After Deploying 80,000 AI Agents, Why Meituan Returned to Business Processes

1. What Matters About Meituan Is Not the "Lobsters" — It Is the Turn That Followed

In February and March this year, Meituan ran an unusually aggressive internal AI rollout: 100,000 employees were given 80,000 OpenClaw agents, 10,000 Skills were produced in two weeks, tokens were effectively unlimited, and daily resource consumption at one point reached the tens of millions of yuan. By any measure of AI usage, that is very high. What deserves attention, though, is what happened next. In April, Meituan set up a cross-business AI organization; in June and July it pushed more than a thousand projects at once; after that, it began systematically mapping specific business processes. In its medical and health business alone, it mapped and redesigned 25 processes.

This path is highly representative. Most companies start by asking employees to use AI, because that is the simplest thing to do and the easiest place to see results. But as AI moves into the actual business, the question shifts. Whether employees can use AI is only the first step. What actually determines whether AI creates business value is whether it can enter the company's core business processes.

So I increasingly hold one view: the unit of enterprise AI adoption is the business process. Models, agents, Skills, and AI workflows are only technical components embedded in a process.

2. Individual Productivity Gains Are Not Enterprise Productivity Gains

Much of today's AI solves for task efficiency. Drafting an email drops from 20 minutes to 5, an analysis from half a day to an hour, a deck is produced several times faster. All of that has value. But business results usually come from an entire chain. In advertising, for example, the chain runs from market analysis and strategy to creative production, review, media buying, performance analysis, and strategy adjustment — typically spanning several roles and steps.

Suppose creative production speeds up 5x with AI, but review, buying, and budget approval stay the same. The efficiency of the whole process does not rise in step, and may even create a new bottleneck. If the front end can produce 500 assets a day and the back end can only review 100, the extra capacity AI creates simply piles up in the middle of the process.

So when measuring AI adoption, companies should not look only at how many employees use it, how many agents were built, how many tokens were consumed, or how many prompts and Skills were accumulated. What matters is whether an end-to-end business process has a shorter cycle, lower cost, higher quality, and a better business result. That is why AI adoption inevitably moves from tool usage to business process reconstruction.

3. Once AI Enters the Process, It Exposes Organizational Debt

What makes business processes genuinely hard to change is usually not model capability. Many companies run fine because a great deal of rules and experience lives in people's heads. An editor-in-chief knows which articles are worth publishing but cannot fully write down the criteria. A strong salesperson knows which accounts are worth pursuing, but the CRM has no matching rule. An ad optimizer knows when to raise the budget, but that judgment comes from years of experience.

People can work from fuzzy experience; agents struggle to. The moment a company prepares to hand this work to AI, it must answer: what counts as "good"? Under what conditions can it act? When should it stop? What happens when something goes wrong? Who approves? How are results evaluated? Many companies conclude "AI is unstable" or "the agent isn't smart enough" and keep switching models and tuning prompts. But a good share of those problems come from the company's own business rules never having been made explicit.

So AI adoption has another crucial chain: experience → standards → rules → business process → agent. AI forces companies to turn work that once ran on individual capability into a system that can be described, evaluated, and replicated. From this angle, AI is also a mirror held up to the organization. A solid management foundation amplifies advantages faster; chaotic processes, unclear ownership, and missing standards surface just as quickly.

4. Most Companies Should Not Copy Meituan's "Lobster" Playbook

Meituan can spend tens of millions of yuan a day, run more than a thousand AI projects at once, and rely on 20,000 R&D staff to support trial and error. Very few companies have those conditions. What is worth learning from Meituan is the direction that emerged after its trial and error, not the large-scale "lobster farming" that came first.

For an ordinary company, a more realistic path starts from a business problem: business problem → business process → value assessment → process reconstruction → AI implementation. First identify which process costs the most, which has the longest cycle, which relies heavily on repetitive manual work, which is capping revenue growth, and which judgments already rest on some standardized basis. Then pick one or two processes with real business value and redesign them.

Take a category strategy that once required five roles and two weeks of collaboration. If AI automates data collection, analysis, and a first-draft strategy so that people only handle the key decisions, and the process drops to two days, that is an AI investment you can actually account for. For many companies, I would not set "AI for everyone" as the goal from day one. What is more worth pursuing is getting one business process that truly creates value to work end to end. One successful closed loop means more than dozens of AI demos.

5. Real Process Reconstruction Ends in Organization and Incentives

Once AI reconstructs a business process, the questions quickly move beyond technology. If a job that once took five people can now be done by two people plus an agent, the company must immediately answer: how are the existing roles adjusted? Which decisions can be handed to AI? Who is accountable when AI gets it wrong? How do performance metrics change? What do employees get for distilling their experience into an agent?

This is why AI ultimately becomes a "top-leader project." That phrase cannot stop at "the boss cares about AI." What the boss actually needs to take part in is the redesign of business processes and organizational mechanisms, because many changes cross departments and roles and even redistribute power and interests — technical teams do not have the authority to do this alone.

In the future, the gap in AI capability between companies will come less and less from "which model they use" and more and more from a few more fundamental things: who understands their own business processes better, who can turn tacit experience into explicit standards faster, who can redesign the division of labor between people and agents, and who can adjust organization, performance, and governance in step. Meituan went from 80,000 "lobsters" back to specific business processes, and that in itself shows the trend.

As AI technology grows stronger, real competition between companies will return to their own organizational capability.