。因此，在这一点上，我们行业正在进行场景的探索。
However, for many industries, it is not only customer recommendation, but also the ability to obtain customer-related resources and contacts. Therefore, at this point, our industry is exploring the scene.
案例2 快消行业
Case 2 Fast Elimination Industry
在零售快消行业，大量的营销工作是通过向货架铺货来完成的，有一个访销的过程，这就产生了对货架盘点的需求。因此，越来越多的企业希望通过 AI 的方式进行智能货架盘点，经过分析之后，就知道如何改善和提升自己的产品，进而促进销量的提高。同时，还提供了图像识别，防偷拍，防作弊。
In the retail fast-moving industry, a large amount of marketing work is completed by distributing goods to shelves, and there is a process of visiting and selling, which generates the demand for shelf inventory. Therefore, more and more enterprises hope to conduct intelligent shelf inventory through AI. After analysis, they will know how to improve and upgrade their products, thus promoting the increase of sales volume. At the same time, it also provides image recognition, anti-candid photography and anti-cheating.
简而言之，通过 AI 技术，既可以提升终端工作人员的工作效率，也能让公司更好地监督一线人员，促进商业行为的完成。
In short, AI technology can not only improve the work efficiency of terminal staff, but also enable the company to better supervise front-line personnel and promote the completion of business activities.
陈宏志：用友不在于数据量大，而在于拥有的小样本，以及小样本多模态的 SaaS 服务。
Chen Hongzhi: UFIDA does not lie in the large amount of data, but in the small samples it has and the multi-modal SaaS service with small samples.
在这样的情况下，基于用友云，实际上是数据治理体系+大数据平台，构成数据中台的整个能力。基于多年的行业积累，用友服务的企业多为中型和大型客户，因此我们有了公有云、专属云和私有云。这是因为大型企业是部分私有，也就是价值链的某些点可以做公有化。
Under such circumstances, based on UFIDA cloud, it is actually a data governance system + big data platform, which constitutes the entire capability of the data platform. Based on years of industry accumulation, UFIDA serves mostly medium-sized and large customers, so we have public cloud, exclusive cloud and private cloud. This is because large enterprises are partially privately owned, that is, some points in the value chain can be publicly owned.
同时，在企业内部整个价值链来说，像制造型企业的生产经营、制造的过程，与公网是物理隔离的状态，但采购、营销这些可以在公有云上订阅 SaaS 服务，因此混合云的模式，也成为了满足客户需求的部署。
At the same time, in terms of the entire value chain within the enterprise, the production, operation and manufacturing processes of manufacturing enterprises are physically isolated from the public network, but purchasing and marketing can subscribe to SaaS services on the public cloud, so the hybrid cloud mode has also become the deployment to meet the needs of customers.
在有了大量数据的情况下，用友并不是以数据量大取胜，也不是以技术上的一些优势取胜。
With a large amount of data, UFIDA did not win with a large amount of data, nor did it win with some technical advantages.
对于企业级用户，尤其是面向有一定传统基因、传统血液的企业而言，我们是通过对业务、对场景的理解，往下进一步深耕的，比如像 CV 技术、NLP、语音交付技术，我们会对它进行一个集成。
For enterprise users, especially for enterprises with certain traditional genes and traditional blood, we have further deepened our understanding of business and scenes, such as CV technology, NLP and voice delivery technology, and we will integrate it.
第一个突破点是开发智能机器人、智能收单、智能报账 RPA，包括从发票到 OCR 保障开户。集成的是技术，核心其实是整个流程的自动化，也就是从整个报账过程，从填报开始到最后报表生成，甚至于最后审查整个流程自动化，都可以通过 RPA 来实现。
The first breakthrough point is the development of intelligent robots, intelligent receipt and intelligent reimbursement RPA, including from invoices to OCR to guarantee account opening. The integration is technology, and the core is actually the automation of the whole process, that is, the automation of the whole process from the beginning of reporting to the generation of the final report, and even the final review, can be realized through RPA.
另一个突破点是 VPA 产品，可以根据对话交互一系列场景自动进行场景识别，支持财务月结、发票查验、发票认证、采购三单匹配、月结检查、内部交易对账、久其预算报表填报等。
Another breakthrough point is VPA products, which can automatically identify scenes according to a series of scenes of dialogue and interaction. It supports monthly financial settlement, invoice inspection, invoice authentication, matching of purchase orders, monthly settlement inspection, internal transaction reconciliation, Join-Cheer budget report filling, etc.
