After the Agent starts working, the three old rules of SaaS companies have become invalid.
For SaaS companies developing AI capabilities, the initial excitement phase has long passed.
Employees have begun to use AI tools, products are integrated with large language models, and internally built Agents for different scenarios are already deployed. Demos can run properly, and partial work efficiency is also improving.
Yet when Agents officially take over real business tasks, a series of more difficult problems emerge:
Who assigns tasks to the Agent?
What data can it access and modify?
Who is responsible for verifying its output results?
Who takes the fallback responsibility once an error occurs?
When AI takes over a large number of execution tasks, how should the original job positions be adjusted?
Do customers ultimately pay for the software, or for the completed tasks and delivered results?
These problems cannot be solved by simply switching to a more powerful model.
After Agents start undertaking formal work, the organizational structure, R&D and delivery rules that SaaS companies originally built around "people" and "software" all need to be readjusted.
01
Agents Are Officially on Duty
While Management Mechanisms Are Still Designed Exclusively for Humans
The management mechanism of most companies defaults that all work is completed by humans. Job positions belong to people, permissions are granted to people, and performance assessments are targeted at people. When something goes wrong with a task, you can always find the specific person in charge.
After Agents are introduced into the company, the management objects change, but the original mechanism does not automatically update to adapt to the new situation.
Some teams have already used Agents to generate reports, write code, and sort out requirements, but these practices often rely on spontaneous promotion by a small number of employees. There is no unified boundary for what Agents can and cannot do; there is also a lack of stable acceptance standards for whether the generated results can be used directly.
To adapt organizations to the introduction of Agents, at least four questions need to be answered:
First, which tasks can be assigned to Agents;
Second, what system and data permissions can Agents obtain;
Third, who will evaluate the quality of their work;
Fourth, who will be responsible for handling errors and taking fallback responsibilities after Agents make mistakes.
As the number of Agents increases, the work of managers will also change.
Neuters
In the past, project managers mainly arranged the division of labor, progress and collaboration of human employees. Now, they also need to manage the tasks, permissions, context and output quality of Agents. Departmental experience can no longer be scattered in the minds of individual employees, and needs to be precipitated into sustainably callable Skills, Contexts and working rules.
The AI-Native transformation of Minglue has covered different types of teams including functional teams, analyst teams, and R&D teams.
More than 4,300 work units have been formed internally, among which the number of AI agents exceeds the number of human employees. Real efficiency improvement has also taken place: the processing time of financial statements has been shortened from 3 days to 10 minutes, and the production of analysis reports has realized a high degree of automation.
Beyond the digital indicators, a more practical problem emerges: when AI takes over a large number of execution tasks, what exactly do the original employees, managers and business leaders do respectively?
In a recent interview with Wu Minghui, Founder, CEO and CTO of Minglue Technology, Neuters mentioned that the biggest challenge of AI transformation for traditional businesses is organizational restructuring.
New businesses can design positions and processes from scratch, while old businesses have already formed a stable division of labor, assessment and collaboration relationships. After Agents are introduced, they will touch the original work boundaries. Even if the technology runs smoothly, the organization may not be able to adapt immediately.
For SaaS companies, to measure the organizational maturity of AI, we cannot only look at the employee usage rate and the number of Agents. More importantly, it is necessary to check whether the company has redefined tasks, permissions, acceptance standards and responsibility mechanisms around Agents.
02
Many Companies Are Eager to Develop Agents
But They Have Not Completed the Construction of the Previous Three Layers of Capabilities
Agents are very popular now, and many SaaS companies hope to directly enter the Agent development stage in one step.
In reality, the deployment of a set of Agents into production often gets stuck at links outside the model: enterprise knowledge is still scattered in documents, chat records and personal experience; business processes have not been clearly split, and a large number of abnormal scenarios rely on employees to handle on the spot; AI cannot obtain complete context, nor can it call key systems; there is a lack of evaluation for output quality, and there is no rollback and takeover mechanism after errors occur.
Agents seem intelligent enough, but they cannot run stably after being put into real business scenarios.
