Wisely Chen：永聯物流開發如何使用 AI Agent 來加速營運流程自動化

1. In one sentence

A VP with a strong technical background who calls himself "a business process novice" uses AI agents as a macro recorder that can work around small problems by itself, aimed squarely at the old problems large enterprises still have — printed label paper, manually downloaded spreadsheets, reconciliation by hand.

2. The session

The host introduced him as having over fifteen years of technical experience at Google, Apple, Yahoo and GE before moving into traditional logistics and commercial real estate development (this came from the host, and no public material could be found to corroborate it — marked as the speaker's own account, no external source). He is currently VP of Data and AI at Ally Logistic Property (ALP). Verified background on ALP: it is Taiwan's largest institutional logistics real estate developer, currently managing over six hundred thousand square metres of warehousing in Taiwan, driving over a billion US dollars of smart warehousing investment, and expanding into Southeast Asia (Malaysia, Thailand, Singapore, Vietnam).

Last year's progress: from not being able to hire to AI-enabled part-timers

He first recapped what he shared last year about his own startup: traditional transport cannot hire, so use AI to enable the people who are willing to come and work on site, gradually turning them into half data engineers. Looking back a year on, he describes it as a social experiment: of the ten to twenty people he worked with, three converted to full-time roles, one of whom was interviewed by a business publication — and after that interview ran, a number of traditional enterprises (steel logistics operators among them) got in touch asking him to come and explain internally how to accelerate digitalisation with AI.

This year's change: from small companies to the Fortune 500, and the same problems

As the team merged into the Ally Logistic Property group, the business grew five to ten times and the clients shifted from traditional small companies with no digital capability to plenty of well-known listed and multinational brands. He found a contradiction: these international clients do have IT capability, but logistics integration keeps landing last in project priority with insufficient budget and time, so they end up in the same "no system integration" bind. His example: an international household goods manufacturer sends them picking information every day on printed label paper (the speaker's own case, no external source). His conclusion is that whether the client is large or small, it still usually comes down to people wiring things together by hand.

Case one: real-time reconciliation between two WMS systems, replacing manual work with Claude Computer Use

A client was moving goods from warehouse A to warehouse B, the two belonging to different customers on different warehouse management systems (WMS). The schedule was tight, IT estimated three months to build two-way data integration, and the business side simply could not wait. The interim workaround was downloading spreadsheets from both systems by hand and merging them manually, which could only be done once a day — while slot status actually changes hourly, producing frequent conflicts where "system A shows the slot as free and system B already has it occupied".

His solution was Claude's Computer Use: hand the whole sequence of logging in, clicking through in order, downloading the spreadsheet and merging it — all of it previously repeated by hand — to the AI. The whole thing took about one to two hours to build, after which reconciliation could run hourly or even half-hourly. He deliberately left two mistakes in the demo: a field name written wrongly in the prompt (he wrote "outbound order number" where the screen actually says "outbound system order number"), and an instruction left imprecise (never saying where the exported file should go). The AI agent worked both out and corrected them without being thrown. During the demo an instant messaging window unexpectedly popped up and covered the screen, and the agent simply dismissed it and carried on — he made a point of noting that a traditional RPA tool in the same situation would typically freeze and throw an exception.

Case two: report plausibility checks, replacing a part-timer with the Browser-use framework

The second case: before the data team emails reports to management each day, they run a "plausibility analysis" (a given warehouse's revenue should not exceed two million or fall below five hundred thousand, say). This repetitive, dull task used to go to a part-timer checking by hand, but people get tired and there is no guarantee it happens every day. He switched to the Browser-use framework with Claude, having the agent use his own account to open Gmail, search the last seven days for the specific report email, open the PDF attachment, check the numbers fall inside the plausible range, and finally log into Teams under his account to report the result — exactly the same sequence the part-timer used to perform, and management never noticed it was a bot. He explained that Browser-use differs from a purely screen-driving tool: as well as feeding screenshots to the model, it also parses the page's HTML structure, which makes its judgements more likely to succeed.

Why AI agents rather than traditional RPA

He summarised the difference crisply: RPA aims for high determinism and has low tolerance for error and low modifiability, so a change in the process or the screen tends to break the whole thing; an AI agent carries a degree of non-determinism, but that very non-determinism gives it a human-like tolerance for error, letting it handle small problems itself. He thinks this suits two settings especially well: daily repeated, clearly-ruled data downloading and merging, and routine checks like plausibility analysis that need constant watching and where human attention slips.

3. Figures and cases

• ALP manages over six hundred thousand square metres of logistics facilities in Taiwan, is driving over a billion US dollars of smart warehousing investment, and is expanding into Southeast Asia (Malaysia, Thailand, Singapore, Vietnam) (figures from the speaker's slides, with public reporting also corroborating — see sources).
• About ten to twenty part-timers were involved last year, three of whom converted to full-time (speaker's own account, no external source).
• The business grew five to ten times after merging into the group (speaker's own account, no external source).
• Case one's automation took about one to two hours to build, after which reconciliation went from once a day to hourly or half-hourly (speaker's own account, no external source).

4. Lines worth keeping

• "It's not really AI agents; it's a macro recorder accelerating operational automation."
• "They do it this way too — incredible." (his reaction on finding that large enterprise clients also hand over goods via printed label paper)
• "The only thing I have to worry about is whether you've gone down today. That's it." (on why running daily routine checks with an AI agent is more reassuring than giving them to a person)

5. Tools and terms mentioned

Claude Computer Use (letting a model drive the screen, mouse and keyboard directly), Browser-use (an open-source browser automation framework that parses page HTML as well as taking screenshots), WMS (warehouse management system), OMS, TMS, RPA (robotic process automation), ETL, Teams, Excel.

6. Wider observations

What stands out most is that the speaker never dodges how homespun this is — he says outright that it is a macro recorder, while pinpointing exactly what separates a macro recorder from an AI agent: tolerance for error. Traditional RPA is stable until the screen changes and then the whole set breaks; an AI agent, precisely because it is "not quite sure", turns out to be tougher. His use of Claude Computer Use and Browser-use for two different settings (driving a screen versus needing to understand page structure) also echoes 李慕約's later point that "AI agents will end up better at using tools than people" — the difference being that Wisely Chen has already landed this inside daily operations rather than discussing it in the abstract.

7. Sources