另外，在营销和采购侧也有部署。营销侧，主要是运用 AI 技术，对库存消耗情况、损失情况、备货情况进行预测，帮助其更好节约成本和提升销量。采购侧，主要针对工业物料，基于企业级和行业级的知识图谱，更好把物料进行归类管理，尤其是在招投标时，运用企业画像体系，可以在审核过程中节省很多人力成本。
In addition, there are also deployments on the marketing and purchasing sides. On the marketing side, AI technology is mainly used to predict inventory consumption, loss and stock-up to help them save costs and increase sales. On the purchasing side, it is mainly aimed at industrial materials. Based on enterprise-level and industry-level knowledge maps, materials can be better classified and managed. Especially in bidding, the application of enterprise portrait system can save a lot of labor costs in the audit process.
陈运文：达观数据是做自然语言处理服务的。自然语言处理的一个特点，就是与各行各业都相关。在每个行业运用 AI 服务时，都需要在采集大量行业样本数据，这是需要长期训练的。况且自然语言处理技术本身需要和其他业务流程以及场景串联在一起，才能形成一个完整的应用链条。
Chen Yunwen: Philosophical Data is a natural language processing service. One of the characteristics of natural language processing is that it is related to all walks of life. When AI services are used in each industry, a large number of industry sample data need to be collected, which requires long-term training. Besides, natural language processing technology itself needs to be connected in series with other business processes and scenarios to form a complete application chain.
所以，在研发达观自然语言处理产品时，基于上述所说做了很好的考量，尽管应用场景很多，但不是所有场景都要去做，就像市场上已经有很多企业开发聊天机器人、对话机器人客服场景，因此在这一领域我们就不再布局。
Therefore, when developing philosophical natural language processing products, good considerations have been made based on the above. Although there are many application scenarios, not all scenarios have to be done, just like many enterprises in the market have developed chat robots and dialogue robot customer service scenarios, so we will no longer lay out in this field.
达观的主要精力集中在书面文字资料的审核、填写、抽娶搜索、推荐等应用场景，而且书面文字处理是一个细分的 NLP 赛道，这些场景有大量的需求，因此市场机会还是很大的。
The main focus of resilience is on the application scenarios such as the review, filling, selection, search and recommendation of written text materials. Moreover, written text processing is a subdivided NLP track. These scenarios have a large demand, so the market opportunities are still great.
同时，NLP 本身需要和客户的系统做串联，才能形成真正的价值链条。在 NLP 的基础上，我们也开发了自己的 RPA 产品。不同于其他友商的产品，达观 RPA 是把文本处理模块和客户其他的业务操作系统连接在一起，用 RPA 做连接，这样可以很好地让处理系统像数字化园丁一样去完成日常工作。
At the same time, NLP itself needs to be connected in series with the customer's system to form a real value chain. On the basis of NLP, we have also developed our own RPA products. Different from the products of other friends, RPA connects the text processing module with other business operating systems of customers and uses RPA to connect, which can make the processing system complete daily work like a digital gardener.
这些工作既跟财务相关，也包含人事、供应链、物流等场景下的应用。而且，这些场景下的应用都有两个特点：一，需要重复性的大量文档；二，这些应用和客户原有的 ERP、OA 等需要打通链接，才能形成一个完整的操作链条。
These tasks are not only related to finance, but also include applications in personnel, supply chain, logistics and other scenarios. Moreover, the applications in these scenarios have two characteristics: first, a large number of repetitive documents are required; Second, these applications and customers' original ERP, OA, etc. need to get through links to form a complete operation chain.
因此，达观开发的数字化园丁系统，能够为企业降本增效，这也是我们产品的出发点。
Therefore, the digital gardener system developed by Philosophy can reduce costs and increase efficiency for enterprises, which is also the starting point of our products.
总结一下，达观是以自然语言处理为切入点，结合 RPA 等给客户提供数字化员工的企业，以 SaaS 或其他私有化的方式来给客户提供服务，目的是给客户降本增效，提高企业运转效率。
To sum up, resilience takes natural language processing as the breakthrough point, combines RPA and other enterprises that provide digital employees to customers, and provides services to customers by SaaS or other privatized methods, with the aim of reducing costs and increasing efficiency for customers and improving the operation efficiency of enterprises.
