Last week, while reading the materials from Kingdee's Global Visionary Conference, there was one field I read twice: Responsible Department and Responsible Person.
It's not a product feature, and it's not a model parameter. It's a field on their internal agent operations management platform. The platform manages more than 180 agents, and on the dashboard you can see how many tasks each has cumulatively executed, what they produce, how many credits they consume, and how ROI is calculated—then, one by one, each is labeled with which department and which person it belongs to.
This field is more grounded than all that talk about "AI-native organizations." Because it acknowledges one thing: These agents are no longer features; they are execution entities that need to be brought into the formal establishment and have an owner.
If a feature breaks, contact IT; if a document is wrong, who do you contact?
1. 180 executing entities and 180 switches are not the same thing
In the past, when we implemented an automation, the deliverable was a set of rules, a button, and an operation manual. It would not decide on its own when to run, nor would it modify data when no one was watching. This batch of agents is different: they read emails, recognize documents, generate vouchers, initiate approvals, trigger automatically according to schedule, and can even be woken up by scheduled tasks. Once given write permissions, it is an account that can work on its own.
Once the number goes up, the nature changes. One agent having problems is debugging; one hundred eighty agents running is management. You will start running into very HR-type problems: which one is still in use, which one has long been left unmanaged, which one burns the most points each month but produces little, after which version changed the instructions a certain billing error started appearing, and which department to contact when something goes wrong.
That Kingdee dashboard lays all of this out and adds an even more critical point: the responsible department and responsible person. This is the same principle as when we assign an owner to every interface and every scheduled task during ERP implementation—except the object has changed from a process to something that makes judgments.
The number of agents managed by Kingdee's internal agent operations management platform; the dashboard records, for each agent, the number of tasks executed, outputs, points consumed, and ROI calculations, and indicates the responsible department and responsible person.
By the way, this set of tools is not just for our own use. In contemporaneous public disclosures, as of October 23, Lingji had 3,706 paying customers; IDC MarketScape’s “China Enterprise Resource Management Market AI Agent Vendor Assessment” placed Kingdee in the “Leaders” category, with its total scores on the strategy and capability dimensions ranking first among Chinese vendors. With more customers, vendors must also first manage their own teams well.
II. How to Redraw Roles: Separate "Action" and "Knowledge"
In the November 7 report from Xinhuanet, Zhang Eyu, head of Kingdee's finance line, gave a fairly clear breakdown. In the past, finance was divided by roles such as invoice collection, bookkeeping/entry, accounting, and business finance; the future form is financial experts + digital employees, with digital employees responsible for perception, calculation, and execution ("doing"), and humans responsible for decision-making, judgment, and accountability ("knowing"). In terms of architecture, it comes down to three layers: the underlying intelligent hub layer where agents execute autonomously, the middle governance layer, and the top layer of human value judgment.
I've never been very keen on talk about "AI replacing people," but this distinction is worth remembering. It doesn't say people are exiting; it says the granularity of jobs has changed—from narrowly defined functions to "one expert leading a cluster of agents." People's work shifts from processing item by item to setting standards, watching for anomalies, and signing off.
That's also where the difficulty lies. Responsibility doesn't disappear just because the work is handed to machines; it simply shifts from "the person who made the mistake" to "the people who set the rules and sign off." That's why the division-of-labor mechanism in Manage is stated so clearly: AI executes within authorization, humans retain key judgments and are accountable for the results, and everything is documented throughout. If any one of these three is missing, scaling up becomes risky.
Many teams focus on "can it run automatically," but what actually blocks scaling is another matter: when something goes off track, can you explain within three minutes whose rule it was, who approved it, and at which step it started to go wrong. If you can't answer that, you'll only dare to keep the agent in the pilot.
Row
Digital employees: perceive, compute, execute, and act within authorized scope
know
People: set standards, make judgments, take responsibility, and sign off at key milestones
III. A ledger alone is not enough; there also has to be an exam paper
The ledger solves "who is doing it" but not "whether it is done right". Another move I am willing to record this week is 《Lingee Bench》: China's first Agent evaluation benchmark for enterprise management scenarios, jointly released by Kingdee with large model vendors and universities. It has four evaluation sets—general office, finance, supply chain, and manufacturing—with more than 1,400 questions. It tests not question answering, but whether agents in real business environments can finish the work, do it right, and leave verifiable results. There is also a white paper alongside it; Agent vendors and model vendors can also use it to evaluate their own models and Harness.
Why is this more useful than a spec sheet? Because when enterprises procure AI, what they have to compare has always been very vague: demos all go smoothly, case studies all say "efficiency improvement," and the metrics differ from vendor to vendor. Last week there was a fairly typical scene—a vendor issued a press release saying that of 100,000 confirmation forms a year, 90% required no human intervention; that same day people pressed them on cost, who handles the remaining portion, how many people were originally needed, and when it went live. Not one of the four denominators was provided. This isn't nitpicking; it's that the client side has finally learned to look at denominators.
