Many business leaders have recently noticed a clear trend: foundational models are becoming larger, with increasingly comprehensive capabilities. Coupled with enhanced reasoning technologies improving efficiency and gradually lowering costs, along with the conveniences provided by the open-source ecosystem, and products like AI Agents that can be implemented—enterprise digitalization seems to have suddenly hit the gas pedal, and a significant transformation is imminent.
However, when it comes time to take action, many people feel confused: where should digitalization actually head? The underlying logic has already changed—it is no longer about ‘one algorithm solving one scenario’ through scattered attempts. The goal of B-end products has shifted from ‘optimizing a specific link’s efficiency’ to ‘upgrading the operational capabilities of the entire business chain.’ In simple terms, today’s digital products must evolve from being ‘pure tools’ to ‘powerful instruments that can integrate into the business and drive growth.’
This change first forces key decision-makers in enterprises to upgrade their understanding. For instance, many people ask: is a large model a ‘cost center’ that incurs expenses, or a ‘profit center’ that can generate revenue? How can we quickly find suitable large models to optimize business processes in the short term? How should organizational and product mechanisms be adjusted to adapt to new business models?
In fact, our expectations for AI Agents have long surpassed mere ‘showcasing technology’; we hope they can help enterprises break through growth bottlenecks. However, the reality is that many enterprises are stuck in the pitfalls of ‘uncertain scenarios,’ ‘ineffective implementations,’ and ‘difficulties in scaling replication.’ The core issue is quite simple: they have not grasped the essence that ‘technology must serve the business.’
Next, I will break down how enterprises can truly generate business value from AI Agents from three dimensions: ‘setting direction, finding paths, and enhancing value.’ Let’s start with the first step, ‘setting direction.’ If you find this information useful, feel free to share it with friends who might need it.
01
Set Direction: Don’t Chase Technical Trends
First, focus on ‘high-value scenarios.’
Many enterprises stumble when deploying AI Agents: either they aim to cover ‘the entire process’ all at once, or they follow the trend of creating general large models, resulting in high investment and slow returns. In fact, in the early stages, it is not necessary to be overly ambitious; the key is to find the ‘anchor points’ where large models can precisely empower the business.
1. First, clarify: the ‘business penetration capability’ that enterprise-level AI Agents must possess.
Let’s distinguish: consumer-level AI can chat with you and write copy, but enterprise AI Agents are different—they must truly integrate into the business, achieving ‘understanding of the business, implementability, and security.’ Therefore, enterprise-level AI Agents must have business penetration capabilities, which are reflected in three dimensions: the autonomous closed-loop capability of the Agent’s tasks, the understanding of business context, and data controllability and security. For example, the ‘autonomous closed-loop capability’ does not mean it can only execute a single command; rather, it can break down steps and adjust tools around business objectives. Take ‘conducting a monthly sales review’ as an example: the Agent will pull sales data from the CRM system for various regions, compare it with last year’s and last month’s data in the BI tool, and if it finds that sales in a certain region have suddenly dropped by 20%, it can automatically analyze whether it is a channel issue or if promotions have not kept up, ultimately consolidating this into a report with improvement suggestions.
Similarly, the ‘business context understanding capability’ is crucial; the Agent must understand the ‘know-how’ and processes within the industry. For instance, in manufacturing, if AI needs to complete a ‘BOM change,’ it must first understand that this change will affect procurement plans and production schedules. In retail, if AI needs to complete ‘user segmentation in private domains,’ it must know which dimensions to segment based on different business forms and what outreach strategies to use for different levels of users. This requires fine-tuning with industry data or integrating ‘RAG business knowledge plugins’; otherwise, the Agent may easily end up in awkward situations, such as when you ask it ‘how to handle customer complaints,’ and it responds with a bunch of generic etiquette, which is meaningless.
Moreover, the ‘data security and controllability capability’ is the bottom line for enterprises. AI Agents interface with core data, such as customer contact information and financial data, so they must support ‘private deployment,’ automatically desensitize sensitive data, and ensure that every operation—who accessed the data, what report was generated—has traceable logs for compliance auditing. Otherwise, if compliance lines are crossed, it could lead to significant losses.
2. Next, filter for ‘quick implementation and high ROI’ priority scenarios.
Different industries have different pain points, and the value density of AI Agents varies. Considering the current maturity of technology, scenarios in manufacturing, retail, and enterprise services are the most worthwhile to prioritize for trials, as they can be implemented quickly and yield tangible returns.
The first is manufacturing: focusing on ‘supply chain collaboration’ and ‘production anomaly response.’
Those in manufacturing know that the biggest fear is ‘bottlenecks’ and ‘late problem detection.’ For example, in the past, adjusting production plans required manual coordination with procurement, warehousing, and logistics departments, leading to slow information synchronization and often resulting in situations where ‘procured materials have not arrived, but the production line has already stopped.’ Product defects were detected through manual sampling, and by the time issues were discovered, they often involved batch production, leading to significant waste.