我认为，在中国的人工智能应用市场化，未来还有很多机会，但目前面临的挑战也非常多。如何让 AI 技术实实在在落地应用起来，这与产业链的数据、场景是否具备能力、SaaS 这种服务形态下的周期长短，以及数据的安全私密性、在线服务的稳定性，息息相关。
In my opinion, there are still many opportunities for the marketization of artificial intelligence applications in China in the future, but there are also many challenges at present. How to make AI technology apply to the ground is closely related to the data of the industrial chain, whether the scene has the capability, the cycle length under SaaS, the security and privacy of the data, and the stability of online services.
王守：爱因互动是一家做企业级智能交互的公司。也就是所谓的 conversation as a service，通常叫做对话即服务。
Wang Shou: Aiyin Interactive is a company that does enterprise-level intelligent interaction. The so-called conversation as a service is usually called conversation as a service.
这个行业为企业提供的是整体服务，服务核心是和企业客户对话。我们的产品则是将对话落到人身上，属于 NLP 细分领域。我们的服务场景是为企业和企业的客服、企业的售前销售，以及企业用户和员工之间发生交互、发生对话的场景。
This industry provides enterprises with overall services, and the core of service is dialogue with enterprise customers. Our products fall on people for dialogue, which belongs to NLP subdivision. Our service scenario is for the customer service of enterprises and enterprises, the pre-sales sales of enterprises, and the interaction and dialogue between enterprise users and employees.
具体应用上，在客服领域，全中国大概有800万至1000万的客服人员，辅助和替代相关客服人员，应该是比较大的市常在销售领域，某些销售事实上可以用机器对话去替代，不能用机器对话去替代的可以让对话机器人去扮演客户，训练销售话术和对产品知识掌握程度，这是一个很有意思的方向，也是目前在我们业务里增长迅速的领域。
In specific application, In the field of customer service, There are about 8 to 10 million customer service personnel in China, Assist and replace relevant customer service personnel, It should be that relatively large markets often live in the sales field. In fact, some sales can be replaced by machine dialogue. What cannot be replaced by machine dialogue can let dialogue robots play the role of customers, train sales skills and master product knowledge. This is a very interesting direction and is also an area that is growing rapidly in our business at present.
与其他做自然语言处理企业不同的是，要为客户建立起基于他的客户、用户、潜在销售线索的整体画像，从对话中建立多维度的更加丰富的动态的用户画像。
Different from other natural language processing enterprises, it is necessary to establish an overall portrait of customers, users and potential sales leads for customers, and to establish multi-dimensional and richer dynamic user portraits from dialogues.
比如，在做辅助销售过程中，每跟客户说一句话，后台对用户的画像就会更加完善，就能对用户的意图有更好的判断，而后对于产品的推荐，或对于问题的回答就会有一个更好的、更精准的判断。
For example, in the process of auxiliary sales, every time you say a word to the customer, the backstage portrait of the user will be more perfect, and you will have a better judgment on the user's intention, and then you will have a better and more accurate judgment on the product recommendation or the answer to the question.
以上就是我们的产品定位，用自然语言处理技术（NLP），包括知识图谱技术、推荐技术等服务企业级客户，为企业提供智能交互服务。
The above is our product positioning. We use Natural Language Processing Technology (NLP), including knowledge mapping technology and recommendation technology, to serve enterprise customers and provide intelligent interactive services for enterprises.
同时，我们的产品都是以 SaaS 化形式去进行。对于有合规要求的，进行了私有云部署，对于一些没有合规性要求的，一直在推 SaaS 模式。
At the same time, our products are all carried out in SaaS form. For those with compliance requirements, private cloud deployment has been carried out, while for those without compliance requirements, SaaS mode has been pushed.
最后，分享一点心得。坦白来讲，无论是对话，还是自然语言处理，其实离真正的智能还有非常大的距离。所以，我们定义了一个指标，叫人工替代率，就是在某一个场景下如果用机器的对话能够替代掉多少人工。
Finally, share some experience. Frankly speaking, whether it is dialogue or natural language processing, it is still a long way from real intelligence. Therefore, we have defined an indicator, called the manual substitution rate, which is how much manual can be replaced by machine dialogue in a certain scene.
