How can software companies prepare for the second half of AI?

Mr. Yao Shunyu's article "The Second Half" elaborates on the strategic shift in the "second half" of artificial intelligence (AI) development. Yao Shunyu joined OpenAI in August 2024 as a research scientist. He graduated from Tsinghua University's "Yao Class" with a bachelor's degree and received a doctorate from Princeton University. The core points of the article are as follows:

From "training models" to "defining tasks": In the first half of AI development, research focused on developing new training methods and models, such as Transformer, deep reinforcement learning, and large-scale pre-training models, which promoted breakthroughs in AI in many fields. However, as the technology matures, Yao Shunyu pointed out that the second half of AI will shift its focus from "solving problems" to "defining problems." This means that researchers need to think like product managers, clarifying what problems AI should solve for whom, and how to measure "whether the problem is solved well."

Generalization ability of reinforcement learning: Yao Shunyu emphasized that reinforcement learning (RL) finally has the ability to generalize. In the past, the application of RL was often limited to specific environments and difficult to promote. Today, by combining language and reasoning, RL can solve a wider range of tasks, such as software engineering, creative writing, mathematical problems, mouse and keyboard operations, and long-form question and answer. This progress marks a new stage of development for AI.

Importance of evaluation mechanism: In the second half of AI, the evaluation mechanism becomes more important than training. Yao Shunyu pointed out that traditional evaluation methods often ignore the importance of environment and prior knowledge, resulting in algorithms performing well in specific environments but not performing well in actual applications. Therefore, redesigning the evaluation mechanism to ensure that it can truly reflect the performance of the model in real tasks is the key to the future development of AI.

Introduction of product thinking: Yao Shunyu proposed that AI researchers need to introduce product thinking and focus on the effects of AI systems in practical applications and user needs. This means that researchers should not only pay attention to the performance of the model, but also consider how to transform technology into useful products to meet the dynamic changes of the market.

You can read the original article through the following link: ? The Second Half – Shunyu Yao (https://ysymyth.github.io/The-Second-Half/)

Inspired by this article, I wrote this article with AI to answer the following questions: What is the core of the first half of AI in the era of big models? Why do we say that we have entered the second half of AI? What is the core of the second half of AI? How can software companies prepare for the second half of AI? Thanks to Mr. Yao Shunyu.

What is the core of AI in the first half of the era of big models ?

Over the past decade, the " first half " of artificial intelligence ( AI ) has been driven mainly by breakthroughs in large models and new algorithms . Various milestones have emerged one after another: Deep Blue defeated the chess champion, AlphaGo defeated the Go champion, GPT-4 surpassed most humans in legal and mathematical exams , etc.Supporting these brilliant achievements is a series of core technological advances : from classic search algorithms, deep reinforcement learning, to large-scale pre-training and reasoning technology, each breakthrough has brought a leap forward in AI capabilities ..

In the first half, the Transformer architecture and GPT series models were landmark milestones in the evolution of AI . The Transformer model proposed in 2017 introduced the " self-attention " mechanism, which greatly improved the model's ability to conduct parallel training and understand long sequence dependencies, laying the foundation for subsequent pre-trained large models.Subsequently, Transformer -based pre-training models flourished: BERT achieved a breakthrough in natural language understanding tasks, and GPT-3 demonstrated amazing Few-Shot learning capabilities with 175 billion parameters in 2020 , allowing the model to solve new tasks with zero or few samples . The improvement of the general capabilities of these foundation models in fields such as natural language and computer vision has driven AI performance to rise steadily on multiple benchmark tasks..

As you can see, the core competition in the first half is " whether you can train a better model " and " whether you can get a higher score on the benchmark test " . The most influential papers recognized by the academic community often propose new models or training methods, such as AlexNet (deep convolutional network), Transformer , GPT-3 , etc.These works have proven their value through their remarkable achievements on public datasets. However, from another perspective, this also reflects the limitations of the evaluation criteria in the first half : everyone is more concerned about the improvement of the model itself, and real tasks and evaluation benchmarks are often just tools to test the model.After all, when AI capabilities are still at a relatively low stage, improving the intelligence of the model can often lead to improvements in various tasks, so the focus is on how to improve the performance of the model itself.