AI Agents can connect the data flows of ‘production orders, material procurement, inventory, and quality inspection’: once the production plan is modified, it automatically pushes procurement needs to the procurement department while checking inventory levels; during production, it monitors equipment data and product appearance in real-time, and if it detects defects such as dimensional deviations, it immediately issues a warning and can halt the production line. One home appliance company that implemented this reduced its production plan adjustment cycle from 3 days to 4 hours, and the defect rate dropped by 18%, resulting in considerable cost savings.
The second is retail: primarily targeting ‘omnichannel marketing’ and ‘user operations.’
The pain points in retail are straightforward: marketing often feels like ‘scratching an itch through the boot,’ and creating materials is labor-intensive and time-consuming. In the past, writing marketing copy required manual adaptation to different styles for platforms like Douyin, Xiaohongshu, and private community groups; recommending products relied on fixed rules—such as recommending B after A was purchased—making it difficult to achieve ‘personalized experiences.’
AI Agents can conduct precise operations based on user tags; for example, for maternal and infant users, it will consider ‘baby age’ and ‘purchase frequency’: recommending ‘pregnancy care knowledge + discounts on maternity wear’ to newly pregnant users, and ‘complementary food tutorials + infant product bundles’ to users with 1-year-old babies; it can also automatically adjust the style of copy without manual oversight—using lively short sentences for Douyin and adding emojis and scene images for Xiaohongshu. A certain chain coffee brand saw a threefold increase in marketing material production efficiency after trying this, and the conversion rate of private domain users increased by 0.8 percentage points—this 0.8% is significant, translating to over a million in additional monthly revenue for them.
The third is the enterprise service sector: addressing ‘knowledge dispersion’ and ‘inefficient cross-department collaboration.’
Those in enterprise services often encounter two types of troubles: new employees must sift through dozens of documents to understand the business; cross-department projects rely on meetings to synchronize progress, often missing key milestones, such as ‘the R&D department’s product prototype is not ready, but the marketing department is waiting to launch the promotion plan.’
AI Agents can serve as the ‘central hub of enterprise knowledge’: integrating product manuals, process specifications, and historical project cases. When a new employee asks ‘how to handle customer complaints,’ it directly provides steps and scripts without needing to sift through documents; it can also act as a ‘project collaboration assistant,’ automatically synchronizing progress across departments—such as whether the R&D department’s prototype is completed and when the design department’s drawings are due, and it will remind the marketing department to produce the promotion plan within 2 days as the deadline approaches. It is reported that a tech company that implemented this reduced new employee training time by half and decreased cross-department project delays by 25%.
02
Find Paths: From 0 to 1 Validate, From 1 to 100 Replicate
Avoid Long Cycles of Trial and Error
1. From 0 to 1: Complete the closed loop in 3 months, validating value at minimal cost.
The core of this stage is ‘not being overly ambitious, but focusing on a single high-frequency pain point.’ Specifically, this can be approached in three steps: ‘scenario focus → lightweight resources → effect quantification,’ which requires minimal investment, yields quick results, and builds confidence within the team.
First step, scenario focus: choose ‘high repetition’ scenarios; do not aim to create an ‘AI assistant usable by the entire company’ right away, but prioritize tasks that ‘a specific role must perform daily, which are repetitive and time-consuming.’
For example, in the finance department, ‘invoice review’—with thousands of invoices each month, accountants must verify amounts, check tax rates, and ensure the header is correct, which can lead to errors due to prolonged focus, and they cannot review many in a day; similarly, in the customer service department, ‘common question responses’—over 60% of inquiries are about ‘order tracking,’ ‘how to apply for after-sales,’ and ‘how to use coupons,’ leading customer service personnel to repeat the same phrases daily, which is particularly exhausting. These types of scenarios have the advantage of abundant data and clear rules, requiring no complex technology, and can be launched in 1-2 months, making them ideal for ‘practicing’ and validating value.
Second step, lightweight resources: leverage ‘low-code + mature platforms,’ and do not build large models from scratch. Many bosses think that to create an AI Agent, they need to ‘hire a large team and build a large model,’ but this is entirely unnecessary. Initially, we can ‘stand on the shoulders of giants’—using existing enterprise-level AI Agent platforms like Alibaba Cloud and Microsoft Azure, we can connect to the company’s existing systems through ‘low-code configuration,’ such as customer service systems and financial software; if industry-specific capabilities are needed, we can install ‘industry plugins,’ such as financial review plugins or customer service script plugins, without writing code from scratch.