对于这个定义，是根据数据来分类的。如果人工替代率只能达到30%，没有办法直接上线做，就采用辅助的方式；如果人工替代率达到70%以上，那此场景就可以考虑使用以机器人为主、人工为辅的方式；如果场景替代率到了90%、95%以上，则可以完全用机器人替代。
For this definition, it is classified according to data. If the manual replacement rate can only reach 30% and there is no way to do it directly online, the auxiliary method will be adopted. If the manual replacement rate reaches more than 70%, then this scene can consider using robots as the main and manual as the auxiliary. If the scene substitution rate reaches 90% and 95%, it can be completely replaced by robots.
但很遗憾的是，目前能够完全替代90%、95%以上的产品非常少，绝大部分在30%-70%之间，所以这就需要做一些梳理或产品的设计，使对话机器人的产品能够更好地辅助客户。
However, unfortunately, there are very few products that can completely replace 90% and more than 95%, most of which are between 30% and 70%. Therefore, it is necessary to do some sorting or product design so that the products of dialogue robots can better assist customers.
话题二 RPA（Robotic Process Automation）的场景应用及未来发展前途
Topic 2: Scene Application and Future Development of RPA (Robotic Process Automation)
陈宏志：用友在这一领域，应该说在2017年就有了 RPA 的前身，那就是语音助手小友。当时，用友一直在做财务软件，因为移动端的加持，企业经营决策者提出一个诉求，无论何时何地，可以用一句话把平时常看的表、图，用不超过三句话的语言迅速查询到。在这种情况下，用友便做了语音助手小友。
Chen Hongzhi: UFIDA has the predecessor of RPA in 2017 in this field, that is, voice assistant Xiaoyou. At that time, UFIDA had been making financial software, because of the blessing of the mobile terminal, the business decision makers of the enterprise put forward a demand, no matter when and where, they can use one sentence to quickly query the tables and maps they usually look at in a language of no more than three sentences. In this case, UFIDA became a voice assistant friend.
后来，随着协同办公软件友空间、友云采、友云售等公有云产品生态建立起来，我们发现 RPA 与传统 AI 技术和传统模式识别技术结合以后，能够实现一系列流程自动化。
Later, with the establishment of public cloud product ecology such as collaborative office software Youspace, Youyun Mining and Youyun Sales, we found that RPA can realize a series of process automation after combining with traditional AI technology and traditional pattern recognition technology.
以前，不论是 MES，还是 ERP，用友传统信息化系统的很多都是手工完成，NLP 技术相对弱一些，但是在流程自动化和流程智能化方面相对强一些。因此，将两者结合，再加上前端的语音交互机器人，就可以把整个报账、审批、银企直联，以及现在在工业现场经常会用到的智能报单等一系列应用，全部结合起来。然后，用一个统一的设计封装，就等于把 AI 技术包装成了一个结合用友云特色场景去做的一系列机器人。
In the past, many of UFIDA's traditional information systems, whether MES or ERP, were completed manually. NLP technology was relatively weak, but it was relatively strong in process automation and process intelligence. Therefore, the combination of the two, coupled with the front-end voice interactive robot, can combine the whole reimbursement, approval, bank-enterprise direct connection, as well as a series of applications such as intelligent declaration forms, which are often used in industrial sites. Then, using a unified design package is equivalent to packaging AI technology into a series of robots combined with UFIDA cloud characteristic scenes.
在从 SaaS 不断细分之后，我们最终是朝着群体智能方向发展的。现在产品已经开始支持机器人打群架，也就是说机器人式的机器人协作。
After continuous subdivision from SaaS, we finally developed towards swarm intelligence. Now the products have begun to support robot group fighting, that is to say, robot-style robot cooperation.
比如，在供应链金融产品中，可以通过机器调查被粉饰过的数据，再根据我们建的模型，把报表进行还原重现，最后给银行的是一份相当于企业级的健康体检报告，而这个健康体检报告是机器人自己写出来的。这就相当于用一个 RPA，把整个完整的贷款融资审批的报告生成过程生产出来。
For example, in supply chain financial products, the whitewashed data can be investigated by machines, and then the report forms can be restored and reproduced according to the model we have built. Finally, the bank is given an enterprise-level health examination report, which is written by the robot itself. This is equivalent to using an RPA to produce the entire report generation process for loan financing approval.
