What 10 years in Shanghai taught me about AI in SMBs

I came back to France in early 2026 after ten years working in Shanghai. Ten years in start-ups, international groups, and above all alongside Chinese SMBs that look like nothing we know here.

This text isn't a geopolitical treatise. It's an honest list of what I saw work over there, and what could save an SMB months. Some lessons can be taken as is. Others need adapting. One or two should be discarded outright. I say so each time.

1. Execution speed always beats technical sophistication

In Shanghai, a manager who has an idea on Tuesday tests it on Friday. Not the next month. Not "when the spec is ready". On Friday.

What I saw concretely: a restaurateur added WeChat mini-program ordering in three days. A textile SMB set up production tracking with QR codes in a week. An insurance broker plugged a document extractor onto incoming emails in ten days. None of these solutions was "clean". All were running. All were earning money or time from the second week.

The French reflex is the opposite: we think about architecture for three months, scope for two months, develop for six months. By delivery, the idea is no longer relevant — and the market has moved on.

The rule I brought back: if a POC can't run in 2 weeks, the problem is poorly broken down. Clean architecture comes after, on what proved its value.

You see the trap: in France, we want clean before having proven utility. In Shanghai, utility is proven before aiming for clean. On my automation engagements, I now apply the second method — even when the client was used to the first.

2. Automation is a reflex, not a project

In China, no one says "we're launching an automation project". Automation is daily hygiene: as soon as a task is done manually twice, someone writes a script that evening or the following weekend.

I saw 5-person teams that had automated, in less than a quarter:

• Daily bank reconciliation
• Batch invoice generation (500+ per month)
• Ingestion of supplier orders from emails
• Customer follow-ups via WeChat and email
• Synchronisation between their homemade ERP and accounting
• Weekly reporting sent as PDF to the boss every Monday at 7 a.m.

None of these processes would have justified a "project" in France. Each would have been worth two scoping meetings here — meetings that never happened in Shanghai because delivery was already done before the first meeting ended.

For an SMB, the translation is simple: stop waiting for the big yearly automation project. Do small ones, one by one, ship them within the week, and let them accumulate. My most useful engagements are often the fastest to wrap up — 3 to 4 weeks on a targeted process, and 5 to 15 hours per week recovered by someone on your team.

This is also the takeaway of my article "5 daily tasks you didn't know you could automate": most don't make it past the executive committee in France because they're "too small". That's exactly why they should be tackled.

3. AI embeds in the workflow. Not in a chatbot.

In France, 80% of AI projects I see starting begin with "let's put a chatbot on our site". In Shanghai, I saw very few chatbots. On the other hand, I saw AI everywhere:

• Sorting and scoring product photos before listing
• Extracting data from scanned invoices
• Automatic generation of product descriptions in three languages
• Anomaly detection in order flows
• Drafting replies to customer reviews (with human validation for sensitive cases)
• Classifying incoming emails by urgency and theme

None of these applications is "sexy". None would make a tech press headline. All have a ROI measurable in months.

The chatbot is the visible part and the least profitable. AI that works is the kind you no longer see, that has disappeared into a workflow. That's exactly what I explain in this article on AI agents vs chatbots — value comes from action (modify data, route a ticket, make a decision), not from conversation ("Hello, how can I help?").

Simple rule: if your AI project starts with an interface, it's probably the wrong angle. Start with the thankless task no one likes, and stick an LLM behind it.

4. Patch first, industrialise later

Here's a method I saw ten times in Shanghai, almost never here:

1. A manager identifies a concrete operational problem
2. Someone (often them, often their assistant) hacks together an Excel + emails + copy-paste solution in 48 hours
3. The solution runs 2 to 3 months in manual production
4. Once proven useful, a developer is called in to industrialise the part that works

This approach is considered "dirty" in France. Yet it's the most economically rational: we only industrialise what passed the field test. We don't fund six months of dev for a hypothesis we haven't validated.

The translation for an SMB: if you're hesitating to launch a tool, start with the ugliest manual version possible. Shared spreadsheet, emails, human process. Let it run 4 to 6 weeks. If usage holds, build the real version. If not, you've saved 8 weeks of development and a five-figure budget.

For my clients, I often apply this framework during scoping. I regularly refuse to write code when the manager isn't yet sure what they want — we do manual first, measure, and then build the tool around what worked. Slower in appearance, much safer in reality.

5. Mobile isn't a feature, it's the main entry point

In China, 95% of SMB interactions happen on mobile. Dashboards, purchase order signatures, urgent notifications, weekly reports: everything goes through WeChat or a custom app. The desktop became the secondary channel.

In France, I still see dashboards designed first for a 27-inch screen, with the mobile version added at the end of the project "because we have to". Result: the dashboard is opened once a week in a meeting, then forgotten.

It's backwards. An SMB manager has their phone 14 hours a day and their computer 4. The dashboard giving them the 3 figures that matter on mobile, consultable between two client meetings, is more useful than the PC behemoth nobody opens outside monthly meetings.

Simple rule: if you can't steer your business from your phone in 10 seconds, it's not a dashboard — it's a report.

6. What NOT to copy

Now the reverse part. Not everything I saw in China is good to take. Here are the practices I refuse to import on French engagements:

Wall-to-wall employee surveillance

In Shanghai, many SMBs track everything every employee does: clicks, screen time, movement in the premises, conversations. Legal in China, illegal in France (GDPR, Article L1121-1 of the French Labour Code), and above all counterproductive in the medium term. A surveilled employee does the bare minimum — and the good ones leave at the first competitor who respects them.

Opacity on data

Chinese SMBs collect massively, often without informing the user or client. This is forbidden in France, and rightly so. AI data security is a topic I address first on engagements — before tech, before budget, before functional scope.

The premium on "impressive features"

Chinese tools are often packed with gimmicks — animated stickers, generative AI everywhere, micro-interactions — to impress the end user. It works in mainstream B2C. In B2B SMB, it's noise: what counts is reliability, predictability, and the ability to last five years without supervision.

The "we'll see about maintenance later"

Chinese speed has a flip side: many internal tools run without documentation, with a single person who knows how it works. When they leave, it's a catastrophe. I've seen SMBs lose a year of growth because a single dev resigned without handing over. In France, I systematically deliver documentation, source code, and a handover session. It's one of the rare places where the French model — transmission, traceability, durability — is clearly better. Keep it.

Synthesis: what an SMB can do starting Monday

If you run an SMB and want to act on these lessons, here's where to start:

None of these actions requires a six-figure budget. The most expensive costs a day of dev and three coffees.

And if you want to see what my price ranges look like for engagements outside this scope, pricing is public here — no opaque quotes, no runaway hourly rates.

Where to start

If you want me to look at one of these lessons applied to your specific case, 30 minutes are enough. I'll tell you which make sense for you, which are off-topic, and which to start with. No forced engagement — if it's not yet the right time, I'll tell you frankly.

What I care about is that the tool I deliver lasts at least five years, not that I bill an extra day. It's the only legacy I really kept from the French model.

To go further

• Related service: Workflow automation
• Related articles: Automate your SMB processes — where to start · Integrating AI in an SMB in 2026