Digital Employees in Production: 92% Decision Agreement at MINISO, +59% Livestream Sales at Winner

Two companies at the Feishu CLI Developer Day in Shenzhen shared numbers, not slideware: Winner Group (稳健) and MINISO (名创优品). Both run "digital employees" in production — agents with a defined job, real permissions, and measurable performance. This post pulls apart what they actually built, the standard MINISO uses to qualify a digital employee, and the lark-cli building blocks so you can stand up your own.

MINISO's five-part standard for a digital employee

Before the cases, the most useful takeaway: MINISO's definition of a qualified digital employee. It needs all five:

• A clear job definition — it has a position, not just a prompt.
• Company-specific skills — tuned to your processes, not generic.
• Quantifiable KPIs — you can grade its work.
• Standardized tool permissions — it can only touch what its role allows.
• Long-term memory that keeps iterating — it accumulates context over time.

That last three are where lark-cli does the heavy lifting: permissions, measurable output written to a Base, and a memory built from your own meetings and docs.

MINISO: 92% decision agreement, 70% less manager comms time

MINISO pulls the company's full set of group chats and meeting records through the CLI, distills the decision logic of core scenarios, then validates it with multi-model checking and multi-version "horse racing" to keep the digital employee's decisions aligned with a real manager's.

Result: the management-type digital employees now in production reach 92% decision agreement with real managers, auto-answer employees' routine questions, and cut managers' daily communication time by over 70%.

The "learn how we decide" loop:

The distillation prompt:

Winner Group: three production workflows

Winner Group shared three concrete builds, all on the same "CLI connects, LLM judges" split:

• 618 daily business report, automated — the CLI pulls sales, ad-spend, and complaint data from every channel; the model generates a structured daily report that drives the daily promo review. Ended with both sales and profit beating target.
• 24-hour livestream monitoring — built by a business person on the aily agent. Auto-generates a data report every hour and pushes real-time alerts on anomalies. Big-promo livestream sales hit 120% of target, +59% YoY.
• Agent-as-judge — in an internal data-analysis contest, the agent auto-joined the meetings and scored every team's roadshow against a preset rubric. Its scores matched the human judges' to a very high degree.

The livestream-monitor skeleton (the most copy-able):

The one line that ties it together

Winner Group's own summary is the sentence worth keeping: Feishu CLI owns connection and execution; the LLM owns understanding and judgment. A digital employee is just that split, given a job description. Define the role, wire the pull-and-write-back loop with lark-cli, grade its output against a human, and iterate the memory. That's the whole playbook.