陈运文：RPA 的本质是模拟人在键盘和鼠标上的操作，软件安装在电脑上以后，可以自动接管键盘和鼠标输入的操作。
Chen Yunwen: The essence of RPA is to simulate people's operations on the keyboard and mouse. After the software is installed on the computer, it can automatically take over the operations of keyboard and mouse input.
比如，让机器人完成审核合同、阅读财务报表、阅读简历、阅读公司业务文档资料，然后对这些文档资料进行填报、审核、处理等相关工作。原先这些工作都是由人来做，效率比较低，现在可以让计算机来完成。类似的场景还有非常多，尤其是财务、人事领域，有很多工作都可以让机器人替代。
For example, let the robot complete the review of contracts, reading financial statements, reading resumes, reading the company's business documents, and then fill in, review, process and other related work on these documents. Originally, all these tasks were done by people, which was relatively inefficient. Now computers can do them. There are many similar scenes, especially in the fields of finance and personnel. Robots can replace many jobs.
因此，达观要做的是一方面降低各行各业使用 RPA 的门槛，让使用 RPA 系统的客户以非常简单的方式开发个机器人，让机器人上线运行。这是一个让 AI 使用门槛降低，让更多潜在使用方以非常便捷的方式，开发出一个满足自己所需要的程序。这非常考验产品能力，但也是未来技术发展方向。
Therefore, what resilience needs to do is, on the one hand, to lower the threshold for all walks of life to use RPA, so that customers using RPA system can develop a robot in a very simple way and let the robot run online. This is a very convenient way to lower the AI usage threshold and allow more potential users to develop a program that meets their needs. This is a great test of product capability, but it is also the direction of future technological development.
也就是说，降低 AI 使用门槛，让客户以最简便的方式，如拽一个浏览器的框、在框里输入网址，机器人就操作，这种用无代码或者代码的方式，是今天 RPA 应用的主流。
That is to say, lowering the AI usage threshold and allowing customers to operate in the simplest way, such as pulling a browser box and inputting a web address in the box, the robot will operate. This way of using no code or code is the mainstream of RPA application today.
另外，我们 RPA 里集成了一些自然语言处理或 OCR(Optical Character Recognition,光学字符识别)的模块，把它做得尽量简单易用，让使用方直接拖拽一个 OCR 操作流程，就能自动识别商业票据。
In addition, we have integrated some natural language processing or OCR (Optical Character Recognition) modules in RPA, making it as simple and easy to use as possible, so that users can automatically recognize commercial bills by dragging an OCR operation flow directly.
王守：关于 RPA，我们主要体现在交付上。不过，我们的 RPA 其实也是服务于对话机器人的。
Wang Shou: As for RPA, we mainly embody it in delivery. However, our RPA actually serves dialogue robots.
比如，在保险行业做保险产品，人身险产品的保险责任和条款分拣让机器人完成，其实是服务于后续关于保险问答的对话机器人。或者是一份保险产品的文档在后台上传之后，机器人就会自动把知识和保险责任拆解出来，同时给到前端的对话机器人，对话机器人就能够使用其来回答相关问题。
For example, when making insurance products in the insurance industry, the sorting of insurance liabilities and clauses for personal insurance products is completed by robots, which are actually dialogue robots serving the follow-up questions and answers on insurance. Or after a document of an insurance product is uploaded on the back stage, the robot will automatically disassemble the knowledge and insurance liability and give it to the dialogue robot at the front end, which can use it to answer relevant questions.
还有一种机器人，是给企业的销售做质检，因为很多公司销售往往是在微信上和客户进行沟通，其实这对质检带来很大的障碍。我们的一款RPA 机器人，能够直接在手机和电脑上进行截图，将其变成一张张图片，再通过 OCR，还原成相关的对话，同时进入到对话机器人里，来进行质检一些用户画像抽娶销售话术合规性检查等。
There is also a kind of robot, which does quality inspection for the sales of enterprises, because many companies often communicate with customers on WeChat, which actually brings great obstacles to quality inspection. One of our RPA robots can directly take screenshots on mobile phones and computers, turn them into pictures, and then restore them to relevant dialogues through OCR. At the same time, it can enter the dialogue robot to carry out quality inspection, some user portraits, smoking and sales speech compliance checks, etc.
话题三 数据的结构化
The Structure of Topic 3 Data
林松：做人工智能（AI），数据是根本。如果脱离了数据，很多东西其实做出来有一种“无米之炊”的感觉。
Lin Song: To be an artificial intelligence (AI), data is fundamental. If we are separated from the data, many things actually feel like "cooking without rice".
