When AI Product Managers Forget Business: Lessons from a Live‑Ecommerce Case

01 When I Use the Wrong Tool, My Effort Becomes Embarrassment

Last month a friend who runs a live‑ecommerce shop, drunk and frustrated, shouted at me that "AI" was the most annoying word for his team. I realized I had been chasing the latest large‑model releases—Gemini 2.0, Claude 3.7 Opus, Sora video generation, Midjourney V7—while ignoring the real pain points of the business: exhausting short‑video production, clueless content planning, manual live‑stream recap, and no insight into why some videos succeed.

I had been immersing myself in AI groups, debating MoE architectures, RLHF, and multimodal fusion, believing that mastering technology equated to being a good AI product manager. When my friend asked, "Can your big models help me sell more girls' dresses?" I had no answer.

The lesson was stark: the worst AI mindset is to study AI itself without asking what the business actually needs.

02 A Holiday in the Live‑Stream Room Taught Me Real Product Management

During the Chinese New Year I stopped chasing AI news and spent a few days on my friend’s shop, observing how he selected products, filmed videos, streamed, and performed post‑mortems. I documented his SOP (Standard Operating Procedure):

Content creation – how many videos per day, how many outfits, optimal lighting.

Content distribution – which platforms, posting times, headline wording.

Live‑stream flow – start time, interaction minutes, order‑closing prompts.

Data review – which metrics to watch: GMV, conversion rate, UV, and how to record them.

Seeing the SOP made me realize I was not there to sell AI, but to find where AI could actually help.

03 Why AI Deployments Fail: We’re Too Self‑Absorbed

Many AI product peers fall into the same trap. Examples:

Six months on an intelligent客服 (customer‑service) bot, yet frontline agents report three out of ten answers are wrong, creating extra work.

Huge investment in a data‑platform that looks like a sci‑fi dashboard, but only the data team uses it.

An AI writing tool promises "one‑click viral content" but users spend two hours editing, longer than writing from scratch.

The root cause is never asking the business what they truly want. For a客服 system, do they need fewer complaints or higher revenue? For a BI tool, do they need flashy charts or direct action recommendations?

04 AI That Business Understands

After spending days in the live‑stream room, my technical metrics (accuracy, recall, F1) were replaced by business metrics: GMV, conversion rate, return rate, repurchase rate. I asked whether AI could help calculate profit.

Example: a customer complains that a product is "slow" and wants a new model. A traditional AI客服 replies with a generic apology. By adding a "listening ear"—keyword extraction plus order‑history lookup—I could suggest an upgrade with a subsidy. Result: 3.5% of complaining users clicked the link and bought a higher‑priced product, generating revenue ten times the annual licence fee of the客服 system.

The business leader then asked, "Can we equip the sales team with the same assistant?"

05 Turning Expert Tools into Fool‑Proof Assistants

Data analysts often receive a BI dashboard they never use. Instead, I embedded the analysis into a chat bot (DingTalk/WeChat Work). Users ask, "Which channel sold best last week?" The bot replies in 30 seconds with a concise action recommendation and the exact script keywords to reuse. Monthly active users of the data platform quadrupled.

06 Stop "Free Creation" and Build an Assembly Line

AI content generators that aim for "viral" pieces often produce bland text that requires more editing than original writing. The mistake is treating AI as an artist. The right approach is to let AI handle the repetitive pipeline steps and let humans focus on review and selection.

For a mother‑and‑baby community, the workflow becomes:

Procurement – AI crawls ten authoritative parenting guides daily.

Chopping – Split each guide into 20 independent knowledge‑card snippets.

Cooking – Re‑phrase each snippet into three tones: expert, mom‑friendly, and engagement‑driven.

Plating – Auto‑match cards to relevant community groups.

Human effort is reduced to reviewing, fine‑tuning, and publishing, boosting efficiency while keeping quality stable.

07 The Real Battlefield for AI Product Managers Is Not Academic Papers

My earlier mindset chased model updates and technical jargon, believing that made me professional. The new mindset asks the business: "What process? What pain point?" This shift turns AI from a self‑indulgent showcase into a profit‑driving engine. As Serai AI founder Wang Yu said, "AI’s true value is to make professionals more professional—designers stop doing rote work, product managers stop endless meetings, and can focus on strategy." So, AI product managers should ask themselves: Are you studying AI or studying the business? Studying AI makes you "knowledge‑rich but useless"; studying the business makes you "solution‑oriented". When you stop selling AI and start helping people work less and earn more, business units will chase you for the next version.