One e-commerce company was particularly clever when developing a customer service AI Agent: they assigned only 2 product managers to set rules (for example, ‘if a user asks about an order that cannot be found, first ask for their phone number, then pull data from the order system’), and 1 developer to connect to the existing customer service system, completing the launch in 3 weeks, initially covering 10 common questions, resulting in a 40% reduction in manual handover rates—such a small investment yielded immediate results, making the boss willing to continue investing.
Third step, effect quantification: use ‘business metrics’ to speak, and do not discuss ‘how advanced the technology is.’ After going live, do not just tell the boss ‘the AI Agent is very intelligent’; this is useless—real business data must be presented. For example, for the customer service AI Agent, look at ‘how much the manual handover rate has decreased’ and ‘how much the average response time has shortened’; for the financial invoice AI Agent, track ‘how much the average daily review volume has increased’ and ‘how much the error rate has decreased.’
One enterprise’s finance department saw the average daily reviewed invoices increase from 50 to 200, and the error rate drop from 3% to 0.5%, resulting in a payback period of just 3 months—such results speak for themselves, and the boss will naturally support further advancement.
2. From 1 to 100: Scale from 6 to 12 months, building a ‘business moat.’
Once single-point validation is successful, it is time to upgrade from ‘usable by a specific role’ to ‘usable across departments.’ The core of this stage is ‘do not reinvent the wheel’; instead, ‘accumulate capabilities, adapt the organization, and run the data flywheel’ to ensure that the AI Agent truly integrates into the business rather than becoming an ‘additional burden.’
First, capability accumulation: build an ‘AI Agent technology platform’ to enable new scenarios to be implemented twice as fast. Do not start from scratch for every scenario—take previously validated capabilities, such as ‘task breakdown logic,’ ‘system interface connections,’ and ‘industry knowledge templates,’ and turn them into ‘modular components.’
For example, commonly used ‘equipment data access components’ in manufacturing and ‘user tag analysis components’ in retail can be directly assembled when implementing new scenarios, eliminating the need for development from scratch. One automotive parts company did this, and after building a platform, the implementation cycle for new scenarios like ‘supplier risk warning’ was reduced from 2 months to 2 weeks, with R&D costs dropping by 60%—this is the value of ‘accumulating capabilities,’ making it easier as time goes on.
Second, organizational adaptation: let ‘business departments lead, with technical departments supporting,’ and do not let technology work in isolation. The most common issue during the scaling phase is that ‘the technology department creates something that the business department does not use’—the technical team believes ‘the features I developed are advanced,’ but the business team feels ‘this thing does not fit into my daily work and is cumbersome to use.’
To solve this problem, it is essential to clarify that ‘business leaders are the owners of AI Agent applications.’ For example, for the sales department’s AI Agent, the sales director should take the lead: they know best what the sales team needs (for instance, ‘it should automatically track customer follow-up progress and remind when follow-up deadlines approach’), and they can drive the team to change processes; the technical department is responsible for ‘whether it can be implemented,’ such as optimizing model performance and resolving system integration issues. One group company specifically established an ‘AI Agent special team,’ with one representative from each business line joining, covering sales, supply chain, and finance within six months, without any ‘disconnect between technology and business.’
Third, data flywheel: use ‘business data to feed back into the model,’ making it smarter over time and forming a barrier. What is the greatest advantage of an enterprise? It is ‘its exclusive business data’—this is something that general large models cannot compete with. What we need to do is utilize this data to establish a closed loop of ‘data collection → model optimization → effect feedback,’ allowing the AI Agent to become increasingly effective.
For instance, an AI Agent in the retail industry collects user click and purchase data on recommended content, gradually discovering that ‘pushing parenting knowledge + products has a 30% higher conversion rate than simply pushing products,’ and will adjust its recommendation strategy accordingly; in manufacturing, as more equipment failure cases are accumulated, risk warnings will become increasingly accurate.
One machinery company improved the accuracy of its equipment failure warnings from 75% to 92% after six months of data iteration, preventing several production losses. Once this data flywheel starts turning, the effectiveness of the AI Agent will continue to strengthen, gradually becoming the enterprise’s unique ‘technical barrier’—others will find it hard to replicate because they do not have as much exclusive business data as you do.
03
Enhance Value: AI Agents Are Not Just Tools
But Innovators of Business Models
At this point, some may still think, ‘Isn’t an AI Agent just a tool for improving efficiency? It’s good enough to save some manpower.’ However, what it can bring is a fundamental transformation in three dimensions: ‘efficiency, data, and management’—not just ‘cost reduction,’ but also helping enterprises ‘increase revenue’ and even change the logic of business operations.
1. Efficiency Reconstruction: From ‘people finding processes’ to ‘processes finding people,’ freeing people to focus on high-value tasks.
