Huawei Releases Enterprise AI Whitepaper, Proposing 'Business Defines Technology' and DIMAK Framework

Over the past year, AI has been moving from a tool that answers questions into real enterprise business, covering areas such as office assistants, code generation, marketing content production, and supply chain analysis. But a recurring phenomenon is that while a certain role is noticeably accelerated, the entire business process does not necessarily benefit in sync.

An e-commerce AI product photography case reported by QbitAI is quite representative. Manager Peng, who provides AI solutions for e-commerce merchants, said that his team has integrated capabilities such as AI image generation, model swapping, and background swapping into the product listing process, and that the product owner can complete operations such as material processing and product listing at about 20 times the original speed. At present, the system has been adopted by 8 stores, with each store paying about 20,000 yuan per month.

But Manager Peng also emphasized, "it is still best to have a human take a look"—whether generated materials meet requirements still needs manual review, and this capability is mainly concentrated in the product catalog production stage, while its connection with subsequent processes such as approval, product selection, and operations has not yet been fully established. In other words, AI saved time for one role, but how that benefit is transmitted to the entire business process remains an unresolved question.

Facing this gap, Huawei's first-layer proposition is to "let business define technology." In Huawei's view, the first question in enterprise AI construction is not "which model to choose," but what business problem the enterprise truly needs to solve.

Many AI projects in the past started from technical capabilities and first looked for applicable scenarios; but for enterprises, valuable applications need to start from business goals and find the key links that affect business outcomes. The report gives an example: an AI system can help sales organize customer information faster, but if that information does not enter sales decisions, customer follow-up, and order processes, the value ultimately generated is still limited.

Huawei Chief AI Application Expert Li Hongkai told QbitAI that local efficiency gains may also cause bottlenecks to shift: "Your link became more efficient, causing the next link to become the blockage." Therefore, enterprises need to start from the complete business process, find the key links that truly constrain overall output, and then judge where AI should be used. The whitepaper suggests that business, finance, and technology leaders must first jointly confirm the current business baseline, the business metrics they hope to improve, why AI is needed, and who is responsible for the final result.

Finding a valuable scenario is only the first step. After a scenario runs successfully, enterprises also need to consider: how the data, knowledge, and methods formed in solving this business problem can be preserved, reused, and continuously improved, and then expanded to more businesses.

DIMAK, proposed by Huawei, is an engineering system for the construction, supply, and continuous improvement of enterprise intelligent capabilities. It consists of five mutually coordinated engineering domains: Data, AI Infrastructure (Infra), Model, Agent, and Knowledge. Around a specific task, data engineering provides trustworthy business facts, and knowledge engineering organizes policies, processes, and expert experience into retrievable rules and cases with clear applicable conditions; model engineering is responsible for model access, adaptation, and evaluation; agents are responsible for organizing data, knowledge, models, tools, and human-AI collaboration; and AI infrastructure ensures the stable operation of the entire capability set.

The experience left after a task is completed is also continuously used: what evidence the task used, what humans changed, where anomalies occurred, and what the final result was can all be recorded; after verification, effective experience enters knowledge services, failed records are used for subsequent evaluation, and mature methods are accumulated into reusable task methods. The whitepaper summarizes the core of DIMAK as "decoupling the technology lifecycle from the enterprise capability lifecycle"—models and frameworks can be continuously updated, while verified data services, knowledge assets, and task methods become capabilities that enterprises can accumulate over the long term.

The power grid equipment evaluation practice in the whitepaper gives this capability accumulation a concrete form: a certain power grid enterprise organized professional guidelines originally intended for human reading into knowledge assets with clear sources, valid versions, and clear scope of application; the operation and inspection agent combines equipment ledgers, defect and monitoring data, generates evaluation suggestions based on the guidelines, and professional personnel review and confirm them. The evaluation basis and results are saved together, and manual revisions, defect handling results, and equipment status changes continuously feed back as the basis for the next round of evaluation.

For tasks that need to further change order, account, or production status, a formal operation must be completed after a judgment is formed, which involves the connection between intelligent capabilities and the enterprise's existing business systems. If subsequent steps still require personnel to query, enter, and submit across systems one by one, efficiency gains in the analysis stage may not shorten the complete business process.

Huawei's proposed H-shaped dual towers are precisely for connecting these two parts: the enterprise AI system (intelligent tower) uses DIMAK as its engineering foundation and is responsible for understanding goals, analyzing changes, comparing options, and organizing tasks; the enterprise IT/OT system (existing network tower) is the business and production system already running in the enterprise, responsible for providing authoritative business facts, verifying and executing operations, and recording formal results; the horizontal connection in the middle of the "H" connects data, knowledge, permissions, instructions, and results, allowing the same task to move from judgment to execution and then return the results to the intelligent system.

The bank wealth management maturity case in the whitepaper demonstrates this collaboration: a large commercial bank started from customers' arrangements for maturing funds. After obtaining necessary authorization, the mobile banking agent combines product maturity status, account funds, risk level, and existing holdings to form multiple options and explains differences in term, return, liquidity, and risk; the bank's original systems provide authoritative facts such as accounts and holdings. Only after the customer determines an option and completes suitability confirmation and necessary strong authentication can the agent invoke business capabilities such as fund transfers and wealth management purchases within the authorized scope, while transaction verification, account changes, and holdings records are still completed by the bank's original systems. The whitepaper discloses that this practice has entered authorized execution after customer confirmation within a limited scope.

A task does not end once an instruction is issued. According to the architectural requirements of the H-shaped dual towers, after a purchase request is accepted, the system still needs to confirm based on the formal records of the original business system whether the transaction is completed and whether accounts and holdings have been updated; if the result is unclear, the status should first be verified to avoid duplicate operations; if execution fails or business conditions change, the task needs to be readjusted or handed over to personnel. These real execution results then return to the DIMAK system and become the basis for the next round of capability improvement.

In addition to the above cases, the whitepaper also records implementation practices in multiple fields such as finance, automotive, pharmaceuticals, materials chemistry, and electric power. Different industries have different business goals, business processes, data conditions, and risk boundaries. These cases can provide references, but how to build specifically still requires enterprises to make judgments based on their own business.

Huawei AI System Scientist Yao Jun said at the launch event: "The last mile of the business value stream is built by the enterprise; we cannot replace it." This also determines Huawei's role: sorting out common problems that repeatedly appear in different projects to form reference methods, engineering frameworks, and common tools, reducing the cost of repeated exploration by enterprises and ecosystem partners.

Huawei's proposed implementation path emphasizes starting from core scenarios, gradually completing verification and deployment into production, and then expanding along the value business stream; after entering a specific industry, enterprises, Huawei, and ecosystem partners need to advance together—enterprises clarify business goals, business rules, and responsibility boundaries; partners participate in the adaptation and implementation of specific scenarios; and Huawei provides reference designs and engineering foundations that can be used jointly. Returning to the choice of the e-commerce merchant at the beginning, enterprises can purchase already mature capabilities and place their own construction investment in business rules, data connection, and process collaboration. How to allocate this specifically still needs to be determined by business scale, existing conditions, and expected returns.