刚才谈到的两个场景，可以说一下所需要的数据。比如，在给客户推荐线索时候，包括给客户自己录入到系统里的数据进行打分，需要的数据是超出我们本身系统里用户使用产生的数据。还有，刚才谈到要给客户提供推荐的线索，很多推荐的线索并不是把 A 客户的数据推荐给 B 客户，而是需要全中国企业的工商数据，并加入更多企业经营的维度分析。
For the two scenarios just mentioned, we can talk about the required data. For example, when recommending clues to customers, including scoring the data entered into the system by customers themselves, the data needed are beyond the data generated by users in our own system. Also, just now we talked about providing recommended clues to customers. Many recommended clues do not recommend customer A's data to customer B, but need industrial and commercial data of all Chinese enterprises and add more dimensional analysis of enterprise operations.
因此，当我们做数据产品时，第一件事就是要跟现在拥有工商数据的平台进行合作，像企查查、天眼查、启信宝等平台。通过这样平台，可以获得企业完整信息。同时，根据产品核心特色，针对所需要的特征值再去补充对应的企业所需要的细节信息。
Therefore, when we make data products, the first thing to do is to cooperate with platforms that now have industrial and commercial data, such as enterprise search, eye search, Qixinbao and other platforms. Through such a platform, complete information of the enterprise can be obtained. At the same time, according to the core characteristics of the product, the required characteristic values are supplemented with the detailed information required by the corresponding enterprises.
在有了这些数据之后，才能根据平台上面的每一家客户自己形成成交企业的数据集，去进行对应的模型训练，才能产生最终针对企业比较有效的推荐。所以，真正最后的推荐其实是针对每一家企业，都有一个属于自己的推荐模型。
After these data are available, the data set of the transaction enterprise can be formed according to each customer on the platform, and the corresponding model training can be carried out, so as to generate more effective recommendations for the enterprise. Therefore, the real final recommendation is actually aimed at each enterprise and has its own recommendation model.
第二个场景是关于零售。在这方面，我们数据就更少了，现在每启动任何一家企业的智能货架识别，都有前期一个非常长期的数据学习语料的准备过程。
The second scene is about retail. In this respect, our data is even less. Now every time any enterprise starts intelligent shelf identification, there is a very long-term preparation process for data learning corpus.
首先，需要这家企业先提供所有的商品产品，和对应的图象数据、包装、外观所有的图像数据。同时，我们需要在对应的场景下大量的货架拍摄数据，通过数据对比去进行对应的数据学习和训练，最终不断去调整模型。并将这样一个数据识别的准确度，或置信度达到95%以上。当客户觉得准确度达到业务平时要求时，最终才能够在业务上真正使用起来。
First of all, the enterprise is required to provide all the merchandise products, and the corresponding image data, packaging, appearance and all the image data. At the same time, we need a large number of shelf shooting data in the corresponding scene, through data comparison to carry out corresponding data learning and training, and finally constantly adjust the model. And the accuracy, or confidence, of such a data identification reaches more than 95%. When customers feel that the accuracy meets the usual requirements of the business, they can finally use it in the business.
李一帆：首先因为深度学习算法对数据需求很大，而且要求是标注的数据，所以其实不仅要原数据，还要有标注，这是两个维度。
Li Yifan: First of all, because the deep learning algorithm has a great demand for data and requires labeled data, it actually requires not only the original data but also labeled data. These are two dimensions.
先说标注。数据标注通常的解决办法就是先通过标注公司的人工来形成数据，先建一个基础的模型，再去标注AI，最后就用AI来做。
Let's talk about labeling first. The common solution to data labeling is to form data by labeling the company's labor, build a basic model first, then label AI, and finally use AI to do it.
第二个维度是原始数据，原始数据最关键的是能够打通业务流程，业务接口能够有原始数据。
The second dimension is the original data. The most important thing for the original data is to be able to get through the business process and the business interface can have the original data.
第三个维度是当客户越来越多之后，如何守住客户合作过程中，或者通过其他渠道采集的数据。
The third dimension is how to keep the data collected in the process of customer cooperation or through other channels when there are more and more customers.
