How Shanghai Jahwa Turned WorkBuddy Into a Cross-Department AI Operating Model

Shanghai Jahwa’s WorkBuddy rollout is notable less for the headline efficiency figure than for the way the company scaled it. Instead of deploying an AI assistant as a generic chatbot, teams reportedly converted recurring, rules-based work into tested workflows and reusable “Skills.” That allowed knowledge held by individual employees to become a shared operating resource across the organization. 4

The results are substantial, but they should be read carefully: the efficiency figures come from reported company case studies rather than an independent productivity audit.

From a two-team pilot to nine departments

WorkBuddy reportedly started with Shanghai Jahwa’s IT and HR teams in April 2026 before expanding into routine production across nine departments. The rollout covered functions including supply chain, brand marketing, R&D, finance, HR, and IT, with the company reporting average employee-efficiency gains of more than 4.5×. 4

The scaling pattern had four parts:

1. Find repeatable work. Each function identified processes involving recurring data collection, reconciliation, formatting, analysis, or reporting.
2. Translate business rules into workflows. Employees broke down how the work was performed and used WorkBuddy to create scripts or structured Skills.
3. Validate against real historical work. Teams checked outputs, corrected errors, and added exceptions discovered during testing.
4. Reuse the result. Once a workflow was reliable, it could be run repeatedly and shared with other employees or departments.

This approach made adoption a process-design exercise, not simply a software installation. Employees who understood the business remained responsible for judgment and quality control while the agent handled much of the repetitive execution.

How employees used the agent

The reported use cases extended well beyond drafting text. In the most recent month cited, 67% of users applied WorkBuddy to specialized business work, 63% to data processing and analysis, and 57% to automation and operations or maintenance. These figures are reported usage shares, not measures of how much time each activity saved. 7

Examples across Shanghai Jahwa included:

• R&D: researching cosmetic ingredients and applicable regulations.
• Finance: turning reimbursement-review rules into a repeatable process and generating investment-screening reports.
• HR: creating probation-expiry reminders for managers and employees.
• Supply chain: automatically refreshing shop-floor dashboards across multiple workshops.
• Brand marketing: producing consumer insights, analyzing content, and generating marketing materials.

The marketing team also reportedly converted about 70% of its weekly presentation-style reviews into HTML pages. The pages could be shared across the company, updated online, and viewed without waiting for a new static deck to be distributed.

The supply-chain workflow that made the model concrete

The clearest example came from logistics billing. Shanghai Jahwa’s B2C business used Tmall, JD, and Douyin, with warehouse, handling, delivery, and other charges spread across seven raw billing categories. Over 18 months, the data set reached roughly 60 million rows. The existing Excel-based monthly process took five to six person-days, and large files could become difficult to open or process.

Zhang Ying, a supply-chain operations manager with an economics background and no coding experience, reportedly approached the problem by decomposing the monthly process into smaller steps. WorkBuddy generated Python scripts for those steps, which Zhang repeatedly tested against the 18 months of historical data. When testing exposed exceptions or new business rules, she added them to reusable Skills.

The reported changes were:

• Monthly billing analysis: two person-days to 15 minutes.
• Butterfly-chart base table: one person-day to two minutes.
• Visual dashboard: 1.5 person-days to three minutes.
• Complete monthly workflow: five or six person-days to under 20 minutes.
• Highest reported gain for an individual task: up to 150×.

The result was not just a faster spreadsheet. The finished dashboard could be filtered by dimensions such as brand, platform, expense category, and average order value, while managers could open a link to view refreshed results instead of waiting for a new report to be rebuilt.

Why the rollout scaled

The case suggests that enterprise-agent adoption is more likely to spread when the first use cases have three characteristics:

• They happen frequently enough to justify automation.
• Their rules can be explained and tested.
• Their output can be checked against existing records or business expectations.

Logistics reconciliation fit that pattern particularly well. The process was repetitive and data-heavy, but it still required an employee who understood the billing structure, recognized exceptions, and could decide whether the result made business sense. WorkBuddy accelerated the mechanical work without eliminating that domain oversight.

The same principle applied to finance, HR, R&D, and marketing. Each team supplied the context and judgment; the agent helped package the repeatable parts into workflows that could run again and, in some cases, be shared more broadly. 4

Productivity multiplier, not automatic job replacement

Tencent’s argument from the deployment is therefore narrower—and more practical—than a claim that AI replaces office workers. The reported gains came from domain experts specifying rules, validating results, correcting exceptions, and turning successful processes into reusable organizational assets. 4

That distinction matters. A chatbot can provide a one-off answer, but a tested Skill can preserve how a company performs a recurring task. The value compounds when several departments create and reuse workflows rather than keeping every automation as an isolated experiment.

At the same time, the headline numbers should not be generalized automatically to every company or role. The available evidence describes Shanghai Jahwa’s reported results in selected workflows, and it does not establish an independently verified 4.5× increase in productivity across all employees. The more defensible lesson is the operating model: start with measurable, repeatable work; keep experts in the loop; validate against real data; and share what works.