In general, the core characteristics of the first half of AI are: algorithm-driven, model-oriented . Through continuous innovation of network structure and training paradigm (such as deep learning, Transformer , large-scale pre-training, etc.), the intelligence level of AI systems has been improved by orders of magnitude. At this stage, AI is more like a powerful tool developed to compete in various artificially designed tests.

Why we have entered the second half of AI

Today, the industry generally believes that we are standing at the starting point of the " second half of AI " . Behind this judgment is a dual transformation of technical paradigm and practice orientation . OpenAI researcher Yao Shunyu pointed out in his article " The Second Half ": "AI has entered the halftime break " , and the second half that follows will be fundamentally different from the first half.

First, from the perspective of technological milestones, we have found " an effective AI recipe " that has enabled AI to achieve breakthroughs in a wide range of tasks . In the past, reinforcement learning ( RL ) was often seen as the ultimate path to general intelligence, but its generalization ability has always been limited. However, recently, by combining large models with RL (such as the use of human feedback reinforcement learning RLHF in dialogue models ), the same paradigm has begun to be able to solve a variety of complex tasks . For example, a fine-tuned GPT-4 model can actually complete many tasks that were originally unrelated to each other, such as programming, creative writing, solving math problems, cross-software operations, long-form question and answering - this scenario was almost unimaginable a year or two ago, but it has recently become a reality. This shows that a paradigm shift has occurred in the field of AI : large model pre-training provides powerful prior knowledge, supplemented by a small amount of task-specific training or feedback, AI can perform well in multiple fields. This marks a qualitative change in the game centered on " continuously improving model algorithms " in the first half , laying the foundation for the second half..

Secondly, the " chart-topping " competition model of the first half is changing. Yao Shunyu pointed out that the " universal formula " of large models has been standardized and industrialized , and continuous expansion can make great progress on most benchmarks.. This leads to two results: on the one hand, researchers' hard work in designing a new model for a specific task may only bring a 5% improvement, while the next generation of general large models (such as the models called "O series " within OpenAI ) may improve by 30% without targeted optimization . On the other hand, even if we build a new and more difficult evaluation benchmark, large models can often " pass " quickly . In other words, the previous cycle of " improving algorithms- > improving scores- > publishing papers " is losing its effectiveness, and the practice of " rolling model parameters and rolling benchmarks" has a smaller marginal contribution to promoting actual results.

More importantly, the economic value of AI has not kept pace with the technological achievements . Although AI has defeated humans in Go, exams, and Olympic competitions, there has not been a huge change in macroeconomics and productivity commensurate with these breakthroughs.Yao Shunyu called it the " utility problem " of AI : AI performs well in labs and competitions, but its productivity improvement in the real world is not obvious. This reminds us that the past standards for evaluating AI success ( such as competition scores and benchmark accuracy) may be out of touch with actual business value, and the AI ​​research paradigm needs to shift from pursuing indicators to pursuing real utility.Because of this, more and more people in the industry realize that we are entering a new stage. The focus of the second half is no longer to simply improve the performance of the model in a closed environment, but to truly integrate AI into real-life scenarios and create tangible value .

In summary, there are three main reasons why we are entering the second half of AI : the general large model paradigm has matured , making model improvement easy and universal; the traditional competitive innovation dividend has decreased , and new ideas are needed to break the bottleneck; and the gap in real value has forced AI to move from the laboratory to productization. As Yao Shunyu said, the key to the second half is that we must change our mindset and skill set , switching from a researcher's mindset to a mindset closer to a product manager.. Only in this way can we win in the new competition.

What is the core of the second half of AI ?

