At present, various AI products emerge in an endless stream, but how to solve problems such as high application thresholds and little efficiency improvement for enterprises is still an unavoidable topic on the road to AI. The emergence of AI assistants may be able to give us some help.
Preface
Since OpenAI released ChatGPT, various AI products have emerged one after another, and they are changing our lives little by little.
- Search AI can give smarter and more helpful answers
- Assistant AI can convert video conferences and automatically generate summaries
- Design AI can generate pictures based on prompt words, and even generate good PPT with one click
- ……
Although AI has played a great role in improving efficiency in personal learning and work, there are still many problems, the biggest of which are these two:
- The threshold for application is too high: The threshold is too high for ordinary people to find the right tool according to their own problem scenario and then learn how to use it
- Enterprises have little efficiency: These tools are more often applied to individual scenarios, and do little to improve the efficiency of repetitive tasks in enterprises
Integrating AI efficiency improvement scenarios and solving repetitive tasks through agents are two hurdles that enterprises cannot bypass to improve efficiency.
DingTalk, an enterprise tool, has been working hard in the past 1 year, and AI+ has become the main theme of DingTalk in the past 1 year.
On April 18, DingTalk officially launched the AI assistant market, allowing more enterprises to embrace AI to improve efficiency.
Title: PaaS (Copolit) [Copolit)、AI助 (Agent)、AI助呬 (AI Agent Store), 它𱔼一起构构 I'm going to say
- DingTalk AI PaaS system: It provides the underlying PaaS capabilities of large model calling, proprietary model training, and enterprise application access, which can access more enterprises to allow them to develop Copolit and Agent products based on AI PaaS.
- DingTalk AI (Copolit): integrates AI efficiency improvement scenarios and provides a variety of AI functions such as meeting summary, content generation, and document assistance, making it more convenient for users to summarize and create.
- AI Agent: Solve repetitive work, can realize various workflows with one click, and can greatly improve the efficiency of people in repetitive work.
- AI Agent Store: Further reduce the cost of using AI for enterprises, so that more enterprises can use AI at a low cost to improve efficiency
DingTalk naturally has AI application scenarios, and it is not a matter of holding a hammer to find a nail by reversing the model to make AI assistant products. What are the enhancements of building an AI agent on DingTalk and building it directly on the base model?
1) Capability enhancement: The AI assistant can be deeply bound and combined with DingTalk, which means that the AI assistant is not only an independent AI product, but can seamlessly connect with the existing functions and data of DingTalk, so as to provide richer and more personalized services.
2) Traffic or rationality issues: DingTalk itself has the needs and scenarios of all walks of life, users naturally exist in scenarios, and there are needs in scenarios. This is in stark contrast to the current problems such as GPTS and large models, which often lack clear user needs, and users only look for AI when they have needs. On DingTalk, the user's needs have been clear, and the AI assistant can provide solutions more accurately.
3) ToB market characteristics: It is difficult for ToB to exist in a single phenomenon-level application, but tens of thousands of roles and industry assistants to meet specific user groups. This determines the DingTalk AI assistant market, and does not make full recommendations, but only recommends selected AI assistants, AI assistants with industry attributes, action capabilities, and professional knowledge to ensure that each user can get the service that best meets their needs.
The AI assistant market has completed the last link of DingTalk's intelligent office, but for itself, the future challenges have just begun.
The current AI assistant has the initial agent ability to handle simple RPA tasks, but it lacks the integration of external systems.
When the RPA function and SaaS access of the AI assistant are more perfect, users only need to send task instructions to the AI assistant, which can help you create meetings, initiate meeting invitations, summarize meeting minutes, and follow up To-Do matters after the meeting.
The all-powerful AI super assistant that looks back at us in the future.
1. How does DingTalk develop AI?
Any good product evolves little by little, let's see how DingTalk AI has done in the past 1 year.