Traditional automation tools, such as RPA, can only optimize ‘individual links,’ such as automatically entering data, but AI Agents can ‘reconstruct entire processes.’ Think about it: in the past, employees processing sales orders had to first log into the CRM to check customer information, then log into the ERP to record orders, and finally log into the BI for analysis, with the entire process being ‘people chasing processes’; switching systems alone took a lot of time.
Now? AI Agents will automatically pull customer data, generate orders, complete analyses, and only push ‘decision-making nodes’ to employees—such as ‘this customer’s order amount exceeds 1 million, requiring your approval.’ One fast-moving consumer goods company saw its sales order processing efficiency rise from ’30 orders per person per day’ to ’80 orders per person per day’ after implementing this.
More importantly, employees no longer have to perform repetitive data entry and checking tasks, allowing them to spend time on ‘communicating with customers’ and ‘identifying needs,’ which is a high-value activity—equivalent to one person doing the work of 2.5 people in the past, while also generating more revenue; this is the true ‘efficiency revolution.’
2. Data Activation: Turning ‘sleeping data’ into ‘profitable assets’; data becomes valuable when it moves.
Many enterprises have a wealth of ‘sleeping data’: sales data in CRM, inventory data in ERP, production data in MES, but these data are siloed and do not know how to interconnect—such as when sales data increases, the procurement department does not know to stock up; production data shows that a certain product has sufficient capacity, but the marketing department does not know to promote it more.AI Agents can serve as ‘data hubs,’ connecting data from ‘people, goods, finance, and assets’ to achieve a closed loop of ‘data → insights → actions.’
For example, one retail enterprise’s AI Agent integrated ‘user consumption data’ from CRM, ‘inventory data’ from WMS, and ‘activity data’ from the marketing system: it discovered that a certain store’s snack sales had increased by 50% in the past 7 days, but the inventory only had enough for 3 days, immediately pushing a suggestion to ‘restock 200 units’ to the procurement department while reminding the store to implement ‘limited-time discounts’ to avoid stockout losses. After implementation, this enterprise reduced its inventory turnover days from 30 to 22 and decreased stockout rates by 15%—once data comes to ‘life,’ it can help enterprises spend less and earn more.
3. Management Upgrade: From ‘experience-driven’ to ‘engineered management,’ bidding farewell to ‘doing things based on mood.’
Traditional enterprise management often relies on ‘the experience of veteran employees’—for example, when sales follow up with customers, veteran salespeople know which questions to ask and how to advance, but new salespeople may overlook key needs; quality inspectors determining whether products are qualified may have different standards, leading to ‘execution deviations.’
AI Agents can transform ‘excellent experiences’ into ‘standardized processes,’ making management more controllable. For instance, it will guide new salespeople to follow steps based on SOP (Standard Operating Procedures): ‘first communication should clarify customer budget and needs, follow up on any changes in needs after 3 days’; it will also track project progress based on PDCA (Plan-Do-Check-Act)—if a certain region’s sales do not meet targets, it will automatically analyze the reasons, such as ‘the new customer conversion rate is 10% lower than average,’ and suggest ‘enhancing new customer training’; it can even synchronize the completion status of each department’s OKRs in real-time, such as ‘the marketing department only achieved 80% of new customer acquisition in Q3, and needs to prioritize offline activities.’
One tech company that implemented this saw a 30% increase in departmental collaboration efficiency, and the target achievement rate rose from 75% to 90%—this is the benefit of ‘engineered management’: no longer relying on ‘feelings,’ but on ‘data and processes,’ with standardized execution leading to better outcomes.
By 2025, AI Agents will no longer be a question of ‘whether to implement’ but rather ‘how to do it well’—they have already become the ‘infrastructure’ of enterprise digitalization. For business leaders and B-end product managers, the key to deployment has never been about ‘how advanced the technology is,’ but rather ‘whether it can solve business pain points and generate quantifiable value.’
So stop worrying about ‘whether to develop large models in-house’; instead, think about ‘which scenario to start with’—if you choose the right scenario and follow the right path, AI Agents can become the ‘new engine’ for enterprise growth. After all, when your competitors are already using AI Agents to improve efficiency, reduce costs, and expand revenue, being a step behind could mean falling behind an entire era.
PS:
This article’s insights are primarily derived from Jiazi Guangnian, whose founder shared a keynote at the Alibaba Cloud Qixi Conference: how Agents drive industry transformation. Since the emergence of ChatGPT, the industry has been filled with fresh developments daily. The first phase mainly focused on model evolution, but at this point, AI applications are blossoming, even creating many previously unimaginable product opportunities and business miracles.
We are in the midst of this, often feeling a mix of emotions. On one hand, there is anxiety about being left behind by the times. On the other hand, there is excitement, as we feel that the next opportunity may belong to our industry. You know that when two types of thinking coexist in a person’s mind, it inevitably requires reconciliation and balance. Regardless, let us continue to learn, maintain our thoughts, and create meaning together!