# After the 'Raising Shrimp' Rush: Why Enterprise AI Cannot Run on Mass Adoption Alone
Canonical URL: https://theunclej.com/blog/meituan-raising-shrimp
Markdown URL: https://theunclej.com/blog/meituan-raising-shrimp.md
Description: ![](https://assets.theunclej.com/uploads/3679df9f572694d6dd7f838a21a508e9.webp)
Category: ai-organization
Tags: ai-organization, piece-00000149
Published: 2026-10-03T03:04:54.892Z
Updated: 2026-10-03T03:04:55.165Z

---

![](https://assets.theunclej.com/uploads/3679df9f572694d6dd7f838a21a508e9.webp)

> Evidence posture: This Field Note is based on a republished transcript of a public talk attributed to Wang Puzhong, together with identifiable event coverage and Meituan product material. It is not an official Meituan post-mortem, an independent operating audit, or proof that AI × Organization OS has been validated.

The Chinese-language account behind this Field Note uses a colloquial phrase, _yang xia_—literally, “raising shrimp”—for an internal wave of experimentation with OpenClaw agents. Give people an agent. Let them try it. Watch the activity spread.

The nickname is vivid, but it can distract from the more consequential question: what changes after the organization has got everyone using the thing?

According to a republished transcript of Wang Puzhong’s public talk, Meituan’s early-2026 experience unfolded in four stages. In February and March, the company ran a broad internal OpenClaw adoption push. Wang described it as an organization-wide introduction to AI, but also as a period in which cost pressure, permission and security concerns, and erroneous outputs began to affect real operations. From April, he said, individual business units established AI organizations. By June and July, projects were being run in parallel to clarify the method; the account says that some internal products and workflows were beginning to operate more smoothly.

The transcript names five mismatches in enterprise AI adoption: cognition, efficiency, scenarios, evaluation, and workflow. The last is the hinge. A model may be in many hands and still remain outside the operating flow where a company creates value, makes decisions, and handles errors.

That is why this is not a story about “Meituan failing at AI.” It is a public process account of the distance between adoption and organizational capability. Usage, demos, and token consumption can show that an organization has encountered a new capability. They cannot, on their own, tell us which workflow has changed, which business owner accepts an AI output, which errors are barred from entering operations, or who owns the cost of repair.

This is where the account meets the question behind AI × Organization OS: **AI activity is not organizational capability.** The practical move is from mass exposure to operating design. Choose a consequential workflow. Define the business result and the conditions for accepting it. Set boundaries for permissions, cost, exceptions, and repair. Then let business, HR, and technology redesign the operating system together.

But this account cannot supply the missing answers for us. The available materials do not fully disclose workflow baselines, the scope of reported costs, the final point of accountability, or the authority to reject, pause, and repair AI outputs. This Field Note therefore makes one limited claim: broad adoption alone is not enough. It does **not** claim that Meituan lacks these mechanisms, that its outcomes have been independently audited, or that any single framework has solved the problem.

The point is not to forbid broad AI use. Organization-wide experimentation can build literacy. The point is what comes next: move quickly from enthusiasm to a real workflow, real acceptance conditions, and real responsibility. Otherwise the activity arrives first. Capability may not.

## Sources, interpretation, and unknowns

| Layer | Scope |
| --- | --- |
| Attributed account | The five mismatches, four-stage account, timing, cost/permission/error pressures, and business–organization–technology framing are attributed to the [republished NetEase transcript](https://www.163.com/dy/article/L4A17KI00517N211.html). |
| Contextual corroboration | [Southcn's event report](https://news.southcn.com/node_810c33d731/3652f2f829.shtml) confirms Wang spoke at the 2026 Xipu Conference and used a business/organization/technology systems framing. [Meituan's product material](https://www.meituan.com/news/NN260327175004721?source=relativeNews) provides only contemporaneous context on merchant-facing AI work. |
| AI × Organization OS interpretation | “AI activity is not organizational capability” is this article's analysis, not a Meituan fact. |
| Still unknown | Cost accounting scope; workflow baselines; acceptance, pause, repair, and accountability mechanisms; and whether any outcomes were repeatable at organizational scale. |

## Read the Chinese Field Note

[美团“养虾”之后：AI 为什么不能只靠全员使用](/zh/blog/meituan-raising-shrimp)