What 118 Minutes With an E-Commerce Team Revealed About AI Adoption Gaps

On March 11, I took the high-speed rail from Hong Kong to Shenzhen's Futian district -- an 18-minute ride across the border -- to spend two hours with an e-commerce team I used to work with. The company sells smart health devices globally, with 40,000+ monthly orders on their flagship product and nearly 380,000 Amazon reviews across their catalog. They operate in the US, Japan, Germany, and other European markets. By any measure, this is a company that understands scale.

What I found when I walked into that conference room was a team that had adopted AI further than most companies I train -- and was still leaving enormous value on the table.

The Setup

This was not a cold engagement. I used to work at this company. The CEO invited me back to do a share session with the team, an informal knowledge exchange covering AI applications across their e-commerce operations. Six departments showed up: Design, Operations, TikTok content, Sales, HR, and ERP Product Management. No slides for the first hour. Just conversation.

I wanted to understand where they actually were before recommending anything. So I asked each department a simple question: what are you using AI for right now, and where does it break?

The 70-80% Number That Hid a Deeper Problem

The design team gave me a number that would impress most organizations: 70-80% of their visual marketing materials are now AI-generated. They are using ComfyUI-based platforms with various image generation models. On paper, this is advanced adoption.

Then I asked about their hit rate. How much of what the AI generates is actually usable without manual editing?

About 50%.

That means for every two images they generate, one goes straight to Photoshop for rework. Products deform. Textures -- particularly complex ones like patterned surfaces -- lose accuracy. Video generation is worse: product proportions shift between frames, angles become inconsistent.

The root cause was not the AI tools themselves. It was three compounding gaps that I see in almost every team that has moved past the initial excitement phase.

Gap 1: No Prompt Library

When I asked the design team if they had a standardized prompt library, the answer was immediate: no.

Every designer was writing prompts from scratch, every time. Different people used different terminology. There was no shared vocabulary for styles, no documented templates for common product shot types, no way to replicate what worked yesterday.

I introduced them to the SOBAR framework -- Style, Object, Background, Enhancement, Retouch -- which I have used with other product-focused companies including Chow Tai Fook's design team. The idea is simple: break every image prompt into five consistent components, document the best values for each, and store them in a shared spreadsheet. When a new team member joins or when anyone needs to generate a product lifestyle shot for the Japanese market, they pull from the library instead of starting from zero.

Companies that build these libraries typically double their usable output rate. That 50% hit rate could become something closer to 80-90%, not because the AI got better, but because the inputs got more consistent.

I also showed them how to reverse-engineer competitor imagery. Take the top-ranked Amazon listing in your category, feed the hero image into Gemini, and ask it to deconstruct the image using the SOBAR framework. What style is it? What background? What lighting? You extract the formula, add it to your library, and adapt it for your own products. Legal, practical, and immediately useful.

Gap 2: Wrong Model, Wrong Tier

The operations team told me they use DeepSeek and Gemini for data analysis and market research. When I dug deeper, I found they were using default settings -- no model selection, no distinction between fast and thinking modes.

This is like owning a car with six gears and only ever driving in second.

The difference between a base model and a reasoning model on the same platform is not incremental. For data analysis tasks, switching from DeepSeek's default to its Thinking mode can be the difference between a hallucinated market summary and a genuinely useful competitive analysis. The team did not know these tiers existed.

I demonstrated a complete research pipeline: start with Copilot for initial web research, move to ChatGPT's reasoning mode for synthesis, then use a research agent for deeper analysis. Same question, dramatically different output quality -- because each step used the right model for the right task.

The design team had a parallel version of this problem. They were generating images with older model versions when significantly better options were available on the same platforms they already paid for. Upgrading from their current image model to the Pro tier, and from their video model v1.5 to v3, would have cost them nothing extra and improved output quality substantially.

Gap 3: No Data Security Framework

When I opened the floor to questions, the first hand that went up was about data security. "What happens to our product data when we upload it to these AI tools? Can competitors access it?"

This is the question that stops organizations in their tracks. I have written about the Traffic Light Protocol before -- it is the single most effective framework I have found for unblocking AI adoption in security-conscious environments. I shared the same model with this team:

Green: Public data, general knowledge queries. Use freely with any AI tool. Yellow: Internal company data that is not customer-sensitive. Use with cloud AI tools, but strip identifiers and be selective about what you upload. Red: Customer PII, proprietary formulas, unreleased product designs, legal-risk data. Do not enter into external AI tools. Period.

The team was also evaluating private deployment options -- running open-source models on local hardware. My advice: master the cloud tools first. Private deployment is a significant infrastructure investment. If you cannot get value from ChatGPT and Claude with proper data hygiene, running your own models will not solve the problem. It will just add server maintenance to the list.

The Live Demo That Changed the Conversation

About an hour in, I shifted from discussion to demonstration. I opened Claude Code -- an AI agent that can orchestrate multiple tools autonomously -- and gave it a single instruction: research this company's CES 2026 product lineup and produce a 15-second promotional video with voiceover.

The agent searched the web for product announcements, selected relevant images, wrote a script, generated voiceover audio, and assembled everything into a finished video. The room got quiet. Then it got very loud.

This is the gap between where most companies are -- stuck at Stage 1 of AI maturity, using individual tools for individual tasks -- and where the technology is heading. AI agents do not replace human judgment. They compress the mechanical work so dramatically that the ratio of strategic thinking to execution shifts in your favor.

For an e-commerce team producing content across six markets, the implications are significant. A content pipeline that currently takes a week of coordination across designers, copywriters, and translators could be compressed to hours of human review and refinement, with the agent handling the first-pass assembly.

The 10 Areas I Prepared (And What Mattered Most)

Before the visit, I had researched 10 specific AI application areas for e-commerce: Amazon listing optimization, review mining, ad copy generation, multilingual content, visual content, customer service automation, demand forecasting, competitive intelligence, social media marketing, and internal operations.

I came with stats. AI-optimized Amazon listings convert 15-30% higher. AI sentiment analysis hits 89.7% accuracy versus 74.3% for rule-based systems. AI demand forecasting reduces stockouts by up to 65%.

But the numbers that resonated most with this team were not the aspirational ones. They were the immediate ones: listing creation dropping from 4 hours to 30 minutes. The ability to process 380,000 product reviews in hours instead of never. Content localization costs falling 50-80% per market.

For a company selling the same products in the US, Japan, and Europe, multilingual content is arguably the highest-ROI AI investment. Each market needs not just translation but localization -- different tone, different SEO keywords, different regulatory language, different cultural context. Traditional translation costs $50-150 per language per SKU. AI localization brings that to $5-15, with same-day turnaround instead of a week.

What I Took Away

This visit reinforced a pattern I keep encountering. The companies that have moved furthest with AI adoption are often the ones with the most to gain from structured training -- precisely because they have already cleared the psychological barriers. They are not asking "should we use AI?" They are asking "why is our AI output inconsistent?" and "how do we scale what is working?"

The answers are usually not about better tools. They are about better systems: prompt libraries, model selection knowledge, data classification frameworks, and documented workflows that make individual experimentation into organizational capability.

This team had the willingness and the tools. What they needed was the methodology to connect the two. That is the gap I spend most of my time filling.

I train corporate teams on practical AI adoption -- from initial awareness to embedded workflows. If your e-commerce or cross-border team is using AI but not seeing consistent results, I have probably seen your exact problem before. Connect with me on LinkedIn.