Baiwang: The Biggest Mistake in Enterprise AI Is Redoing All Processes with AI

From customer service, marketing, and R&D to finance, HR, and procurement, nearly every traditional software category is looking for its own AI version: forms that people used to fill in are now generated by AI, materials that people used to read are now recognized by AI, buttons that people used to click are now invoked by Agents, and those enterprise processes that have already been transformed by digitization and automation many times are also beginning to be repackaged as "AI processes."

This change is of course valuable. Take review as an example: expense reimbursement, supplier onboarding, contract checks, and verification of HR materials are all high-frequency, repetitive tasks with relatively clear rules that cannot easily afford errors. Using OCR to read information, large models to understand semantics, and rule engines to make judgments can indeed greatly reduce manual operations.

The answer is probably no.

Because what agents truly bring is not "in the past ten people clicked the mouse ten times, and now AI clicks for them," but rather forcing enterprises to rethink a more fundamental question: if we were to design a job from scratch today, with large models, Agents, data, and automated execution capabilities already available, would we still design the process we have today?

The former is layering AI on top of old processes; the latter is what AI-Native truly means—not making AI adapt to the original way of working, but redesigning the work itself according to AI's ability to understand, judge, schedule, and execute tasks.

Many traditional enterprise processes are complex not because the business naturally requires such complexity, but because under the software capabilities of the past, enterprises had to break a complete job down into a large number of standard actions that systems could handle.

Take finance and tax work as an example. From the occurrence of a transaction to the final completion of processing, it often has to go through data collection, invoice organization, system entry, manual verification, voucher generation, report consolidation, declaration confirmation, payment, and archiving. In the era of traditional software, every step required an interface, an operator, and clear inputs and outputs, so a complete complex workflow gradually became fixed.

If so-called AI transformation only replaces manual entry with OCR, manual verification with large models, and manual clicks with RPA, enterprises will of course gain efficiency improvements, but what they are running is still the process formed in the past—only part of the work has been shifted from people to machines.

The true AI-Native logic should be the reverse: enterprises first propose business goals—"handle this month's taxes properly," "judge whether this transaction has risks," "tell me what potential problems there are this year"—and as for how many invoices need to be read, how many systems need to be called, how many verifications need to be performed, and how many intermediate tables need to be generated to accomplish the goal, the system should complete as much of that as possible on its own.

In the past we asked, "How many steps have been automated?" In the future, the real question should be, "How many steps still need to exist?"

From "operating software" to "delivering results," Baiwang bets on a large model for finance and tax

From this perspective, looking again at the fact that Baiwang has already delivered more than 500,000 digital employees, what is truly noteworthy is not just the number 500,000 itself, but that the product logic behind it has begun to shift from "intelligent features" to "reconstructing the way work is done."

According to the design of its AI finance and tax system, business owners or finance personnel no longer need to first determine which software module to enter and which menu to click. Instead, they can directly raise business-level requests, and the AI CFO understands the intent, breaks down the tasks, then schedules specialized agents for bookkeeping, declaration, risk, and analysis to execute collaboratively, and finally delivers the results and the matters requiring enterprise confirmation back to the user.

Take Shuixiaozhi as an example. It does not solve just one link among bookkeeping, tax filing, or invoice processing, but attempts to take over the complete chain from data collection, intelligent bookkeeping, and report generation to tax declaration, payment, and archiving. After enterprise authorization, the system can collect business data such as input and output invoices, bank statements, and receipts, and uniformly archive the resulting vouchers, reports, declarations, and payment results, forming a traceable finance and tax evidence chain.

Behind this actually corresponds to an increasingly important divide between AI-Native products and traditional SaaS: traditional software sells features, while AI-Native products ultimately deliver results.

For enterprise managers, what truly matters is no longer "how much faster bookkeeping has become," but that a job that used to require finance personnel to constantly operate and connect across multiple systems and multiple nodes is beginning to be packaged into an intelligent service that can continuously complete tasks and deliver results.

This change is especially evident in enterprise risk management.

Traditional review is essentially an after-the-fact mechanism: the business has already occurred, the vouchers have already been formed, the materials have already been submitted, and then the enterprise looks for problems through one or more layers of review. Therefore, when AI enters the enterprise, the most direct approach is naturally to automate these reviews.

Baiwang's exploration in finance and tax risk detection has already shown this shift. The system no longer merely checks a single invoice or a single report, but puts cross-year financial, tax, invoice, business registration, judicial, and other data together, observes changes in the enterprise's tax burden ratio and operating indicators through the time dimension, and combines industry data to find anomalies, forming a panoramic risk profile of the enterprise and further moving risk discovery forward to the self-check and rectification stage.

As a result, the enterprise's risk process begins to shift from "business occurs—materials submitted—review initiated—problems found" to "business occurs—data continuously enters the system—agents continuously understand the operating state—when anomalies appear, proactively alert or even trigger handling."

After 500,000 digital employees, the key to agent competition is no longer "quantity"

When enterprises talk about Agents today, it is easy to develop a new kind of quantity anxiety: others have built ten Agents, so should I build twenty; finance has an Agent, so should procurement, HR, and sales each also be equipped with a batch?

But the number of Agents itself can hardly constitute true enterprise competitiveness.

An agent that can truly enter an enterprise's core business requires not only model capabilities, but also business facts, industry rules, historical data, permission systems, and system interfaces that can actually execute tasks. Precisely because of this, Baiwang's long-accumulated enterprise transaction and fiscal and tax scenarios have instead gained new value in the agent era. Its services cover approximately 30 million enterprise users, and after 500,000 digital employees entered real scenarios, they can also form data and feedback loops through continuous use, driving further iteration of knowledge, rules, and agent capabilities.

In this sense, what the 500,000 truly validates is not "how many Agents Baiwang owns," but whether agents can move beyond the Demo, enter real business processes, and ultimately become part of how enterprises work.

Therefore, to judge the value of AI in the next stage, enterprises perhaps should establish three new standards: not looking at which positions can be replaced by AI, but at which business results can be directly handed over to AI; not looking at how many steps have been automated, but at whether AI truly reduces process nodes, waiting time, and manual handoffs; and not understanding Agents as a new operating entry point for traditional software, but redefining the division of labor among people, AI, and professional services.

Observing Baiwang again according to this standard, what is truly worth discussing is no longer that it "redid fiscal and tax with AI," but that it is gradually transforming the way of working previously built around software modules, manual operations, and professional personnel into a new service structure that starts from business tasks, uses an AI CFO to dispatch professional agents, takes real business data as the factual basis, and relies on professional personnel as a fallback for complex matters.

This is perhaps the direction that truly deserves expectation for enterprise agents.

In the AI era, of course, a large amount of work will be redone, but where technological revolutions truly create value has never been in copying the old world once more efficiently, but in making the processes, positions, and organizational methods once taken for granted in the old world be re-examined.

Enterprises that can answer this question may have obtained the next ticket to the AI 2.0 era.