The core concept of the second half of AI can be summarized in one sentence: " Defining the task is more important than solving the task . " At this stage, AI is no longer seen as a tool for showmanship, but as a product . We need to use the mindset of building products to promote the development of AI . This means the following key points:

1. Clearly define truly valuable problems. In the first half, we often say " given a task, find a way to make the model do it well " ; and in the second half, we need to think the other way around - " What is the most worthy task for AI to solve at present? "  As Yao Shunyu pointed out, the second half will shift from " solving problems " to " defining problems " . Because as the capabilities of large models become stronger, whether or not many problems can be solved is no longer a problem. The key is what problems are meaningful to solve. Defining tasks includes clarifying the real pain points of users, business demands, and whether AI is suitable for solving the problem. In the context of enterprise products, this often means that product managers need to go deep into business scenarios, find problems that can bring significant value after being solved , and clearly describe them as task forms that AI models can accept. For example, instead of letting the team spend a lot of effort to improve the chatbot's chat accuracy by 1% , it is better to define a more business-oriented goal, such as " let the customer service robot effectively identify user emotions and provide comfort measures in the product return and exchange scenario " . Such task definitions are directly oriented to real needs and have more direct value. 

2. Reshape the evaluation mechanism. Evaluation is more important than training. In the second half, it is not enough to have a powerful model. We must establish a scientific and effective evaluation system to measure whether AI has truly achieved its expected goals. . Traditionally, evaluation is often limited to offline metrics (accuracy, F1 score, etc.) or performance on a closed test set. But as Yao Shunyu points out, many default evaluation assumptions do not conform to reality:For example, in real interactions, users and AI have repeated conversations, while benchmark evaluations usually assume that the model outputs the answer in one go. Another example is that real tasks are continuous and have memory accumulation, while the evaluation set treats each sample independently.Therefore, in the second half, we need to boldly question and redesign the evaluation plan .Possible approaches include: introducing human participation in evaluation (for example, letting real users interact with AI to judge the effect);), test environments that simulate real business processes, design long-term continuous tasks to test AI 's continuous learning and adaptability, etc. The goal of innovation in the evaluation mechanism is to ensure that the performance indicators of the model are related to real business utility , so that AI is no longer trained for " getting high scores on exams " but optimized for " excellent performance in actual applications ." As Yao Shunyu said: " We must fundamentally examine the existing evaluation methods and create new evaluations to force real innovations beyond the formula . " In this process, evaluation is even more important than training - only with good evaluation criteria can we know whether the trained model is truly useful.

3. Build AI with product thinking : AI is a product, not just a tool. The second half requires us to view the AI ​​system as a continuously iterating product , rather than just a one-time algorithm tool. This means paying attention to user experience, maintainability, reliability, and value delivery . As a product, AI needs to consider how to integrate into the user's workflow, how to present results, how to obtain feedback and continuously improve. This is no different from traditional software products. Yao Shunyu emphasized that to succeed in the second half, a mentality and skills closer to that of a product manager are needed. . For example, when designing a new AI -driven SaaS feature, the team should not only care about the accuracy of the model prediction, but also how the user will understand and use it after seeing the prediction - this involves product-level considerations such as explainability (why does the user get this result?), interaction design (how to prompt if the model is uncertain?), and failure protection (is there a backup plan when an error occurs?). Because of this, some people have proposed "AI is a product manager " , which means that the AI ​​project team needs someone to balance technology and experience specifically, so that AI can truly land and become a product that users are willing to use. Only when AI is delivered to end users in the form of products and continues to create value can we say that AI has gone from the laboratory to the real world. The winner of the second half lies in: whoever can turn AI technology into a product that users pay for the fastest will be able to lead in the new competition. As Yao Shunyu said, players in the second half will create billions or even hundreds of billions of dollars in value by building useful intelligent products - AI is no longer a tool for showing off, but a product force that changes the industry landscape.

In summary, the core of the second half of AI is not to show off how powerful the model is, but to choose the right track and run with value . Defining the right problems, using appropriate indicators to evaluate progress, and using product thinking to polish the experience, these three together ensure that AI truly solves practical problems and wins the favor of users. The second half of the competition is the ability to transform AI into products and productivity , which will profoundly determine the future success or failure of software companies and industries.