中台基建 (AI PaaS)——钉钉AI (Copolit)——AI助理(Agent)—AI助理市场(Store)——埋 I'm sorry, I'm sorry Hey, he's right.
Since April 23, DingTalk has built an AI PaaS platform based on the LLM capabilities provided by Tongyi Qianwen, so that DingTalk has scalability in AI.
It can provide LLM capabilities and access more SaaS applications through the middle platform.
AI PaaS平台是AI助理的地基,AI助理的LLM能力、SaaS接口、钉钉接口都依赖于AI PaaS平台来提供。
The AI PaaS platform is the starting point of DingTalk AI and the end point of how much it can do in the future.
Based on the AI PaaS platform, DingTalk has launched the DingTalk AI (various Copolit) functions, which are based on AI to achieve summary and creation capabilities, which can allow you to browse meeting content more quickly and allow you to assist in creation through AI, but it has little effect on the efficiency of repeated workflows.
If Copolit is used to assist people to complete their work, then Agent is to complete their work on behalf of others, and there are still big differences between Agent and Copolit in terms of use scenarios, and their complexity is not the same.
The document assistant Copolit is enough to cover the demands of most users by providing meeting summaries, document assisted writing functions, and AIGC content generation.
The B-side scenario is less versatile, and it is difficult to solve everyone's problems through fixed agents.
In January this year, DingTalk released an AI assistant (Agent), which can be based on DingTalk tools, RPA, SaaS and other capabilities, allowing users to solve its efficiency problems in various scenarios by independently creating AI assistants.
Followed by the DingTalk AI assistant market in April, more creators' AI assistants have been exposed, and users can use AI assistants at a lower cost.
The AI assistant market (Store) has completed the last link of DingTalk's intelligent office, but it still has a lot of homework to do.
Add better RPA functions, access more SaaS systems, and improve the memory ability of AI assistants, so that more high-quality AI assistants can be created.
In my opinion, the development of DingTalk AI is very steady and methodical.
They didn't want to make a big move all at once, but first took the time to build an AI PaaS system, and first laid the foundation of a high-rise building, but this foundation is actually difficult for people to see its value in a short period of time.
You don't have such a large amount of business, you don't have such a big demand, why not hurry up and quickly put on the function to meet the needs of users.
I think DingTalk also has to face a lot of such questions when doing AI PaaS, just like when Alibaba Cloud first started, you have to stick to what you believe in and spend time trying to do it well.
When a large number of vertical enterprise models need to be accessed, when a large number of SaaS software is integrated, and when AI assistants can complete super complex workflows with one click, people can understand how important a good middle office system is and what is the value of DingTalk's early investment in AI PaaS.
First, lay the foundation, then integrate AI efficiency improvement scenarios, so that enterprise users can quickly use various AI capabilities on DingTalk, and finally use enterprise creators to overcome the most difficult agent scenarios.
二、钉钉AI Copolit:侧重协作
The first AI capabilities provided by DingTalk focus on summarizing and creating, and it focuses on integrating the ability to ask thousands of questions from Tongyi and listen to and understand.
In the current market, summary and creation capabilities are not scarce, and DingTalk's integration is more about allowing users to use these capabilities more conveniently in work scenarios.
In terms of summarizing, it supports the ability of meeting and document summarization, and in the creation of AIGC content generation, document assisted writing, conference poster generation and other capabilities.
Make it more convenient for users to summarize and create, this is the problem that DingTalk AI Copolit solves.
For the solution of repetitive scene problems, it is left to the AI assistant agent.
DingTalk AI Copolit has done a good job in integrating AI efficiency improvement scenarios, but there are some scenarios that it is not well embedded in, such as the dialogue and search of the basic chatbot, which are two relatively high-frequency scenarios.
From the analysis of DingTalk's product logic, on the one hand, it does not have a very good product in the search scene, and on the other hand, it hopes that the AI assistant will undertake these functions.