What the evaluation benchmark fills in is precisely this gap: replacing 'feels pretty good to use' with a set of alignable questions. Of course, benchmarks led by vendors themselves have questionable independence and cannot be used as third-party audits. But at least it points in a direction—in the future, accepting AI should not just rely on vendors reciting their own cases; there needs to be a set of questions that both sides can rerun.
A quick reminder: Benchmark is a general baseline and cannot replace your own acceptance testing. The same set of questions may score 95 in someone else's repository, but when applied to your setup with historical dirty data and custom approval flows, it might not even reach 60. Benchmarks are for screening vendors; your own baseline is for deciding whether it can go live—don't mix the two up.
IV. Context Is Not Decoration; It Is the "Instruction Manual"
For more than a hundred agents to each do their jobs without overstepping, prompts alone cannot hold it together; they need to know how this enterprise operates. The numbers given at the conference were quite specific: the first batch of 70 ERP connectors and 1,127 Tools; the enterprise context covers more than 30 types of data sources, over 30,000 business objects, more than 800,000 semantic fields, and 170,000 business rules. The permission system, approval rules, relationships between business objects, and historical handling practices are written as preconditions for AI actions, rather than letting the model guess from scratch.
Zhang Eyu's travel expense reimbursement example thoroughly explains the difference between 'connecting in' and 'getting it done': after an employee initiates a reimbursement, the agent does not just generate the document; it also has to go back to ERP to verify whether this reimbursement form actually exists, whether it exceeds the department budget, whether finance has completed the review, and whether the payment has already been deposited into the employee's account. Only after all four steps are verified does it truly enter the business process.
Embedded in this is a rewording of ERP's positioning: traditional ERP is a system of record, recording what has happened; AI-native organizations need a system of action, pushing things forward. But the prerequisite for being able to push them forward is understanding the definitions, rules, and boundaries.
| Past: Functional Perspective | Now: Executing Entity Perspective |
|---|---|
| Deliver a rule, a button | Deliver an account that triggers on its own and makes judgments |
| If an error occurs, ask IT to check the logs | If something goes wrong, it must be possible to trace it to the business owner and approver. |
| Acceptance checks "whether the functionality is available" | Acceptance checks “whether the work is finished, done correctly, and leaves verifiable results” |
| The cost is hidden in the implementation quote | Costs are incurred monthly based on the number of executions and points consumed. |
For teams using Odoo, this lesson is actually a bit more convoluted. The advantage of the open-source route is that the data is in your own database, and permissions and processes are fully visible and modifiable; the cost is that there is no vendor to maintain that ledger for you—whether 180 agents or 3, who owns the execution entity, which models it can read, which documents it can write, how much it costs per month, and under what conditions it is disabled all require you to build your own table and designate the owner yourself. The community edition never had built-in AI agents to begin with, and those entry points connected through custom modules or external automation platforms are actually more likely to be missed in registration. I have seen this more than once: when reviewing half a year after the project went live, no one could clearly say who configured a certain scheduled task in the middle of the night, or whether it is still running.
V. Three Clumsy Methods for Digitalization Leaders
First build the roster, then talk about scaling the numbers. Give each agent a record: business owner, readable data scope, writable document types, monthly cost cap, deactivation conditions, and reviewer. If there is no responsible person, don't grant write permission. This costs nothing, but it can head off 80% of the wrangling later.
Bring the exam into your own acceptance process. Pick a high-frequency process (reconciliation, invoice matching, or expense reimbursement all work), use historical data from your own database to define a set of baseline questions, and require the agent's output to include verifiable results; treat vendors' general benchmarks as screening tools and your own baselines as go-live gates. Also record who discovered the time it ran wrong.
"Human accountability" must be written into the rules, not just on the architecture diagram.Of the three layers, the one most easily skipped is the middle layer. Who approves turning it on, who reviews the output, who has the authority to stop it with one click, and whether an upgrade that changes the instructions requires re-acceptance—set the rules first, then enable the feature. If the rules are not set, the number can only remain at the scale of three to five pilots.
Go back to that field at the beginning. On the dashboard of 180 agents, the most valuable thing is not the ROI column, but the "responsible department and responsible person." Because ROI is the result, responsibility is the premise. In the enterprise, there is now an extra team that does not draw a salary, does not get off work, and does not need onboarding. Do not rush to calculate how much money it saves—first make clear who manages it, whom to go to if it does something wrong, and on what basis it is judged to have done it right.
Before these three things are made clear, the faster the scale goes up, the uglier the accounts look.
This article is compiled and analyzed based on public information: Xinhua Net’s “How Is AI Reshaping Financial Management? Kingdee Explores New Paths for Enterprise Management” (2026-11-07), reports by China Business Herald Network and Phoenix Finance on Lingee Bench and the IDC China Future Enterprise Awards, Eastmoney’s “Kingdee Lingji Fully Upgraded, Making Every Enterprise an AI-Native Organization” (2026-11-04), and China Economic News Network’s “From Tool Stacking to Multi-Agent Collaboration” (2026 Global Visionary Conference). Data as of 2026-11-08, for reference by enterprise digitalization leaders, and does not constitute procurement advice.
Zhongdi Information · Specializing in Odoo and digital solutions for enterprise management