How software companies can prepare for the second half of AI

Facing the second half of the AI ​​wave, SaaS and enterprise software companies need to make all-round preparations from organization to strategy to seize opportunities and meet challenges. The following aspects are crucial:

Organizational structure: Building a cross-functional AI product team

In the second half, software companies first need to adjust their organizational structure and form cross-functional AI product teams . Traditional software development teams may isolate AI development in R&D departments or laboratories, but the second half requires deep integration of AI and products, so the team needs to bring together multiple roles to collaborate, including product managers ( PMs ), AI engineers / data scientists, back-end developers, data engineers, and evaluation and operation and maintenance personnel. An ideal AI product team should have both product roles that understand business and user needs , and technical roles that are proficient in algorithm models , while adding evaluation and feedback functions to ensure that the model effect meets expectations.

Specifically, the following key roles can be included:

• Product Manager ( PM ): Responsible for defining the vision and goals of AI products, deeply understanding user needs and business pain points, and translating these into specifications for AI functions. PMs need to balance technical feasibility and user value in AI projects to ensure that the team does the " right " thing.

  
  

• Data Scientist / Algorithm Engineer: Responsible for model selection, training, and optimization. They build or call large models or algorithms that are suitable for the problem, and continuously debug to improve performance. This role focuses on the intelligence level of AI , and in the second half, they also work closely with PMs to determine task definitions and indicators.

  
  

• Backend / Application Development Engineer: Responsible for integrating AI capabilities into the company's product applications, realizing model reasoning services and docking with existing systems. This includes building APIs , microservices, and necessary engineering optimization to ensure that AI functions can run stably and reliably in the production environment.

  
  

• Data Engineer: Responsible for preparing and managing data pipelines to provide high-quality training data and continuous input data for AI . In the second half, data quality and freshness directly affect the effectiveness of AI . Data engineers must ensure that the data infrastructure supports continuous iteration of the model.

  
  

• Evaluation and quality control personnel ( AI QA/ Evaluation Engineer): This is an emerging and important role in the second half of the game. They are responsible for designing and executing model evaluation plans and monitoring the quality of model output. They are equivalent to testing and supervising AI products. They use pre-defined evaluation indicators to continuously test the performance of the model in real scenarios and find problems in a timely manner. For example, Morgan Stanley was able to successfully deploy the GPT-4 assistant. Behind it, a strict evaluation framework was established to test the model performance of each application scenario to ensure that the output quality meets the high standards expected by financial advisors.This framework verifies each AI function point before it goes online and introduces expert feedback to continuously improve the model ..

  
  

• Domain experts or business analysts: In some vertical software companies, it is best to also include industry experts, such as doctor consultants in medical AI products, supply chain experts in ERP system AI projects, etc. They help the team correctly define problems and review AI decisions , avoiding technical personnel working in isolation.

  
  

Through cross-functional teamwork, enterprises can open up the entire process of AI development : from requirements definition, data acquisition, model development to product implementation and continuous monitoring. This organizational approach has proven to be the key to the success of AI projects. Team members are jointly responsible for the final results of AI products, which can avoid the disconnection phenomenon in traditional organizations where " R&D produces models but the products are not well used . " Software companies should create such a collaborative culture so that AI research and development is no longer an island in the laboratory, but is integrated into the main line of product development.

Product strategy: Focusing on problem definition and evaluation indicators, embedding AI into business scenarios

With the right team, you also need the right product strategy to guide the direction of the AI ​​project. Software companies in the second half of the game should adopt a " problem definition-driven " product strategy , that is, first identify the core problem to be solved for users, then decide which AI technology to use to solve it, and design evaluation indicators and product solutions around the problem.

1. Clarify business problems and application scenarios. The first step of product strategy is to define the problem . Companies should sort out the most valuable and painful aspects of their own products or customer business processes, and evaluate whether AI has opportunities for improvement. The key is to select high-impact, feasible problems as entry points. For example, an ERP software company found that customers spent a lot of manpower on inventory management and often made inaccurate forecasts, so " inventory demand forecasting " is a high-value problem that can be solved with AI . At this stage, the team should avoid using AI for the sake of using AI , but focus on use cases where AI can bring qualitative changes. As mentioned earlier, the second half emphasizes that " defining tasks is more important than solving tasks " - software companies need to ask like product managers, " What AI functions do users really need ? " instead of having a model first and then looking for applications. 