However, I think that in the early stage of integrating AI efficiency improvement scenarios, it may be a better choice to use Tongyi Qianwen as an assistant to make up for unexpected scenarios when designing products, so as to give users more choice, and later iterate products according to user behavior.
3. AI Agent: Efficiency improvement
The AI assistant agent can realize various workflows with one click, which can greatly improve the efficiency of people in repetitive work.
Is this an over-the-top compliment to the Agent?
Let's take a look at how the agent works:
Lilian, OpenAI AI 安全团队 leader
Agent = 大模型（LLM）+ 工具使用 + 规划 + 行动 + 记忆
Planning through large models, having the ability to recall tools to take action, so that the AI can complete the job.
The DingTalk AI assistant allows users to send tasks to the AI assistant through dialogue, and it completes the task by analyzing the task and calling the DingTalk capability.
For example, if the user asks it to create a new meeting, the user only needs to tell it the meeting information, and it will call the meeting tool to generate a meeting information by itself after analysis, and wait for the user to confirm again, and the meeting will be added to the calendar.
Users can invoke various functions of DingTalk through the AI assistant to complete tasks in the form of dialogues.
Users want to get started directly and can choose from the AI assistant market, and the first batch of nearly 200 AI assistants on the DingTalk AI assistant market are available for users to choose from.
Tongyi Farui can provide users with professional legal knowledge and let users have a paralegal.
You can also use the Suno Lyrics Generator to generate a song that the user wants.
If you want to use an AI assistant to solve more complex scenarios, you can create an AI assistant yourself.
AI assistants can be created in two ways: application capabilities and workflows.
At present, the RPA function in DingTalk has been supported on the mobile terminal, and some cumbersome steps can be collected first, and then the AI assistant can be realized through the RPA function.
Workflow scenarios allow users to set up business execution actions based on their own workflows.
This is a workflow that can go to 1688 to search for products, after configuration, users only need to give it product information, it can go to 1688 to search for products, and organize and send them to users in the form of interactive cards.
Knowing this, we will find that DingTalk AI assistant currently implements agent in some basic scenarios, and it actually has little effect in many complex scenarios.
So how can it become a super AI assistant to help enterprises improve efficiency?
Start with more tools and memory at your disposal, and then plan on your own.
After reading the Agent, we will find that the real place that determines its future is precisely the AI PaaS platform, the tools it needs need to be accessed by the AI PaaS platform, the memory capabilities required need to be supported by the AI PaaS platform, and the self-planning capabilities required need to be trained on the AI PaaS platform.
It is impossible to do function superposition purely on the AI assistant, design a good workflow plan, and then through the capabilities of AI PaaS, you can create more valuable AI assistants.
Although the current AI assistant can only improve the efficiency of some simple repetitive scenes, this is a good start, but it still needs to be optimized.
Fourth, what exactly does AI bring?
Taking the meeting scenario as an example, let's see how AI can improve the efficiency of meetings and what are its shortcomings.
A meeting usually consists of three parts: pre-meeting invitation, meeting minutes, and post-meeting To-Do follow-up, so what kind of help can AI give us?
Create meetings with one click, record meeting content in real time, organize meeting minutes into documents, and create To-Do items.
1. How can AI improve the efficiency of meetings?
Start with a pre-meeting invitation, what should I do if I don't want to go to the calendar to add a meeting and want to be lazier?
Directly type and tell the AI assistant to create it for you, you need to send it the meeting time, name, participants, and meeting room, and it can create it for you.
You can use the "Flash" function in the meeting, which can record the meeting content in real time and generate intelligent minutes after the meeting, which is convenient for participants to review later.
After the meeting, we can send the content of our meeting summary to the AI assistant, which will automatically organize and generate a document.
When the meeting is over, we can assign tasks through the AI assistant, just need to tell it the name, the executor, and the deadline:
Once the AI assistant is created, the task will appear in our to-do list:
AI assistants help a lot in meeting scenarios, but people still have to do a lot of steps, so what can AI assistants do in the future to be more efficient?
