♪ Take notes ♪The rules of honesty at this station.
I think the corporate AI transition can be divided into five phases.
Knowledge first, personal effectiveness, process optimization, architecture optimization, and finally AI.
These five stages are not simply from ease to difficulty, nor do they mean that each enterprise must go through it in its entirety. It’s more like AI getting into a business. Initially, the boss knew what AI was, then the employee started using AI itself, then AI entered the company ’ s workflow, and then the company began to adapt its own technical and organizational structures in order to stabilize the processes. It is only then that the company ’ s products, jobs and methods of collaboration are designed around AI itself.
A lot of companies come up to buy an AI platform, or a corporate knowledge base, and I feel wrong. You don’t even know what AI can solve, and no one really does work with AI, and the platform that comes up at this point is probably just a new setup.
The architecture should not be designed in a vacuum. The structure should grow from proven individual usage and workflow.
Phase one, cognitive first.
Knowledge first means that management needs to have a basic understanding of AI capabilities.
There are two types of management.
One type of management begins at the grass-roots level. He himself had done what the employee was doing, knew how a form came from, knew why a client was late in entering into a deal, and knew where the complexity of what staff said was “the process is complicated”.
Such managers tend to understand more quickly what AI can do. Just show him a few real cases, he can immediately think about his business.
Another management is not familiar with grass-roots work. He is responsible for managing, talking about clients and redeploying resources, but does not personally perform the tasks that staff do on a daily basis.
In such cases, managers ’ judgement of AI is easily blocked by staff attitudes. The employee may say that this is not possible, that the data cannot be given to AI, that the process is too special, or that the content generated by AI is not useful at all.
Those words may not be wrong. Staff may be concerned about AI replacing themselves, or the short-term workload may have increased after the introduction of AI. There is also a situation where the most rational choice is, of course, to refuse to change when it is done without additional gain and when it is broken and liable.
So cognitive first isn’t about telling the boss what a big model is, nor is it about showing a couple of fancy cases. What really needs to be done is to provide management with basic independent judgement.
He should at least be able to judge:
- What is primarily the reading, collation, judgement and output of information;
- What jobs can be done by AI first, then people check;
- What involves real responsibility must remain manual;
- Which jobs should not be touched now;
- Is an AI project really worth it, or is it just a demonstration?
At this stage, what we can provide is corporate AI training and transformational diagnostics, and first and foremost for management.
We’re going to go into the business, we’re going to look at real jobs, we’re going to talk to managers, we’re going to have employees, and we’re going to make a list of all of them. It is important to clarify which scenes are suitable for immediate trial, which ones require data preparation, which ones are too risky and which ones look advanced but need not be done.
The sign of success in this phase is not that all people say they value AI, but that management is able to articulate three worthy scenarios and know the objectives, success criteria, duty bearers and risk boundaries of each scenario.
Phase two, personal effects.
When management has the basic knowledge, we have completed our preliminary judgement, and the next step is personal effectiveness.
There is a need for staff training here, but I don’t feel the tools are working. The training was not complete, as staff members who had completed their classes felt that AI was very good and returned to their jobs without knowing what they should enter and how to hand over their work to AI.
For truly effective training, this post must be used directly for tasks that are dealt with daily.
For example, how executives organize meeting records, how sales investigate clients and write follow-up mail, how finance reconciles forms, how tenderers search for projects, read documents and prepare applications, and how programmers get Agent to modify codes, run tests and check results.
So there are several levels of personal effectiveness.
The first level is generic training. To make staff understand the difference between chat AI and Agent, learn to upload documents, describe needs, judge results, and protect privacy. This layer is the easiest to standardize and the best for large-scale delivery.
The second level is job training. For different jobs, real jobs are broken down into a series of tasks that can be repeated. There is a clear premium for job training, because we are no longer providing a tool but a reorganization of the job.
The third level is the job manual. Instead of keeping 100 hints in mind, staff members were told what tools, what materials should be entered, what steps should be taken, what results would be successful and how they should be recovered from time to time.
We should consider using it as a tool or a SaaS product if there are multiple clients, multiple jobs and the same problem.
