Once the policy is changed, how long until the system catches up?

On November 10, 2026, the website of the Hunan Provincial Department of Industry and Information Technology republished an article on AI implementation in manufacturing. It covered three scenarios—production scheduling, expense review, and material sourcing—all from the "AI + Manufacturing for Growing Enterprises" forum at the Kingdee Global Visionaries Conference. The most valuable sentence was buried in the middle.

Du Yangyang, Chief Operating Officer of Lvneng Huichong, said that after the company adjusted its systems, the relevant rules can be generated in about 30 minutes. The same paragraph also contains two figures: the settlement cycle was shortened by about one quarter, and manual verification work hours were reduced by 50%.

On the same day, EY released the Finance & Accounting Meta Hub AI platform at CIIE: 15 agents, a three-tier architecture, one orchestration hub, and four foundational layers. On one side is the finance office next to the workshop; on the other is the Big Four launch event. They answer the same question—the value of finance and management AI lies not in how much faster document review becomes, but in how quickly the system can keep up after the rules change.

1. An Overlooked Metric: The Delivery Cycle for Rule Effectiveness

In traditional ERP, policies are hard-coded: reimbursement standards, approval levels, account mappings, tax rates and voucher templates, and budget tightness all reside in parameters, approval flows, validation rules, and custom development. Once a policy changes, it must go through requirements, assessment, development, testing, and deployment; minor changes take weeks, cross-module changes take months, and around year-end closing you still have to queue.

The cost of this delay is that there are always two sets of rules: one written in the documents, and one running in the systems, with human labor catching whatever falls in between.

Lvneng Huichong changed direction. They did not upgrade the system, nor add an analytics dashboard—according to Du Yangyang, those tools still require people to make a large number of judgments. They chose “not to build another system, but to add a colleague who can do the work to the finance department”: the reimbursement method remains unchanged; the agent checks policies, identifies risks, and triages in the background, while matters requiring judgment are handed to people.

Two situationsHow to implement institutional changestimescale
Traditional ERP hardcodingGo through requirements—assessment—development—testing—launch; changes involve coordinated updates across multiple areasSmall changes are measured in weeks; cross-module changes in months.
Rule Generation (Lvneng Huichong basis)After the system adjustment, the agent generates a draft of the rules, which takes effect only after manual review.Approximately 30 minutes; the stated basis is rule generation, not online activation.

What it changes is not efficiency, but the distance from management intent to system behavior. The essence of management is translating rules into executable actions; ERP is that machine, and what AI changes is the speed of translation. When that speed is compressed from an iteration cycle to a single shift, rules can for the first time synchronize with the rhythm of the business.

But speed isn't everything. What is generated in 30 minutes is a draft, not a policy that has taken effect; in between, three things are still missing: approval, comparison samples, and rollback points. Treating "generation" as "taking effect" is the most common and hardest-to-detect failure in this kind of project.

II. The Big Four Turned Methodology into a Platform: How 15 Agents Are Layered

EY's Finance and Accounting MetaHub provided a more complete structural reference: split the accounting function into three layers, then connect them through a hub.

LevelCoverage scenariosWho will take responsibility if something goes wrong?
Accounting Operations LayerExpense review, bank-enterprise reconciliation, smart vouchers, month-end closing, accounts receivable management and collection, accounts payable management and payment, intelligent annual report inspectionVouchers can be reversed, reconciliations can be rerun, error correction costs are low, and higher autonomy can be granted
Data Analysis LayerIntelligent finance and accounting insights and management analysis report generation, supporting data queries, metric analysis, and anomaly identificationConclusion deviations require manual review, and the autonomy level should be lower than that of the operational layer
Business Insight LayerRevenue forecasting, budget management, cash flow analysisAffects budget and funding arrangements; humans must make the decisions, and machines perform the simulations.

The three-layer distinction is not about technical difficulty, but about “who is responsible when something goes wrong, and how expensive it is to correct it”: an error in one voucher at the operational layer can be reversed and redone; an error in a judgment at the analytical layer requires human review; an error at the insight layer is an error in budget assumptions and affects decisions. The autonomy level of the three layers on the same platform should differ—it has nothing to do with how smart the model is, but with whether the error is reversible.