这都需要一个非常好的数据管理平台。在我们内部，最近开发了一款产品，它的理念是帮助企业去整理公有库以及私有库。公有库是包括电商数据、线下一些行业的数据。私有库就是通过图形和相应的业务数据的管理，把数据存储下来，有一个内部明确数据陈列的地方。
This requires a very good data management platform. Within us, we have recently developed a product whose idea is to help enterprises organize public and private libraries. Public libraries include e-commerce data and data from some offline industries. Private library is to store data through the management of graphics and corresponding business data, with a clear internal data display place.
另外一点是，如何将专业知识数字化。比如，最近正在做的建筑省图，每个省有七八本建筑规范，每一个专业又有好几本，每一本规范里有数万条，这种数据的整理还没有得到很好地解决。不过，相信在这块有很大发展空间，能够真正把大量的复杂业务规则数据化。
Another point is how to digitize professional knowledge. For example, there are seven or eight building codes in each province, several in each major and tens of thousands in each code in the construction provincial map currently being made. The collation of such data has not been well solved. However, I believe there is a lot of room for development in this area, which can really digitize a large number of complex business rules.
话题四 产品研发与落地中的瓶颈
Topic 4 Bottlenecks in Product Research and Development and Landing
李一帆：瓶颈主要体现在数据、算法、算力三个方面。
Li Yifan: The bottleneck is mainly reflected in data, algorithm and calculation power.
首先数据不够，肯定是不行的。
First of all, the data is not enough, it is definitely not possible.
其次，在算法上，我们的算法还没有到很强，所以有很多缺陷，如需要大量数据，需要算法调优，需要人工维护，算法本身也是一个数据。
Secondly, in terms of algorithm, our algorithm is not very strong, so there are many defects, such as requiring a large amount of data, algorithm optimization, manual maintenance, and the algorithm itself is also a data.
最后说到算力，尤其当我们做线下结合场景时，算力不够。
Finally, when it comes to calculation power, especially when we do offline combination scenes, calculation power is not enough.
这三个掣肘，其实可以通过产品手段去除。产品经理的理念是，通过设计功能去满足客户的需求，这里可以通过让客户妥协、让功能妥协，或者自己让一步，因此需要产品经理从三个维度留一些口子，留一些给员工去操作的开放空间。
These three constraints can actually be removed by product means. The idea of the product manager is to meet the needs of customers by designing functions. Here, customers can compromise, functions can compromise, or one step by oneself. Therefore, the product manager is required to leave some holes from three dimensions and some open space for employees to operate.
陈宏志：我们的最大痛点是人。
Chen Hongzhi: Our biggest pain point is people.
第一，因为用友做得多是 To B 业务，最大的问题是传统企业对 AI 的认识存在两个极端：要么 AI 是人工+智能，要么是 AI 无所不能，其实这两个都不是。
First, because UFIDA does a lot of To B business, the biggest problem is that there are two extremes in traditional enterprises' understanding of AI: either AI is artificial + intelligent, or AI is omnipotent, but neither is.
对于传统产业，最主要是把这些产业知识如何变成 AI 里面的数据，并且把它用上，这个其实是比较大的难题。
For traditional industries, the most important thing is how to turn these industrial knowledge into AI data and use it. This is actually a big problem.
第二，是企业中的人。尤其大的企业数字化转型要驱动一把手去推，但因为盘根错节的组织关系的存在，当他们不愿意去改变的时候，往往呈现的与实际效果就是天壤之别。
Second, it is the people in the enterprise. In particular, the digital transformation of large enterprises should drive the top leaders to push forward. However, due to the existence of intertwined organizational relationships, when they are unwilling to change, the actual effect is often quite different.
话题五 跨部门协作与跨领域协作
Topic 5 Cross-departmental Collaboration and Cross-domain Collaboration
林松：对于我们这类的创业企业来讲，最关键的是人才。一个既懂业务，又能够在数据和人工智能方面比较专业性的复合型人才，能够很好帮助企业把 AI 真正应用到业务层面。我个人看到的是，这种人才在行业里都很稀缺。
Lin Song: For start-up enterprises like ours, the most important thing is talents. A compound talent who not only understands business, but also can be more professional in data and artificial intelligence can help enterprises to really apply AI to the business level. What I personally see is that such talents are very scarce in the industry.