Instead of thinking about a SaaS and then looking for people to use, in the course of training and service, it became apparent that everyone was doing the same thing over and over again, and then the product naturally appeared.
The success of the individual effectiveness phase is marked by at least three AI ways in which an employee can work independently, and we can measure how much time it saves and how many mistakes it reduces. Delivery will be genuine only if the staff member is able to do it himself after leaving the training site the next day.
This is the best business I’ve ever had. It is easy to demonstrate that delivering borders is clear and does not require a system to take over business from the outset, and each training can help us understand a new industry and position.
Phase III Process optimization
Individual efficiency solves how a person works faster, and process optimization solves how a group of people and multiple systems work together to do a thing.
For example, the bidding exercise does not allow staff to read a tender document using AI. The complete process may be:
寻找招标信息
→ 判断项目是否适合报名
→ 登记项目台账
→ 准备报名材料和邮件
→ 人工审批
→ 发出邮件
→ 获取招标文件
→ 跟踪截止时间和下一步动作At this point, the question is no longer whether AI will generate content. It involves the responsibility of multiple jobs, where the data come from, who has the authority to operate, when manual confirmation is required, who is notified of the failure and where the status of each step is recorded.
The difficulty and cost of process optimization will therefore increase significantly.
I had previously felt that process optimization could only provide consulting services to enterprises, that it was difficult to make direct changes and that it was therefore not a good business. Now I feel that this is only half the judgement.
Process optimization is indeed not a business suitable for low-cost standardized delivery. If an enterprise is not clear about the process itself, it is easy for an outside team to go in and become infinite. When you finish one step, you’ll soon find another piece of data that is not standard, and the next one is one that the responsible person is not willing to cooperate.
But process optimization is also a place where there is real value. Because once a high-frequency process is running through, the savings are no longer a 10-minute staff member, but a long-term duplication of time across the department. It also creates strong customer adhesion and industry barriers.
The correct approach is not to come up with a whole company, but to choose only one process at a time with clear borders, sufficiently high frequency and quantifiable results.
We need to turn this process into an observable status machine: what is the input, what is going to be done, who is responsible for the next step, under what circumstances manual approval is required, under what circumstances it fails, and how it will be restored after failure.
The integration of the business list, the automation of the bid rolls and the consistency of financial practices are in fact process optimization.
This stage of success is marked by a real process running from the beginning to the end, with clear input, output, status, responsibility, clearance gate, unusual handling and operational records. It must also be able to prove that the cycle has been shortened, that errors have been reduced or that operations that were otherwise unmanageable can now be addressed.
We should be careful at this stage in choosing projects. It is not the client who says he wants to automate what he wants to do, but it is the client who first decides whether the process is stable, whether the data are available, whether the person in charge is willing to cooperate and whether the results are acceptable.
Phase four: Structural optimization
When an enterprise has only one AI process, there is absolutely no need to talk about AI medium.
But when different sectors start using AI, problems arise. Each department repeats the same information, each system has its own account numbers and privileges, and each Agent does not know what other Agent has done, model costs cannot be measured, errors do not have logs and data leaks do not know who called them.
This is when the architecture needs to be optimized.
The structure here is not just technical but also organizational.
The technical architecture includes the enterprise knowledge base, databases, systems interfaces, model gateways, account numbers and privileges, Agent movement control, log auditing, testing, roll-back and disaster preparedness.
The organizational structure includes who sets out the objectives, who is responsible for AI results, who approves high-risk operations, who maintains the knowledge base, who handles anomalies and how tasks are transferred between different departments.
The real problem for the architecture to optimize is that the new AI process will not start from scratch. A new process should be able to directly re-utilise an enterprise ’ s existing identity, authority, data, knowledge, modelling and log system.
It is difficult for this phase to be taken over directly by external teams. Because it touches the core of data, accountability and systems within the enterprise. Without sufficient management support or a proven process ahead, the architecture project could easily become a costly but unusable platform.
So structure optimization cannot run.
It should grow from multiple successful processes. We start by making specific processes, then by observing which capacities are used repeatedly and then abstracting them into the public infrastructure of enterprises.