The orchestration hub is four things: a unified entry point, MCP routing, cross-Agent task orchestration, and permission-based task scheduling. In management language: people can find it, requests go to the right place, multiple agents relay to finish one task, and every call falls within someone's authorization scope.

The most underrated of the four foundations is the "financial accounting semantic layer." It determines whether the agent speaks the language of accounting: which account an expense goes into, which standard constrains a voucher, and what each metric is under the management basis and the statutory basis. Without it, no matter how strong the model is, it is only guessing; with it, rules can be generated stably and reproduced in audits.

Zhao Jianyun, EY Greater China Technology Risk and Data Intelligence Services Leader, said that value lies not only in the efficiency of individual tasks, but more in combining data, professional knowledge, and scenarios; Director Yang Chen emphasized moving from single-point functions to process collaboration. The two statements are one thing: The selling point is not the number of agents, but in breaking accounting work into verifiable layers.

III. The common premise of the three samples: data must flow first, and only then can rules be introduced

Shenzhen Kinco Electric makes motion control products for the robotics industry. Production scheduling originally had three windows: the first 7 days were in principle fixed, 7 to 14 days allowed limited adjustments, and 14 to 28 days were relatively open. "Small-batch, rapid-response" orders broke through this—once an order entered production, the customer was still raising new requirements. CIO Gu Jianglei's exact words were, "Innovation capability has already overwhelmed organizational capability and supply chain capability." They used Kingdee Lingji to feed in data and align definitions, enabling AI to become familiar with capacity and process constraints and to run simulations when orders changed; material chasing that was originally constrained by fixed cycles can now be adjusted the same day.

Anxinyi E-commerce (Shenzhen) operates in the electronic component supply chain. Global Sales VP Sharon Ho recalls that during the supply tightness from 2021 to 2022, lead times for some materials stretched from 4 weeks to 52 weeks. In a list of dozens of part numbers, there could be 10 discontinued and 5 out of stock; procurement experts had to check inventory and verify historical inquiries and transactions, and even with more than ten years of experience it took 1 to 2 working days. They broke down sourcing, supplier evaluation, price judgment, and order tracking, and integrated knowledge from systems and experience into agents, cutting the time to about 10 minutes. Sharon Ho's summary is "AI is responsible for speed, experts for accuracy, and data for reliability."

About 10 minutes, compared with 1–2 working days. The processing time for Anxinyi's material sourcing and procurement analysis; previously, procurement experts with over ten years of experience needed 1–2 working days to complete equivalent work. This figure is based on disclosure from a forum and does not specify the number of listed models or the distribution of task complexity; it is intended to indicate order of magnitude rather than precise efficiency.

The commonality among the three companies is not that they used the same product, but that their sequence was consistent: first find the most painful problem, then take stock of existing data and processes, and only finally discuss models. The report also points out the premise: all three had digitized relatively early. Zhang Zhibing, General Manager of Kingdee EBC's Overall and Industry Product Department, summarized it as software evolving from a "system of record" to a "system of action."

The "system of record" answers what happened; the "system of action" has to answer what to do next. The former can be delivered as long as its data is accurate; the latter must first turn policies, permissions, and definitions into something a machine can execute—and if this step is done poorly, the more proactive the agent, the faster it goes wrong.

IV. Rules run fast—can accountability keep up?

Turning policies into objects that can be managed with version control is the greatest value. But precisely because it is fast, three things must be designed in from the start.

The first is versioning. A rule needs a policy document number, effective date, approver, and invalidation conditions. Without these, policies will drift within the system, and at audit time no one can say clearly who changed what. Worse still, rules keep running after the policy has been repealed—such "orphan rules" raise no error; they just quietly execute a policy that no longer exists.

The second is replay. An audit trail is not the same as replayability: an audit trail records "what the agent judged at the time," whereas replay means being able to re-run historical documents using the rules in force at that time. In a dispute, the former gives only a conclusion; the latter gives the basis. The two are separate implementations, and acceptance testing must test them as two separate items.

The third is separation of duties. Who can change rules, who can approve them, who can release a given type of document—if all three permissions land on the same person, AI merely upgrades manual fraud into configuration fraud, and makes it harder to detect, because the changes hide in the rules repository rather than on the documents.

The fastest way to fail is to treat "rule generation" as "rule activation." Generation produces a draft; activation must pass three gates: an approval signature, a comparison run against last month's documents, and a version checkpoint that can be rolled back. Skip these three, and the system will process real business with a rule that no one ever approved.

