Enable enterprise AI to transition from "being able to answer" to "trustworthy execution", Qianzhi Intelligence releases the business Agents hub.
Taking datasets as the core and business flow Harness as the trusted base, we enable general intelligent tools to form evaluable, reusable and evolvable connections with the existing business systems of enterprises.
The core product of Beijing Qianzhi Intelligent Technology Co., Ltd. (hereinafter referred to as "Qianzhi Intelligence"), the "Trusted Business Agents Hub", targets the key problems emerging after enterprise AI evolves from dialogue assistance to business execution: how to enable Agents to truly understand business rules, call existing systems, and operate stably, efficiently and with low Token consumption under the premise of verifiable results and traceable processes.
The threshold for enterprise Agents is shifting from model capabilities to business trustworthiness
General intelligent agents are already capable of reading documents, generating code and calling tools, but after being put into real business scenarios, there is still a clear gap between the model being "able to perform" and enterprises being "willing to adopt". A procurement compliance review may involve contract texts, supplier master data, budget accounts, approval records and invoice images at the same time; a financial audit not only needs to identify anomalies in statements, but also understand the interlocking relationships between statements, the caliber of accounting systems and regulatory rules.
The difficulties of such tasks can hardly be solved only by expanding model parameters or adding more prompts. The business logic of enterprises is often precipitated in database associations, existing system rules and expert experience, and scattered across different systems including ERP, finance, OA and procurement. Even if a general Agent completes a task through multiple rounds of dialogue, it is difficult to reproduce the result stably, and cannot guarantee that the execution effect is almost consistent for multiple people and multiple runs.
The technical path of Qianzhi Intelligence is not to replace the existing business systems of enterprises, nor to let the large model recalculate all businesses from scratch, but to add a trusted business hub between general intelligent tools and business systems, so as to solidify the business processing process explored and planned by AI into business flows. Through business connection, ontology automation, Agents solidification and RSI evolution, the platform transforms one-time intelligent exploration into reusable, evaluable and continuously improvable business processes. It separates "rule formulation" from "rule application": business exploration, process planning and formulation of verification standards are completed in one go during process version iteration; daily business processing is only responsible for execution without repeating these actions. Verification is still required during operation, but the verification dimensions and rules can be configured according to scenarios, without sticking to a fixed set. Once enterprise-level Agents face large-scale data, business flow solidification becomes a rigid requirement: without the support of precipitated processes, each call requires on-site reasoning by the model, and the cost will continue to leak in the form of Tokens, directly eroding business profits.
Taking datasets as the core to enable Agents to understand business relationships
The core of this hub is the self-developed "Business Flow Harness" of Qianzhi Intelligence. Through MCP connectors and skill packages, the platform connects mainstream programming and office intelligent tools, as well as enterprises' existing ERP, DRP and data middle platforms, so that the upper-layer model can call underlying business capabilities while retaining the rules and process assets in the enterprise's existing systems.
Different from Agent platforms that take variables or files as the main carriers, the business flow Harness takes datasets as the operating core. Its ProtoDataSet carries structured data and unstructured materials between process nodes, and the Proto prototype supports business semantic query and traceability at any time (metadata defined by ontology objects, attributes, rules, or other methods). As a result, operations such as joint query, perspective, grouping, filtering, and interlocking check can directly become nodes in the business process, without being disassembled into a large number of one-time variable assignments or temporary scripts.
On top of the dataset core, Qianzhi Intelligence uniformly encapsulates data processing, large model processing, machine learning, as well as API and MCP calls into visual and debuggable processing nodes. After business personnel complete a review or analysis with the help of intelligent tools, they can solidify the effective path into a standard business flow Agent. Deterministic tasks in the process are undertaken by data processing and rule nodes, while large models focus on planning, understanding, extraction and generation, reducing the instability caused by repeated exploration.
The ontology development engine further complements business semantics. The seamless connection between the business flow Harness and the ontology engine serves as the optimal base for automated ontology execution. Qianzhi uses formulas close to natural language (syntax compatible with financial and central enterprise final accounts statements) to express business rules, supports automatic association between tables, data acquisition optimization and mapping from rules to database instructions or interface APIs, and can use relational database queries to undertake part of the graph query requirements, reducing the implementation cost for enterprises to build a business semantic layer.
