Enterprise AI: Engineering is closer to value than models.
```
On June 11th, the TaxFriend Agent new product launch was held. Many products, technologies, and market potentials were discussed at the event, but what truly caught my interest were two specific figures.
One is 93.56%, the comprehensive accuracy rate of financial and tax AI staff disclosed by the company. The other is 1 to 460, meaning that a 1 yuan token cost corresponds to 460 yuan of business output.
The former touches the bottom line of enterprises adopting AI, and the latter points to the commercial value of AI applications. Together, they embody the most noteworthy changes in industrial AI: large model capabilities continue to improve, but application competition has already penetrated deep into business processes, engineering systems, and outcome delivery.
From a Single Agent to a Group of Agents
TaxFriend’s newly released Agentic 2.0 features a core upgrade called A2A collaboration. Previously, the AI invoice clerk and AI accountant mainly handled single-point tasks; now multiple Agents work together around invoicing, accounting, tax, compliance, and business analysis.
This change is particularly representative in financial and tax scenarios.
Invoicing and accounting can be broken down into relatively standard tasks, while compliance and business analysis require understanding the company’s industry, transaction structure, expense invoices, and policy changes. A single Agent can complete actions, but it's difficult for one to independently handle full results. After multi-Agent collaboration, the AI consultant provides industry and client information, the AI accountant executes accounting, the compliance Agent conducts risk checks, and the task scheduling Agent connects the workflow.
Thus, AI applications enter organizations. The benchmark shifts from whether an answer can be generated to whether a traceable, reviewable business outcome can be continuously delivered.
TaxFriend disclosed that its AI staff have served over 4,000 organizations and 467,500 SMEs, issuing 1.92 million invoices. This data indicates that vertical Agents now have real business use cases.
Enterprise AI: Engineering Is Closer to Value Than Models
A recurring term at the conference was "Harness."
It can be understood as a comprehensive system beyond the model, including knowledge supply, context management, tool invocation, task orchestration, link tracking, performance evaluation, and security governance. Financial and tax business has clear rules and high cost of errors. Whether the model’s answer resembles human responses is unimportant; the key is where data is drawn, what rules are followed, which systems are invoked, and whether errors are traceable.
This is where industrial AI diverges from generic chat products.
TaxFriend’s management noted that large volumes of foundational financial and tax data processing remain done by traditional computation, with large models mainly used for rule generation, task scheduling, and frontend interaction. This approach is relatively pragmatic. Deterministic computations are handled by mature systems, while semantic understanding and complex collaboration are handled by models. This manages token costs and reduces model hallucinations in accounting results.
Of course, the 93.56% comprehensive accuracy rate still has room before meeting the reliability required for financial and tax business.
Multi-agent collaboration increases workflow complexity; an error in one link may propagate. Therefore, manual review, responsibility allocation, and audit mechanisms remain key parts of product deployment.
Progress in enterprise AI should not only be measured by automation rate, but also by who detects errors, who bears them, and whether they can be corrected in time.
AI Is Changing SaaS Pricing Logic
Traditional financial SaaS mainly charges based on account and duration, with price constrained by software competition and client budgets. With Agents entering the delivery process, vendors are beginning to price based on saved manpower and added business value.
TaxFriend divides its pricing into three tiers. Basic accounting focuses on efficiency improvement, compliance accounting handles risk and tax compliance, and business accounting further provides business analysis. The closer to the business outcome, the higher the customer price. The company also plans to reach end customers directly through business-owner facing memberships, bypassing accounting agencies, with basic membership priced at around 980 yuan annually.
Whether this model will succeed depends on whether end customers are willing to pay continuously.
The company disclosed that in some pilot regions, membership conversion rates rose from about 7% in 2023 to nearly 30%, with renewal rates in Shanghai exceeding 70%. These figures come from specific stores and regions; sample selection, sales methods, and service investment all affect the results. Follow-up should observe renewal rates, customer acquisition costs, and gross margin as the scale grows.
The Hardest Thing to Replicate Is Still the Organization
Another detail easily overlooked at the event:
TaxFriend tests products through joint-operating cloud stores, and management admitted that after rapid store expansion, store manager capability became a constraint.
This shows that AI deployment rarely ends with software installation. Processes must be reconstructed, roles redistributed, and client communication and billing adjusted. TaxFriend owns financial and tax data, industry knowledge, existing clients, and an in-house testing ground; these assets can shorten the validation cycle. Whether they can consolidate into cross-region, cross-store standard capabilities remains to be seen.
From this perspective, TaxFriend’s launch event provided a sample to observe industrial AI. Models are becoming infrastructure, but real application value depends on whether enterprises can connect data, knowledge, tools, processes, and business channels into one chain.
An invoice may be small, but it’s specific enough. AI only truly enters industry when it embeds in such concrete processes, standing up to scrutiny in accuracy, cost, and responsibility.
This article is from WeChat official account "Hard AI". For more AI frontier news, please follow here
Risk Warning and DisclaimerThe market carries risks, and investments require caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions herein fit their particular situation. Any investment based on this is at your own risk. ```