在跨领域协作上，我们采用的是数据团队和产品团队结合起来做事情。因为我们本身是做企业级 SaaS，会有一个 BI 团队，主要是把整个系统里面的数据进行最终的计算呈现。而后，在 BI 团队孵化 AI 小组，这让 AI 小组在 BI 里面会对整个业务产生的各种业务数据有更好的理解。这样一个BI+AI 结合下，再跟产品经理一起去打造 AI 产品。
In cross-domain collaboration, we use the combination of data team and product team to do things. Because we are doing enterprise SaaS, there will be a BI team, which will mainly make the final calculation and presentation of the data in the whole system. After that, the AI team will be hatched in the BI team, which will enable the AI team to have a better understanding of all kinds of business data generated by the whole business in the BI. With such a BI + AI combination, we will work with the product manager to create AI products.
另外，在图像领域里，我们也会和外部一些专业公司进行合作。
In addition, in the field of images, we will also cooperate with some external professional companies.
王守：其实，豆瓣也是一家创业公司。对于豆瓣来说这个问题很不典型，因为豆瓣的数据对内所有都是公开的。
Wang Shou: In fact, Douban is also a start-up company. For Douban, this problem is very atypical, because Douban's data are all open internally.
我补充一点是关于 AI 产品。感触很深的是，AI 产品经理和传统软件开发或互联网开发产品经理的不同。传统软件开发或互联网产品开发，要求的功能是0或者1，有或者没有，产品经理做的决定就是要不要这个功能。
I would like to add one point about AI products. What is deeply touched is the difference between AI product managers and traditional software development or Internet development product managers. For traditional software development or Internet product development, the required function is 0 or 1, with or without it. The decision made by the product manager is whether to want this function or not.
在 AI 产品研发中，不是0和1的标准，而是产品的准确率90%、85%、75%，是客户预期或未来用户的预期，很难判断做到80%满足，还是做到90%满足，亦或其实做75%就够了。
In AI product research and development, it is not the standard of 0 and 1, but the accuracy rate of the product is 90%, 85% and 75%, which is the expectation of customers or future users. It is difficult to judge whether 80% is satisfied, 90% is satisfied, or 75% is enough.
所以，AI 对于产品经理的要求会更高，尤其是他的判断力很重要。我们也是摸索了很长时间，才形成了这样一套机制，整个产品的决策并不仅仅由产品经理来决定。尤其涉及到跟AI相关的产品，要综合算法工程师、开发和运维工程师，以及销售，客户方代表，共同去规划一个产品应该做到什么程度。
Therefore, AI will have higher requirements for product managers, especially his judgment is very important. It took us a long time to form such a mechanism. The decision-making of the whole product is not only decided by the product manager. Especially when it comes to AI-related products, algorithm engineers, development and operation engineers, as well as sales and customer representatives, should be integrated to plan to what extent a product should achieve.
展望 AI 在 SaaS 领域的未来
Looking Forward to the Future of AI in SaaS
李一帆：希望未来3-5年，我们能够将 AI 与SaaS 融合在一起，把产品真正做起来。
Li Yifan: I hope that in the next 3-5 years, we can integrate AI and SaaS and make the products real.
林松：以后 AI 一定会成为整个商业效率提升的最佳助推器，使工作变得更加便捷轻松。
Lin Song: AI will definitely become the best booster to improve the efficiency of the whole business in the future, making the work more convenient and easy.
王守：作为一名 AI 从业人员，我的期待是在未来3-5年，能在认知智能上有本质的突破，如同这几年的 CV 一样。相信如果认知智能、认知计算能得到本质突破，一定会打开一扇新的世界大门。
Wang Shou: As an AI practitioner, my expectation is to make an essential breakthrough in cognitive intelligence in the next 3-5 years, just like CV in recent years. I believe that if cognitive intelligence and cognitive computing can get essential breakthroughs, they will definitely open a new world door.
陈运文：AI 的应用场景只有深挖下去，才能够真正落地。
Chen Yunwen: The application scenario of AI can only fall to the ground if it is dug deep.
陈宏志：希望未来能有更多懂技术、懂业务，而且善于发现场景的产品经理，一起加入到 AI 赋能数字化转型的大家庭中来。未来，大家一起坚持，把 AI 做到更好，帮助人们实现更多愿望！
Chen Hongzhi: I hope more product managers who understand technology, business and scene discovery will join the AI-enabled digital transformation family in the future. In the future, everyone will stick to it together to make AI better and help people realize more wishes!
赞助本站
Sponsor this site