Three actions for the CIO:

First, acceptance metrics must come in pairs. Rule generation time must be viewed in the same table as the false-positive rate, false-negative rate, and replay pass rate; reporting only the first is like watching only the accelerator and not the brake. These four items must be written into the acceptance terms.

Second, write policy versions into the system. Each rule corresponds to a policy document number and effective date; revisions produce a new version, and the old version is marked as invalidated rather than deleted—keeping historical versions is the only way to answer, six months later, the question "why was it judged this way back then?"

Third, run a pre-launch comparison using real documents. Take one full previous month of documents and run the old and new rules side by side, explain every difference item by item, and do not go live with any difference you cannot explain; in the first month after launch, sample and review weekly, and feed disputed samples back as inputs for revision. This set of actions is nothing new, yet it determines whether "30 minutes" is a capability or a risk.

A counterintuitive judgment: the moat is not the model, but the semantic layer and the existing data. Lvneng Huichong can add a "colleague" without rebuilding its system because its account structure, approvals, and voucher structures are clear; Anxinyi can reach a conclusion in 10 minutes because historical inquiries and transactions are searchable. Those with a solid data foundation enter production first; those with a thin one will only see 30 minutes amplify problems faster.

Back to that number from the beginning. What 30 minutes means is not that rules can be changed faster, but that for the first time, policies can be updated at the pace of the business. That is good for management and pressure for governance—the development cycles of the past few months were essentially a buffer period that gave people time to think through whether the policies were right.

With the buffer gone, judgment has to move earlier. The CIO's question thus shifts from "can AI do financial review?" to the sharper one: once the policy changes, how quickly can we prove the system is still right? If you can answer that, 30 minutes is a capability; if you cannot, it merely increases the speed at which risk occurs.


Sources: Hunan Provincial Department of Industry and Information Technology, "Focusing on Real Business Scenarios, Manufacturing Enterprises Accelerate AI Adoption" (published 2026-11-10 09:18, including remarks and data from Step Electric CIO Gu Jianglei, Lvneng Huichong COO Du Yangyang, and Anxinyi Global Sales VP Sharon Ho at the "Growth Enterprise AI + Manufacturing" parallel forum of the 2026 Kingdee Global Visionaries Conference; views from Zhang Zhibing, General Manager of Kingdee China's EBC Overall and Industry Product Department; and the background of the January 2026 "AI + Manufacturing" Special Action Implementation Opinions issued by MIIT and seven other departments); Guandian Net/Tencent News, "Ernst & Young Releases Accounting Meta-Node AI Platform at CIIE, Including 15 AI Agents" (2026-11-10 23:54, including 15 Agents, a three-tier capability architecture, four orchestration hub capabilities, and four foundational capabilities, with views from Zhao Jianyun, Head of Technology Risk and Data Intelligence Services for Ernst & Young Greater China, and Director Yang Chen).

Notes on data basis: 30 minutes, a roughly one-quarter reduction in the closing cycle, and a 50% reduction in manual reconciliation hours are all disclosures made by Lvneng Huichong at the forum, without stating the statistical base, sample period, or calculation method; Anxinyi's 1–2 working days and about 10 minutes are the interviewee's own figures, without disclosing sample size or the distribution of task complexity; the extension from 4 weeks to 52 weeks is the interviewee's recollection of the 2021–2022 supply crunch, not a contemporaneous statistic; Step Electric's three-stage scheduling window is a description of its existing system, not an industry-wide standard. The above figures are intended to gauge magnitude and direction and cannot replace an enterprise's own baseline measurement. This account covered the separation of duties and go-live timing of Odoo 20 accounting agents on 2026-11-02, and the three gates for agents taking over finance and testing on 2026-10-26; this piece focuses on the delivery cycle and governance loop "from policy to rule" and does not repeat their boundary and scenario breakdowns.

Zodioo · Enterprise Digitalization Observatory · 2026-11-11

关于我们

​我们致力于帮助中小企业实现数字化转型,我们的团队由一群充满激情和创新思维的专业人士组成,他们具备丰富的行业经验和技术专长。

扫一扫获取顾问以及手册

Sign in to leave a comment
The real test of permission migration: Is denial still denying?