Turning "trustworthiness" into an executable version release mechanism
For low fault-tolerant scenarios such as finance, procurement, supervision and audit, a successful demonstration does not mean production availability. Therefore, the business flow Harness takes automated trusted evaluation as an independent layer: before the Agent is released, it calls historical data or simulated data for operation evaluation, locates samples with inconsistent results, and then completes iteration by adjusting processes, optimizing prompts and supplementing the sample library, and releases it after meeting the requirements.
After going online, the platform can also precipitate error cases into negative samples, and promote subsequent improvement through difference attribution, sample expansion and regression testing. Qianzhi's multi-agent learning platform has intelligently screened and annotated millions of corpora for customers, and all its capabilities are connected to the business flow RSI (Recursive Self-Improvement) evolution engine in the form of tools and skills. The improvement object here is not the parameters of the basic model, but the processes, prompts, rules and sample assets of the business Agent; every business correction has the opportunity to be transformed into the test basis for the next version release, or the reference sample for the next inference.
This mechanism transforms Agent reliability from relying on personal debugging experience into a testable, traceable and regressionable engineering process. For enterprises, trustworthiness not only means more accurate answers, but also means verification before release, records during operation, and positioning and correction paths after problems occur.
Process reuse responds to accuracy, speed and Token cost at the same time
Another value of business flow solidification is to reduce repeated planning and exploration of Agents in the operation phase. Verified data associations, rule calculations and system calls can be directly reused, and the model does not need to understand the process from scratch, generate scripts and correct errors repeatedly every time. The process thus not only carries business knowledge, but also becomes a technical means to improve operational efficiency.
In the past three years, Qianzhi has completed the review of millions of reports (covering fields including internal control, financial statements, bidding and tendering) and 200 million supporting documents. All the intelligent agents used have an accuracy rate of over 90% when released, and half of them reach 95% (based on the corpus set of intelligent annotation + business expert review). Comparative tests show that after the business flow is solidified, the running speed of the intelligent agent is increased by 10 times, and the Token consumption is reduced to one tenth of the original (the indicators come from the enterprise's own business practice, and the specific performance is related to the business scenario, data base and model configuration).
Accuracy mainly comes from automated evaluation, RSI iteration, trusted release and negative sample learning; speed and Token cost improvement mainly comes from pre-positioning of AI exploration, planning and verification, which are transferred to the release phase (a large number of local open source models are used); development efficiency benefits from ontology semantics, visual orchestration and automated verification. Its goal is not to replace the large model with processes, but to let the model capabilities play a role in a deterministic business framework, solidify AI capabilities with processes, instead of letting AI with high uncertainty directly handle business.
From one-time delivery to accumulable organizational intelligent assets
The trusted Business Agents Hub is applicable to scenarios with large data volume, where structured data and document materials are mixed. In these scenarios, the large model is responsible for material understanding, element extraction and induction, and the business flow is responsible for cross-system data acquisition, rule execution, process trace retention and exception transfer, and the two jointly complete the closed loop from "understanding content" to "finishing work".
More importantly, the solidified processes, corpora, samples and skills can continue to be precipitated as enterprise assets. The experience of business experts no longer only exists in personal minds or one-time conversations, but can be reused, shared and iterated through processes and evaluation sets. For large organizations that need long-term construction of AI capabilities, this assetization capability is closer to the real needs of production systems than single task demonstrations.
Qianzhi Intelligence stated that it will continue to invest around the business flow Harness, ontology automation and trusted evaluation system in the future, and promote the replication of business Agents in more scenarios with high rule density. With the continuous evolution of general models and intelligent tools, what Qianzhi Intelligence hopes to occupy is not another model entrance, but a connection, understanding and execution bridge between the AI brain and business systems, a trusted AI middleware.
Founded in 2022, Qianzhi Intelligence focuses on the R&D of enterprise-level trusted AI agent platforms, supervision and review, and penetrating regulatory intelligent systems. It has previously completed angel round financing and is advancing Pre-A round financing.
The "intelligent system" mentioned here refers to a system built based on the same set of ontology semantics, which provides APIs, digital employees and MCP interfaces at the same time, and can independently select interaction modes and presentation forms. Three types of instructions: traditional business operations, autonomous AI and external AI, share the same set of execution logic and verification rules, taking into account user experience, convenience, efficiency and R&D costs.