2. Convert it into an AI task and design evaluation indicators. Once the business problem to be solved is determined, it needs to be converted into a task form that can be handled by the AI ​​model, and clear evaluation indicators should be formulated for it . For example, the above-mentioned inventory forecasting problem can be converted into a time series forecasting task of " predicting the demand for each SKU in the next four weeks based on historical sales and related data . " Evaluation indicators can be defined as business KPIs such as the error range of the forecast and improvement in inventory turnover . When formulating indicators, special attention should be paid to business relevance in the second half : indicators should be able to directly reflect the improvements that AI brings to the business, rather than just technical scores. For example, for customer service robots, instead of using only the model accuracy, it is better to add indicators such as " user one-time conversation resolution rate " and " user satisfaction score " to evaluate the actual service effect of AI . Yao Shunyu pointed out that we must have the courage to question existing evaluation methods and create new evaluation dimensions. Therefore, software companies should invest time in product strategy to design an indicator system that is close to real utility . For example, for generative AI writing functions, they can design quality scores for human review; for predictive AI functions, they can introduce actual business revenue indicators. This indicator system is both a guide for the development stage and a basis for monitoring after going online.

3. Embed AI functions into product workflows and provide closed-loop feedback. The third step of the strategy is to consider how to present and deliver AI functions in products . AI can only play a role if it is integrated into the user's actual workflow, so software companies need to plan product interactions to make AI outputs visible and available to users. For example, add an AI recommendation module to SaaS applications, or add AI automatic processing options to existing functions . When designing these functions, follow the principles of product usability: provide appropriate user controls (such as allowing users to adjust AI recommendations), clearly display AI results (such as labeling the reasons for recommendations or confidence), and design feedback mechanisms . The feedback mechanism is very important, which enables AI products to form a closed loop of continuous improvement : users can evaluate the quality of AI outputs and correct errors, and the system collects these feedbacks for the next round of model optimization or rule adjustment. AI products in the second half should be like Internet products, constantly iterating based on user behavior and feedback. For example, after Microsoft embedded the Copilot smart assistant in Office office software , it collected a large amount of corporate user usage to adjust the presentation of model recommendations and security settings. This dynamic optimization is the embodiment of product thinking: focusing on user experience and continuous improvement, rather than deploying it once and for all. 

4. Introduce pilot and small-step strategy. In terms of product strategy, software companies can adopt MVP (minimum viable product) Thinking, gradually verify the value of AI functions. In the second half of AI , rapid trial and error is particularly important. You can first pilot AI functions among a small range of users, verify the effect by collecting indicators and feedback, and then gradually expand. For example, a SaaS manufacturer plans to launch an AI- driven data analysis assistant. It can first invite some core customers to conduct internal testing, and evaluate whether the assistant really helps users through indicators (such as % reduction in analysis report completion time) and subjective feedback. If the result is not ideal, adjust the problem definition or model solution in time. Small-step iteration can reduce risks and allow the team to continuously calibrate the product direction in a real environment. This is different from the traditional AI research that emphasizes achieving the best results on an open data set at one time, and is closer to the Lean Startup methodology: start with the end in mind, start from actual value , and quickly verify improvements.

In general, software companies in the second half of the game should let product strategy drive AI development , rather than separate technology R&D from product needs. By focusing on business problems, formulating reasonable evaluation criteria, carefully designing user experience, and rapidly iterating, AI functions can be truly embedded in the product value chain and create sustainable benefits for customers.

Technology stack upgrade: building sustainable RAG , evaluation and interpretability modules

To support the productization of AI in the second half, software companies also need to upgrade their own technology stack . This includes adopting new architectures to improve model effectiveness, establishing a comprehensive evaluation pipeline to ensure quality, and increasing explainability to improve trust. The following three aspects are the focus:

1. Introduce the Retrieval-Augmented Generation ( RAG ) framework to combine big models with enterprise knowledge. Big models are powerful in general fields, but in enterprise business scenarios, they often need to access proprietary knowledge and real-time data. The RAG ( Retrieval-Augmented Generation ) framework is a technical architecture designed for this purpose. It adds a retrieval step before the big model generates answers , extracting relevant information from the knowledge base or database to " feed " the model. Through RAG , the big model no longer has to fight alone, but can use authoritative data within the enterprise to improve the accuracy and timeliness of answers .. For SaaS and ERP vendors, this means that the business data, user manuals, and historical records of their own systems can be connected to the big model to make AI functions more useful to customers. For example, an AI assistant built into a CRM software can first retrieve the customer database and past communication records, and then generate sales follow-up suggestions from the big model - this suggestion will be closer to the actual situation than a pure language model written out of thin air. Another benefit of RAG is that it reduces the risk of false " hallucinations " produced by big models because the model refers to authoritative sources of knowledge . AWS defines RAG as: generating more reliable and well-founded responses by allowing LLM to reference external knowledge basesTherefore, software companies should invest in building enterprise knowledge bases and vector databases , and develop retrieval components (such as keyword search or embedding similarity retrieval) to connect with large models. This RAG framework needs to operate sustainably: regularly update knowledge base content, monitor retrieval results, maintain index performance, etc., so that the AI ​​module always has the latest and most accurate business information and continuously outputs high-quality results.

2. Establish an automated evaluation and monitoring framework ( Eval Ops ). As mentioned above, evaluation is crucial in the second half. Therefore, the technology stack should include a complete model evaluation and monitoring pipeline . Before the model goes online, the team should implement automated testing : prepare a series of test data and scenarios for each AI function use case to evaluate whether the model output meets the expected indicators. This is similar to the unit test + integration test of traditional software , except that the object is the AI ​​model. For example, for a customer service robot, there can be a set of test sets of typical user inquiries and expected answers to calculate the accuracy, politeness, etc. of the model. After the model is deployed, the evaluation should be continued - that is, online monitoring . The model interaction log can be recorded in real time, user feedback annotations can be collected, and these data can be sent back to the evaluation pipeline to calculate the drift of performance indicators. Once the model effect is found to be significantly reduced or deviations occur (for example, the error rate of answers increases, and user negative feedback increases), timely warnings can be issued to trigger model retraining or hot repair. Morgan Stanley 's case shows that a sound evaluation framework is the cornerstone of the successful implementation of AI . They developed a detailed evaluation before deploying the GPT-4 assistant, comparing the differences between the model summary and the expert summary, and calibrated the model by having financial experts score it. After going online, we continue to adjust prompts and outputs through expert feedback.Software companies should learn from this practice and make Eval Ops ( Evaluation Operations ) an important part of MLOps , equipped with special tools and dashboards to have a clear understanding of AI performance . Only when the evaluation mechanism is in place can we ensure that the quality of AI products is controllable and can serve users stably and reliably.

3. Enhance the explainability and transparency of AI results. In enterprise applications, trust is more important than showing off. When using AI , many industry users are concerned about why the model makes such judgments and what the basis is. Software companies in the second half need to provide technical support. This means building an explainability module to make the AI ​​decision-making process and basis as transparent as possible to users and operation and maintenance personnel. The specific implementation method varies from application to application: for predictive models, key feature contributions and decision rule visualization can be provided; for generative AI , the referenced knowledge source can be displayed next to the answer, and the correspondence between the model-generated content and the retrieved data can be highlighted, etc. For example, when an AI assistant of a legal SaaS gives legal advice, it may also list referenced regulations or past cases to win the trust of lawyer users. For another example, when answering complex questions, the conversational robot first presents a list of facts retrieved by the tool, and then answers based on these facts, so that users understand that the answer is not out of thin air. This approach not only improves the explainability of the results , but also helps users to verify whether the results are correct or not. 

Data and research show that explainability is becoming a key requirement for AI implementation: In a 2024 McKinsey survey, 40 % of business executives believe that lack of explainability is one of the main risks of adopting AI .Yet less than one in five companies are actively addressing this issue.Therefore, software companies have the opportunity to establish competitive advantages through leading XAI (Explainable AI ) capabilities when upgrading their technology stacks. This can include: using algorithms such as LIME and SHAP to explain model decisions and create a user-friendly explanation interface; providing an "Ask why" function at the product level to allow users to ask the basis for AI decisions; and establishing an AI output log audit mechanism to facilitate internal tracking of model behavior. These measures can significantly increase user trust in AI functions.When users and decision makers trust AI more, they will use AI more widely , thus truly releasing the value of AI to the business. In short, the software technology stack in the second half is no longer just about pursuing higher accuracy or speed of models, but about filling the three shortcomings of data, evaluation, and explanation , so that AI applications can become a trusted partner of enterprises rather than a black box risk.

