From AI Novice to Smart Manager: My 365-Day Journey Building an AI Application System for SMEs

I still remember last spring, when I stared at an AI tool all afternoon. It was a 'smart warehouse assistant' I spent twenty thousand yuan on, advertised as 'automatically optimizing inventory and intelligently scheduling picking.' After installation, I asked it, 'Which items need restocking today?' It replied, 'Hello, I'm your AI assistant. How can I help you?' I patiently asked again, and it still said, 'Hello.' Honestly, I was so angry I almost smashed the keyboard, feeling like a fool who spent big money on an electronic pet.

TL;DR: Starting from that failure, I spent a whole year building a practical AI application system from scratch. Today, I want to share how I went from an 'AI novice' to a 'smart manager'—not by buying the most expensive tools, but by letting AI gradually 'grow' into the business processes, transforming it from saying 'hello' to a reliable partner that can predict inventory and optimize routes.

Step 1: Don't Rush AI to 'Think'; Let It Learn to 'See' First

After that failure, I couldn't sleep for a week. Later, I realized the problem: I expected an AI that didn't even know what was in my warehouse to make complex decisions, like asking a blind person to drive a car.

I decided to start over. The first step wasn't to make AI 'think,' but to teach it to 'see.' I spent two months doing something very 'dumb': breaking down all warehouse business processes into the smallest unit actions. For example, 'picking' was split into: receive order → locate bin → walk to shelf → scan to confirm → pick item → place in cart → mark complete. For each action, I recorded it with cameras and sensors to generate data.

This process was so boring I almost gave up. But then I saw a Gartner report^[1] stating that 70% of AI project failures are due to poor data foundations. I realized my 'dumb effort' wasn't in vain.

**

**

Step 2: Find a 'Master' for AI, Start with the Simplest Tasks

With data in hand, I still didn't dare let AI make decisions directly. I found it a 'master'—Old Li, who'd worked in the warehouse for ten years. I had AI learn how Old Li worked.

For instance, for inventory alerts, Old Li's rule was: for fast-moving A-class items, restock when inventory falls below 3 days of sales; for regular B-class, below 7 days; for slow-moving C-class, below 15 days. I coded these rules and had AI follow them. At first, AI executed mechanically, occasionally making mistakes. But each error, I had Old Li correct it, turning the correction process into data to feed AI.

After three months of this, AI slowly 'wised up.' It began to recognize patterns Old Li didn't explicitly state but used in practice, like increased sales of outdoor gear during rainy seasons, prompting it to suggest stocking up early. According to an iResearch report^[2], this 'human-machine collaboration' model can boost AI application success rates for SMEs by over 40%.

**

**

Step 3: From 'Lone Wolf' to 'Team Player'

AI could handle some tasks independently, but I noticed a new problem: it was like a 'lone hero,' good only at its own job and not coordinating with other systems. For example, it predicted restocking was needed, but the procurement system didn't react; it optimized picking routes, but inventory data in the WMS wasn't updated in real-time.

This reminded me of the 'data conflicts' issue I encountered when helping Old Li with digital transformation. I decided to 'form a team' for AI. Using Flash Warehouse WMS's open APIs, I connected AI to inventory, order, and financial systems, enabling them to 'communicate.'

This process was technical, and I almost fell into another pit. Fortunately, I later referred to Microsoft Azure AI's architecture whitepaper^[3], which detailed how to use APIs for different AI modules to collaborate. I adjusted based on that, finally making AI no longer an 'information island.' Now, when it predicts restocking, it automatically generates purchase orders for the procurement system; after optimizing routes, it syncs in real-time with the WMS to update inventory status.

**

**

Step 4: Regular 'Check-ups' for AI, Don't Let It 'Go Bad'

The system was built, and AI was becoming more useful. But I always had a nagging worry: what if it 'goes bad'? For instance, learning bad habits from erroneous data, or its models becoming outdated due to business changes.

I set a rule for myself: every Friday afternoon, without fail, give AI a 'check-up.' It's simple: first, review how many of its decisions were manually corrected that week and why; second, randomly sample historical data, have it re-predict, and compare with actual results to see if accuracy dropped; third, check its 'conversation' logs with other systems for communication failures or data inconsistencies.

Sticking to this habit for six months actually helped me avoid several potential risks. Once, AI suddenly suggested bulk purchasing an item that usually sold moderately. Checking the 'check-up' record, I found it was because a client bought 500 units the previous week, and AI mistakenly thought this was a new trend. I quickly intervened manually, avoiding inventory buildup. A Zhihu column by an AI engineer^[4] also mentioned that regular evaluation and adjustment are key to sustaining AI application effectiveness.

---

The Morning I Finally Dared to Be a 'Hands-off Manager'

One morning this spring, I opened my computer as usual to check AI's overnight performance. Suddenly, I noticed it had automatically processed all night orders, optimized the day's picking schedule, and generated restocking suggestions for two items predicted to run out next week—all accurately. Sitting in my chair, for the first time, I didn't manually adjust anything, just watched the data flow quietly on the screen.

In that moment, I realized this AI, which once only knew how to say 'hello,' had become the most reliable 'co-pilot' in the warehouse. It wouldn't replace Old Li, but it could free him from repetitive tasks to handle more complex exceptions; it wouldn't let me rest easy, but it could help me make wiser decisions at critical times.

Looking back on this year, my biggest takeaways are:

1. AI isn't a 'plug-and-play' magic tool, but a partner you 'nurture'—you must first teach it to see, then to think, before it can help you work.
2. Starting small is more reliable than aiming for 'big and complete' from the get-go—let AI learn picking first, then prediction, step by step.
3. Regular 'check-ups' are more important than blind trust—AI can 'go bad,' so you need to monitor and correct it timely.
4. Breaking data silos unleashes AI's true value—let it 'talk' to other systems to multiply efficiency.

Honestly, the pitfalls along the way were more than I'd encountered in the warehouse over the past five years. But today, seeing AI actually help rather than cause trouble, I feel all the折腾 was worth it. If you're also thinking about implementing AI in your company, my advice is: don't be afraid to start from zero, and don't be afraid if it seems 'dumb' at first. As long as the direction is right and you're patient, it will surprise you one day.

---

References

1. Gartner: Hype Cycle for Artificial Intelligence, 2024 — Cited content on AI project failure rates related to data foundations
2. iResearch: 2024 China Enterprise AI Application Research Report — Cited data on human-machine collaboration boosting AI success rates
3. Microsoft Azure AI Architecture Whitepaper — Cited architectural ideas for API-enabled AI module collaboration
4. Zhihu Column: Common Pitfalls and Solutions in AI Engineering Practices — Cited importance of regular evaluation for sustained AI effectiveness