【塞掐】AI 時代「大者恆大」：中小企業還剩多少生存空間？/ Big gets bigger: SME room in the AI era
B · Useful, limited A Taiwanese podcast interview with the CEO of agent vendor Smart Solution (EgentHub) giving a concrete method for turning a master craftsman's know-how into agents; all figures are his own sales claims and there are no client outcome numbers.
Summary based on the full transcript·Original: about 69 minutes to listen
In short
A 69-minute Mandarin episode (uploaded 2 November 2026) of Side Chat on the Taiwanese tech site INSIDE, with Albert, co-founder and CEO of Smart Solution (智慧方案), maker of the EgentHub agent platform. He earlier founded the company behind 優仕網 and 愛逛街, which LINE bought in 2017. His firm turns the know-how of experienced workers in Taiwanese textiles, footwear and precision machining into agents. The method: watch the master do one order, take over, have the master correct the output, and turn the corrections into prompts, now into Skills. He argues large manufacturers adopt faster than small ones and that the gap is widening.
Key numbers
- 200+companies that have adopted the vendor's platform since it opened on 1 July 2024, per the host's introduction
- 10full-time people at the vendor
- 10 to 50 pageslength of an English textile tech pack that experienced staff translate word by word, taking hours to one or two days per order
- ~70quality score (out of 100) of the first AI attempt at the master's task before the master's corrections were turned into prompts
As reported by the source.
Chinese·Video·opens youtube.com
What it reports
- The company opened on 1 July 2024 and reached product-market fit by October 2024. Its first three months in martech and e-commerce failed: those firms have 10 to 30 people at computers and would not buy an enterprise platform. Manufacturers with 500 to 5,000 people converted, and one first contact can bring 100 people to a session.
- Sales start by asking which AI the client uses and what they pay. Most use free tiers, so the team demonstrates on paid 100 or 200 US dollar plans with non-secret order sheets, QC sheets, purchase orders, BOMs and drawings. He calls the firm "like a training company", running over a hundred workshops a year.
- Textile case: orders arrive as English PDF tech packs of 10 to 50 pages. Staff with two or three years' experience translate them word by word in Google Translate into Chinese, Vietnamese, Khmer or Thai; one order takes hours to one or two days. In tests Claude 3.5 Sonnet read the tech packs most accurately, and 3.5 did better than 3.7.
- Method: interview the master, watch half an hour of a three-hour job, take over, reach about 70 points, ask the master how to fix the gaps, and write that into prompts. Today the step is "please generate Skills", followed by reviewing whether the SOP was understood.
- Spread depends on culture: in "warm" organisations the first trained person teaches others; in "cold" ones staff keep it to themselves or fear replacement.
- Pricing is per seat (for example 20 to 30 seats growing to 100); token costs pass to the client with no margin. He calls the role close to a forward-deployed engineer: consulting, sales, training and deployment together.
- Deployment is an on-premise "AI middle platform" on client servers, with cloud or local models, access rights that mirror the org chart ("manage AI the way you manage people") and connectors to ERP and MES. The aim is to stop staff pasting client drawings into personal ChatGPT accounts.
- Resistance: AI makes people busier at first, and higher output without higher pay is a brake. Without company rollout, shadow AI appears ("a six-hour job done in one, handed in slowly"). The host cites a firm whose AI-usage KPI led teams to bounce inflated documents back and forth.
- Big gets bigger: large Taiwanese firms move fast while small ones take three to six months to evaluate. When one of 30 people doing the same task succeeds with AI, the other 29 follow. The host notes that local open models are "99%" Chinese, a trust problem for Taiwanese factories.
- His own setup: several 200 US dollar accounts, Macs that never sleep and a dashboard of "nearly 300" agents run through a chief-of-staff agent; the bottleneck is his own decisions.
How they got it
A hosted podcast interview; the guest describes his own company and clients, named only by sector. No client outcome figures (time saved per order, error rates, adoption) are given. The digest is based on cleaned Traditional Chinese uploader captions.
Limits
- Vendor CEO describing his own sales; no outside verification and no client names.
- The 200+ deployments and 10-person team come from the host's introduction and are not broken down by sector or size.
- The tech-pack case reports the old process, not measured results after the change.
- The episode closes with his personal stock picks and an invitation to learn about his service, so part of it is promotion.
- The library summary calls the 200+ "manufacturing deployments"; the source says 200+ companies without a sector split.
Why it matters for companies
The capture method is cheap and repeatable: observe one real task, let the model attempt it, collect the expert's corrections and keep them as reusable instructions under company control. The governance pattern (on-premise platform, access rights copied from the org chart, low-sensitivity work first) answers the shadow-AI problem most firms have. The warning about usage KPIs is worth taking: a hard adoption target can raise token use while output falls. His claim that scale now favours large firms is an argument, not evidence, but it matches the 30-to-1 spread logic he describes.
Key passages in translation
Chinese original
他就說雖然有可能70分，還有一些就是說它這裡沒弄好，那裡沒弄好我就說，那你跟我講一下怎麼教，你教我 我教AI
He said it was maybe 70 points, with some parts not done right here and there, so I said: tell me how to teach it. You teach me, I teach the AI.
Chinese original
怎麼管人就怎麼管AI
Manage AI the way you manage people.
Chinese original
三十個人在做一樣的事情，只要有一個人導AI成功，另外那二十九個人一起進步啊
Thirty people do the same job; as soon as one of them succeeds with AI, the other twenty-nine improve with them.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com