Business model: Positioning as a service or tool, and formulating a pricing strategy for AI value-added

In addition to preparations at the technical and product levels, software companies also need to re-examine their business models in the second half of AI , especially how to commercialize AI capabilities and how to price them. The introduction of large models may affect the revenue structure and value positioning of existing SaaS . Companies should plan service models and charging strategies in advance to fully capture the value brought by AI .

Servicing  vs.  tooling: Companies should decide whether to offer AI capabilities as a service or simply as an additional tool in their products . “ Servicing ” refers to packaging AI capabilities into a continuously provided service that may be independent of the original software product. For example, an ERP company can provide its AI prediction module to customers as a cloud service interface and charge based on the number of times it is used or the effect, which is closer to the “AI as a service ( AIaaS ) ” model. Servicing usually means that the manufacturer will continue to participate in the operation, update and support of the AI ​​function, which is more like providing a solution service . On the other hand, “ tooling ” is to sell AI as part of the product function. Users get a software license or subscription with AI functions, and the manufacturer does not charge or operate the AI ​​part additionally . For example, if AI functions are directly embedded in the standard SaaS package, users still pay the software subscription fee. Both models have their own advantages and disadvantages: servicing can create new revenue streams for AI and obtain value based on usage, but the disadvantage is that it increases operational complexity; tooling is simple and direct, which helps to quickly promote AI functions, but it may not be possible to measure the value brought by AI alone . 

Many companies are trying to balance these two models. For example, Salesforce has integrated Einstein AI capabilities into its CRM . In the early days, it was provided as part of the high-priced enterprise version, which was equivalent to a tool-based gift. Later, it also launched additional packages for Einstein functions, moving towards the direction of charging for services. Software companies need to choose a path based on their own situation: if AI is a core selling point and the operation and maintenance costs are high, they tend to be service-oriented to ensure that the investment has a return; if AI is just an auxiliary to enhance the competitiveness of existing products , they may choose to integrate it into the product as a tool, increasing the quantity without increasing the price to improve user stickiness. No matter which one you choose, the value proposition of the AI ​​function should be made clear when communicating it to the outside world - whether it is as an independent service to help customers solve specific problems, or as a product upgrade to bring a better experience.

Pricing strategy for AI value-added modules: Pricing is a key detail of the business model. In the second half of AI , the industry is still exploring reasonable AI pricing methods, but some models have emerged for reference: 

• Premium Add -on : This is a more common practice at present, that is, charging an additional fee on top of the original subscription to enable AI features. For example, Notion charges a surcharge of $10 per month for the AI ​​writing assistance feature of its note-taking software , and the AI ​​feature of Miro whiteboard is priced at $20 per user per month . This strategy clearly positions AI as a high-end value-added feature , suitable for users who are willing to pay for more powerful features. According to market observations in 2024, the AI ​​surcharges of many large SaaS companies are roughly around 50% of their basic subscription price , and some even exceed 100% . For example, the enterprise version of Microsoft 365 office suite is about $45 per seat per month , while its Copilot AI assistant is priced at $30 , which is about 67% of the base price ; for example, Google 's enterprise-oriented Workspace basic seat price is $18 , and Duet AI assistant charges $20 , which is even slightly higher than the base price. This shows that these vendors believe that the value brought by AI is comparable to the core product itself. Of course, some vendors choose to bundle AI for free to seize the market. For example, Zapier and Zendesk directly include some AI functions in their products without charging extra..

