--- title: "Who Hit Fast-Forward on the Evaluator's Half Day" date: 2026-11-04 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-digital-worker-19-1104 language: en --- # Who Hit Fast-Forward on the Evaluator's Half Day > In a meeting room, bid documents are stacked half as tall as a person on the table, and five experts sit through a whole afternoon turning pages, checking and scoring. The work is not mysterious, but In a meeting room, bid documents are stacked half as tall as a person on the table, and five experts sit through a whole afternoon turning pages, checking and scoring. The work is not mysterious, but it is draining: dense clauses, many tables, and the need to watch for contradictions between one part and another. Someone gave the same job to a machine to try, and the time it took made the experts pause. This is not science fiction; it is a scene from public-resources trading in Hefei, Anhui. Behind it stands a group of agents called digital workers: they do not chat or write poems, but pick up the most process-driven, most repetitive stretch of work in a company. AI has taken another step from answering questions to getting things done for people. This kind of application also appeared in Wuzhen. According to China Science Daily, at the Light of the Internet Expo of the World Internet Conference, iFlytek put the Xingfan WinWork digital worker platform, a new generation of office computers and smart procurement and bidding on the stand, with organizational efficiency as the main theme. It draws less attention than a robot, yet sits closer to real offices. ## A Room of Bid Documents and an Afternoon Procurement and bidding look simple, but the chain is long. How requirements are parsed, how bid documents are drafted, which criteria are used to score at evaluation: every step has rules, and every step has to leave a trace. According to iFlytek, the smart procurement assistant breaks this chain into more than twenty application modules so that machines can take them on separately. The difficulty is that it crosses boundaries. One bid document touches procurement, legal and finance departments, and connects to several older systems; an assistant that only answers single questions often cannot get a hand in. Bidding is also a heavily regulated setting with detailed compliance requirements, where skipping any step can cause problems, so whether it can enter the process matters more than whether it answers correctly. Policy is pushing the same way. According to Anhui Daily, eight government departments including the National Development and Reform Commission have issued implementation opinions on accelerating the use of AI in the bidding field, proposing that by the end of 2027 more key scenarios be rolled out in more places. Machines entering this field are no longer just a company's own experiment. ## A Quarter Hour at the Evaluation Table The most visible change happened in Hefei. According to Anhui Daily and Hefei's public resources trading regulator, in March 2025, during the evaluation of an elevator procurement project, work that used to take people three and a half hours was turned into an evaluation opinion by AI in 15 minutes, and five experts unanimously adopted it after checking. According to People's Daily Online and other reports, the public-resources trading large model built together with Hefei raised assisted-evaluation efficiency fourteen-fold, and the practice was selected as a typical demonstration case for AI Plus in the bidding field by the National Development and Reform Commission. One caution: this was the first real-world test on a specific project, and should not be taken directly as the average for every project. iFlytek gave a more general account during the World Artificial Intelligence Conference, saying the average evaluation time per project fell from three and a half hours to 15 minutes. The two statements describe the same body of practice, one about a single evaluation and one about an overall caliber; keeping them distinct stops a case being misremembered as the norm. ## Breaking the Process into Many Steps On the corporate side, the accounts have to be kept more carefully. According to China Economic News Network, since 2022 iFlytek and China Energy Investment Corporation have worked closely on intelligent unmanned evaluation, sorting out the rules for drafting, submission, evaluation and review link by link and feeding them to the system. It has run stably on a large number of projects, with an intelligent evaluation accuracy of 97 percent. People's Daily Online gave the same 97 percent figure, and noted that the project was highlighted by the State-owned Assets Supervision and Administration Commission. With two sources agreeing, the number stands. As for how much money is saved each year, public accounts differ; rather than recording a figure that does not line up for readers, it is better to look first at which steps it has actually taken over. According to company business statistics relayed by People's Daily Online, smart procurement and bidding has handled more than 500,000 projects in total, with average evaluation efficiency about 60 percent higher. Figures like these come from the company itself; what matters is the scale and the trend: not one polished demo, but people being freed up little by little across thousands upon thousands of repeated processes. ## Digital Workers Get a Job Description At the Wuzhen stand this approach was given a more vivid name. According to iFlytek, the Xingfan WinWork digital worker platform is built around having a post, understanding the business and being able to take up the job: first it builds an enterprise ontology linking business rules, data and expert knowledge, then it connects to existing systems for contracts, finance and procurement. Xingchen Cowork, shown earlier at the World Artificial Intelligence Conference in Shanghai, was described by CCID and Shangyou News as letting digital workers pull data across systems and produce reports automatically. The two names belong to displays at different events, and the shared point is clear: the assistant no longer stays in a chat box, but is assigned to a specific post. Many companies feel the same: giving employees a Q&A assistant is easy, making an entire business genuinely faster is hard. The obstacle is often not the model, but that one task has to cross several departments and several systems. That is the joining problem digital workers are trying to solve, rather than building one more window that is better at chatting. But digital worker is a position the platform claims, and should not be read as large-scale replacement of human staff. The word more common in public materials is collaboration: machines handle the massive, repetitive, objective comparisons, while people handle complex logic and value judgments. Redrawing a post, rather than removing the post altogether, is what is actually happening now. ## The Final Call Still Belongs to Humans In this design, what matters most is the space left empty. According to iFlytek, once digital workers enter a business process, every step is auditable and able to improve, and humans still make the final call at key points. Contract review, procurement evaluation and financial data processing have fixed, highly repetitive processes, so machines are suited to a first pass; once a trade-off is involved, the wheel goes back to human hands. This is especially true in heavily regulated settings. Bidding is about fairness; however high the efficiency, if the process cannot be explained and responsibility cannot be traced to a person, it will not go far. Being traceable and explainable sounds less striking than a claim of so many times the efficiency, yet it is the threshold that decides whether systems like this are truly used. ## From Answering Questions to Acting for You Look back over this year and the change in AI is not that it talks better, but that it is more willing to act. In the past we asked it a question and it gave an answer; now we hand it a bid document, a set of rules and a deadline, and it walks the repetitive stretch in the middle by itself, putting the result on the table for someone to confirm. For the experts sitting in the meeting room, the most tangible gain may not be replacement, but being able to wrap up by three in the afternoon. The half day saved can go to the hard problems machines cannot chew through. Technology pulling people a little way out of the process makes human judgment more valuable, not less. A quarter hour at the evaluation table measures more than speed. It measures that AI is finally willing to get into the unglamorous, unexciting links where people spend their days. Answering questions was only the starting point; being willing to carry the task through for you is what really counts as taking up the post. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-digital-worker-19-1104 Free to quote with attribution and a link to the original.