2. How can you get stronger next?
Let's go back to this Agent diagram, in order to implement the Agent, you need a large model (LLM) + tool use + planning + action + memory.
Let's take a look at how to make AI assistants stronger from two scenarios: meeting and recruitment collaboration.
But what capabilities does the user want the AI assistant to have in a meeting?
- Pre-meeting invitation: Able to invite according to the meeting requirements put forward by the user, and at the same time feedback to the user according to the invitation situation.
- Meeting minutes: Invoke Copolit's meeting minutes function to generate meeting minutes after the meeting.
- Post-meeting To-Do follow-up: Assist users in generating To-Dos based on user meetings and follow up on actual situations.
This is a capability that users want AI to be able to have in the actual meeting scenario.
But in fact, the current AI assistant has no way to achieve these functions, mainly in memory ability and tool use.
Planning is not important at the moment, because people will spend a little time configuring the business flow, and it is even more important to let the AI assistant have the memory and the ability to use the tools first.
So if you only do one, which one should you do first?
In terms of using the tool first, it is also a problem solving idea to write the time into the workflow without considering the memory ability of the agent.
The number and complexity of the tools supported by workflows can determine the success or failure of an AI assistant.
DingTalk AI assistant can be divided into three steps to gradually enrich its tool capabilities:
- Step 1: Closed loop inside DingTalk to support various DingTalk capabilities to be called by workflows.
- Step 2: Use RPA + AI to make the workflow support jobs in multiple external scenarios.
- Step 3: Access to the SaaS system to integrate and open up the enterprise business, so that the workflow can cover the whole business process.
After these three steps are completed, an AI assistant can be built to create a travel plan for the user according to the budget, upload OA approval, automatically purchase tickets, and calculate expenses to the financial system.
Finally, if the AI assistant has a strong planning ability, the user does not have to worry about complex workflows, so why not do the planning ability first?
Because there is not enough sample data to train planning ability.
This is actually the ability to train a large vertical model, which can learn based on a large number of cases, so as to have the ability of strong planning.
5. Summary
After experiencing the DingTalk AI assistant, I think that many tedious and repetitive processes in the future work may be thrown to the AI assistant to execute, leaving the mechanized things to the robot to complete, and what more creative and valuable things can people do.
From AI PaaS to Copolit to AI Assistant to AI Assistant Store, DingTalk is very stable step by step, but we can also detect many problems:
1. AI PaaS doesn't support as much as native products
The ability of the conference comes from the general sense of listening, but many of the capabilities that the general sense does not have (recording pages for content recording, secondary processing of notes), and relying on the capabilities delivered by AI PaaS will be limited in the application layer.
Is it necessary to do secondary development to solve it, how to deal with this meaningless loss, and when more SaaS APIs are accessed, how to ensure the original application experience? This is a huge challenge faced by DingTalk in the middle platform.
2. There is a lot of duplication in the functions of intelligent document authoring and AI assistant
Both of these are AI assistants, but they are actually two different product capabilities, and it's easy for users to confuse what the difference is.
Is it a question of product design, whether to encapsulate it into an AI assistant, or to separate the agent from Copolit?
3. AI assistants need more capabilities and more complex process design
At present, the AI assistant can support less than 20 kinds of execution actions, so in the face of the different demands of a large number of enterprises, how to reasonably increase the number of capability items, so that it will not be too much and not too little?
The process design of the AI assistant is too simple, and there is only one branch at present, which is basically difficult to meet for complex operation scenarios, and needs to be able to meet more complex workflow design capabilities, and at the same time allow people to get started quickly.
The problem needs to be solved a little bit, the product needs a little bit of polishing, DingTalk AI assistant is a great product, let's get started with it first!
This article was posted by @Super Huang on Everyone is a product manager and is prohibited from reprinting without permission.
Image from Unsplash, based on the CC0 license
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.