The success of this phase is marked by the fact that, when enterprises add an AI scenario, they can build and validate quickly on existing structures instead of repurchaseing a set of tools, re-collating data and redesigning permissions.
At this stage, we can participate in the design of the architecture, or we can provide partial realization, but we should not simply commit to taking over the full AI restructuring of a traditional enterprise.
Phase V. Primary AI
It was not the original AI that installed several Agents for traditional enterprises, nor did it connect the original system to a large model.
The original AI means that the products, processes, jobs and collaborative approaches of a company were designed around AI from the outset. If AI was removed, the company could not even be organized.
I feel that traditional businesses can always get close to the original AI, but it’s very difficult to change it completely. The real AI is more likely to be an AI-era start-up, most likely from an OPC expansion.
One person first uses AI to do what one would need a team to do. When the business increased, he did not immediately recruit a large number of staff on the traditional company ’ s basis, but started by breaking down the work into tasks, processes and status, leaving a small number of people to manage a separate group of Agent.
Future companies, people-to-people forms of cooperation may no longer rely on frequent meetings and meetings. Everyone faced an AI assistant of their own, and the company had a complete and clear knowledge base and database, and the whole company was like a giant-state machine.
Staff do not have to meet every day to report on what they have done. His mission, progress, products, problems and next steps have been systematically documented. Nor does the boss need to keep asking employees where they are now, and he can find out the real state of the company through his own AI.
There is also no need to synchronize the context between staff and staff. When a person completes his or her mission, leaving a record of the mission, a real product, test evidence and a handover report, the next person can continue to work.
It’s actually like GitHub’s collaborative approach.
Two different countries, different professions, and people who have never met, can work together on a project through Issue, Commit, Pull Request, Review and CI. They need not first establish complex interpersonal relationships, but simply agree on issues, interfaces, products and acceptance standards.
My Jarvis master and project manager system is also building on this collaborative approach.
Jarvis did not ask for a continuous dialogue between each and every Agent. Master definition objectives and success criteria, independently implemented by the responsible person, after which the products, evidence and reports are submitted, and the Master is responsible for receiving and inspecting the reports and the next route.
At the heart of this approach is not the opening of additional Agents, but the transformation of complex tasks into differentiating work that can be assigned, accepted, restored and transferred.
Of course, it’s not like people don’t talk to each other anymore. Trust, negotiation, responsibility, creativity and true relationships remain human beings. What AI should really reduce is a huge amount of low-value synchronization, not the elimination of human connections.
The success of AI in its original phase is marked by the emergence of new products, services or new forms of organization that businesses could not have built without AI in the past, and by its ability to generate real income. Business expansion is no longer largely dependent on increasing numbers but on better processes, stronger Agent and clearer knowledge systems.
What should we do at this stage?
If we’re going to turn these five stages into our own business, I feel the route is clearer.
Phases I and II are the most accessible and easy to generate cash flows at the moment. Through management training, enterprise diagnostics, generic training, job training and job manuals, we can enter enterprises to understand real needs, while building on curricula, cases and industry knowledge.
Phase III cannot be completely abandoned, but must be chosen. Processes that are clear, accountable and measurable only. Process optimization, while heavy in delivery, can create high unit prices, customer stickiness and real industry barriers.
For the time being, phase IV focused on understanding and accumulation. Don’t build an AI center for customers to look advanced. It is only when multiple processes have run out and public capacities have really re-emerged that the structure is abstracted.
Phase V should not be used simply to tell a story to a client, but should first be done by ourselves. Jarvis, StarPeer, the Control and Responsible Mechanism, the handover of reports, the knowledge base and the automated processes are all our experiments with the Aichi organization.
That is why we offer corporate AI transformation services abroad, while at the same time testing the end of corporate AI transformation with our own companies.
In the end, we may not sell a tool or a training, but rather a way for businesses to move into the AI era:
管理层获得判断力
→ 员工形成个人AI工作方式
→ 一个真实业务流程完成闭环
→ 多个流程共享统一基础设施
→ 企业形成AI原生产品和组织方式Do not sell platforms to businesses from the start. Let’s make a person really fast, then a process really runs. The platform grew from a successful process and the AI itself from a successful structure.