Picture this: your boss tells you tomorrow you can’t use Slack, Gmail, or even an iPhone. You have to switch to a clunky internal alternative that no one in the industry actually uses. You’d rage. You’d update your LinkedIn. You’d start drafting a resignation letter.
Now imagine that rage is exactly what your company needs to build a world‑class product.
That’s the brutal, beautiful logic behind Alibaba’s recent move. On July 3rd, they sent a notice: no more Claude. No more GPT. No more Anthropic tools of any kind. From now on, every single employee — engineers, marketers, PMs — must use Alibaba’s homegrown models: Qoder and Tongyi Qianwen (Qwen).
Publicly, it’s about data compliance and supply chain security. Privately? It’s the best product decision they’ve made in years.
Most AI companies chase benchmarks. The smart ones chase their own people’s frustration.
When you force your entire workforce to use your own AI, something magical happens. Feedback stops being vague survey data and starts being real, angry, immediate complaints. Your colleagues don’t say ‘the UX needs improvement.’ They say ‘this code autocomplete is absolute garbage and it just cost me two hours.’ That’s not a complaint. That’s a product roadmap.
The tech world calls this dogfooding — eating your own dog food. It’s been Silicon Valley gospel for decades, but in the AI gold rush it’s been dangerously ignored. Most AI labs optimize for MMLU scores and GSM8K benchmarks. They race to top the leaderboard to impress investors, while real users uninstall their apps after a week.
You can’t build a product people love if the people who build it use something else.
Look at DeepSeek. They came out of nowhere, spooking the whole Silicon Valley. Why? Because their parent company, Huanfang Quantitative, made their hundreds of quants and engineers use DeepSeek every single day for real trading work. Those are the most brutal users alive — a model error costs millions. They forced the feedback loop to be dense, real, and relentless. DeepSeek didn’t chase benchmarks; it chased its own employees’ pain points. The result? A product that actually works in the wild.
Tencent pulled the same trick. TAPD, their project management tool, was used internally for twelve years before they ever sold it externally. Twelve years of internal screaming and feature requests. When it finally launched to the public, it instantly dominated because it had been battle‑tested by tens of thousands of Tencent employees. Their new AI workbench, WorkBuddy, followed the exact same path — born inside, used by everyone, then released as a commercial product. They haven’t even set a monetization target. Why should they? It already works perfectly for the people who know it best.
Google Gemini? Same story. 180,000 employees using it across Workspace, search, and internal workflows. Every bug, every response delay, every awkward interaction gets logged and fixed at scale.
Even Li Auto, the electric car company, lives by this. Their CEO, Li Xiang, is famously a super‑fan car guy. He personally tests every inch of the seat, every HUD angle, every millisecond of the infotainment lag. He’s the most successful user of his own product. That’s why their cars feel polished in ways competitors can’t copy.
Now back to Alibaba. Until last week, their own engineers were happily coding on Claude and writing copy on GPT. Qwen was an afterthought inside the company. That’s a death sentence for any AI product. If your own team doesn’t depend on your product daily, you’re flying blind.
By cutting off external tools, Alibaba has forced a high‑density, zero‑delay feedback loop into existence. Over the next few months, tens of thousands of employees will find every single flaw in Qwen and Qoder. They’ll complain on Slack, in meetings, at lunch. And those complaints — as painful as they are — will become the most valuable product data money can’t buy.
Short‑term morale drop is the down payment on long‑term product superiority.
So what should product managers learn from this?
- Become your own first user. Before you write a single PRD, ask: would I use this every day under pressure? If not, start there.
- Replace surveys with behavior. Stop asking users what they think. Watch what they do. Dogfooding turns every frown into a product requirement.
- Ignore the leaderboards. Benchmarks are the entry ticket, not the moat. The real differentiation lives in the unfixable feeling of a fast response, graceful error handling, and intuitive flows. Only deep internal usage can surface those.
Alibaba’s move isn’t a sacrifice. It’s a strategic hijacking of their own workforce to build a moat that no benchmark can match. The pain is the point. The complaints are the code.
If you build something you wouldn’t use yourself, don’t expect anyone else to.
And if you’re an Alibaba employee right now, I’m sorry. But your suffering is the raw material of a killer product.
FAQ
Q: Is forcing employees to use inferior internal tools really a good idea?
A: Short-term, yes it hurts productivity and morale. Long-term, it creates the most realistic, high-frequency feedback loop possible. You get real usage data, not survey fluff. The key is to be ruthless about fixing what they complain about — then the tool becomes genuinely better.
Q: What if the internal tool is so bad it destroys the company's own workflows?
A: That's why you need a strong feedback culture. The goal isn't to let the product stay bad — it's to surface every flaw systematically. If the tool is outright broken, you'll know within days, and the pressure to fix it is immediate. Dogfooding without a commitment to rapid iteration is just torture.
Q: Isn't this just a way to cut costs on external tools?
A: Cost savings are a side effect, not the core. The real benefit is strategic: you build a proprietary feedback engine that no competitor can replicate. External tools give you generic features. Internal pressure gives you a unique, battle-tested product. Apple, Google, and Tencent all proved that dogfooding is the shortest path to a moat.