  
  

• Usage-based : AI  reasoning usually generates considerable computing costs, so charging by usage is one of the strategies that fits the cost structure. In this model, customers pay based on indicators such as the number of API calls , the amount of data processed, or the number of words in the generated content. ChatGPT Enterprise Edition of Open AI and some AI platforms follow this path: basic subscription + excess is charged by traffic. The advantage of usage-based is refined pricing , customers pay for what they use, and the manufacturer can also cover the cost. But the disadvantage is that customer budget forecasts become complicated, which may cause concerns about uncertain costs. Many SaaS vendors will set up tiered packages or quotas to simplify billing when implementing usage-based charging . For example, the AI ​​analysis function of a BI software includes 100 free queries per month, and charges are charged per thousand after exceeding it. This allows customers to have a concept of cost while being flexible in use.

  
  

• Success- based : This is a more radical and innovative model, where fees are charged only when AI produces the expected results. For example, Intercom 's Fin intelligent customer service charges by the " number of conversations successfully answered " - the company only pays if the AI ​​robot solves the customer's problem, otherwise no fees are charged. Similarly, some anti-fraud services charge a commission for each successful fraud prevention, as if " sharing profits " with customers . The benefit of Success-based is that it greatly reduces the psychological burden of customers trying AI , because they don't have to pay if there is no effect. However, the risk is higher for manufacturers, and how to define " success " requires standards that both parties agree on. In addition, under this model, customers often want to understand the basis of AI 's judgment in order to trust the billing results, and therefore have stricter requirements for AI performance and transparency. Overall, this model is suitable for scenarios where results are easy to measure (such as transaction success rate, sales lead conversion, etc.), but it is difficult to apply to complex applications. .

  
  

• Value -based pricing : From a strategic perspective, software companies can set prices based on the value AI creates for customers , rather than the cost or number of features. For example, if an AI feature can help customers save an average of $ 100,000 in costs, the manufacturer may price it at $ 20,000 per year to make customers feel it is good value for money. This requires a deep understanding of the customer's business and sufficient data to support the benefits of AI . Value-based pricing is often reflected in negotiations with large corporate customers and can be flexibly customized. In the second half, some AI enterprise services also tend to emphasize ROI in marketing , charging according to the proportion of performance improvement, etc. 

  
  

When formulating pricing strategies, software companies should also consider the impact of AI on the original business model. AI 's improved efficiency may reduce users' demand for the number of seats (for example, AI customer service reduces the number of manual customer service), thus affecting the traditional SaaS model that charges by user seat. A VC study pointed out that if AI brings a 50% improvement in efficiency and assumes that customers reduce 50% of manual seats, then in order to keep the revenue unchanged, the software price needs to be nearly doubled.Although in reality customers will not completely cut manpower in this way, but may use the saved time to expand new businesses , manufacturers can reflect the efficiency bonus in the price when setting prices. For example, when Microsoft and other companies priced Copilot , they assumed that users would get about 40-50% productivity improvement, so they charged AI fees at about 50% of the base price..

In short, flexible and diverse pricing models will become a major feature of SaaS in the AI ​​era . Software companies need to constantly experiment with which strategies are accepted by the market. In practice, it is better to keep it simple first and then optimize : you can start with a surcharge or premium version, and when you have accumulated user usage data and value feedback, consider more granular usage or outcome charges. In addition, communication should be transparent so that customers know clearly what they pay and what they gain. For example, provide a calculator to estimate ROI , or clearly explain how AI functions save labor costs. It is worth mentioning that a recent survey showed that nearly 90% of medium-sized SaaS companies expect their business models to change moderately or significantly due to AI .This proves that formulating a new pricing strategy in the AI ​​era is a common issue faced by the industry and is also a part of the competition in the second half. Whoever can find the best pricing point to balance customer value and own profits will be able to take the initiative in business.

In summary, the second half of AI has put forward new requirements and opportunities for software companies. Technical executives and product managers need to work together to prepare from the four levels of organization, strategy, technology, and business : form a cross-functional team to integrate AI development with products; define problems and evaluate success based on user value, and truly embed AI into products; upgrade the technology stack to ensure that the AI ​​system continues to be effective and reliable; explore innovative business models to convert AI value into sustainable benefits. Only in this way can software companies remain invincible in the second half of this wave of AI , create a new generation of intelligent software products, meet the ardent expectations of users from all walks of life for AI , and achieve a win-win situation for technology and business.