The transition from Chatbot to Agent is not a simple model upgrade. Every step forward for AI requires access to more context, process nodes and execution permissions.
Chatbot solves the problem of "easy access to information". Employees can query knowledge and generate content, but the results still depend on individual user prompts.
Copilot solves the problem of "higher work efficiency". AI starts to enter positions such as product, R&D and analysis, helping employees complete single-point tasks such as document writing, coding and data analysis. Personal efficiency is improved, but experience may not be reused by the whole team.
Workflow solves the problem of "smooth process execution". Enterprises split a piece of work into relatively stable steps, allowing AI to participate in links such as requirement collection, review, generation and testing. The process can run repeatedly, but human judgment is still required when encountering exceptions.
Agents start to advance tasks around preset goals. They need to obtain context, call tools, collaborate with other Agents, and take actions within a certain scope of permissions.
At this stage, enterprises must establish new production rules: which tasks are allowed to be executed automatically, which nodes require manual approval; how to divide labor among Agents; how to evaluate outputs; how to detect, roll back and trace errors.
Huang Nan, Director of the AI-Native Organizational Transformation Lab of Minglue Technology and Head of AINOL, once demonstrated a real case: after users reported a Bug, multiple Agents participated in requirement processing, review, design, development and testing, and the repair was completed and launched in 1 hour and 49 minutes.
What is easy to remember in this case is the speed, but what really supports this result is the long-term accumulation behind it.
In short, Minglue's AI-Native practice has gone through the promotion process from Chatbot, Copilot, Workflow to Agent. At each new layer, enterprises need to supplement new basic capabilities: how knowledge is formed into Context, how processes are precipitated into Skills, how Agents obtain tools and permissions, and at which nodes human judgment should be retained.
Many Agent projects cannot run stably, and the problem may not lie in the final layer. If the previous knowledge, process and collaboration mechanisms are not completed, even the most powerful Agent can only stay in the demonstration stage.
03
Customers Start to Pay for Results
While the Delivery Model Still Stays at the Stage of Selling Software
Traditional SaaS has a relatively clear delivery model.
Software companies charge by accounts, modules or service years. After the product is launched, the implementation team helps customers configure the system, and subsequent work is mainly completed by the customers themselves.
After Agents are introduced into business processes, the focus of customers is changing.
Customers will continue to ask questions: What work can this Agent do for me? How much time and labor can it save? What if the result fails to meet the requirements? Who is ultimately responsible for it?
The closer AI is to actual work, the higher customers' expectations for results will be.
In the past, what was delivered was a set of usable software. Now, it may be necessary to deliver a task that can be completed continuously, or even a team of digital workforce.
This transformation needs to answer at least three questions.
First, what to sell. Are customers purchasing software tools, the completion of a specific task, or the final business result? Different answers correspond to different product forms, pricing and responsibility boundaries.
Second, how to conduct acceptance. Is the acceptance based on whether the function is launched and how many man-hours are invested, or based on efficiency, quality and business indicators? When an Agent makes an error, what responsibilities should the software company, the customer and the delivery team take respectively?
Third, how to realize reuse. If every time you enter a customer site, you need to re-sort out processes, develop tools and train Agents, the digital workforce will easily become a new round of heavy customization. Project experience must be precipitated into Skills, Tools, data and platform capabilities to be reused in the next delivery.
The value of FDE should also be understood in this entire value chain.
FDE is not just an implementation staff with a new name, but a role that goes deep into the customer site, identifies real business problems, and connects customer scenarios, AI capabilities and delivery results.
Wu Minghui also mentioned that Chinese enterprises have their own unique customer structure and delivery environment, and cannot simply copy Palantir's FDE model.
SaaS companies need to rethink: what exactly are customers willing to pay for, how to accept the delivered results, how to precipitate human experience into reusable capabilities, and how to control the customization cost of Agentic Services.
This article is from the WeChat Official Account "Neuters" (ID: Neuters), Author: Cuiniu Club, published with authorization from 36Kr.