---
title: "The recruiter that never sleeps: how BOSS直聘's AI agent (智能体) lifted hiring efficiency 25%"
date: 2026-11-13
category: AI in Action
site: NeuroAI
canonical: https://neuroai.site/a/na-app-boss-zhipin-ai-recruiter
language: en
---

# The recruiter that never sleeps: how BOSS直聘's AI agent (智能体) lifted hiring efficiency 25%

> BOSS直聘 (NASDAQ: BZ) deployed an AI agent (智能体) on DeepSeek-R1 plus its Nanbeige (南北阁) model — recruiters using it close roles 25% faster.

A hiring manager posts a job at midnight and wakes to a shortlist. Not from an agency — from a large model (大模型) that read 200 resumes while she slept.

This is no longer a pitch deck. On BOSS直聘 (NASDAQ: BZ; HK: 2076), an AI agent (智能体) now scans candidates, drafts outreach, and runs mock interviews at scale.

The company says recruiters who switched it on closed roles **25% faster**.

## What the agent does on each side of the market

On the employer side, the agent reads a hiring brief, searches the pool, and proposes matches with a reason attached. Recruiters who used it saw matching precision rise and their "deal-closed" efficiency improve by **25%**, per the company's internal measure.

On the jobseeker side, the tool is broader:

- An **AI job assistant** is fully rolled out to all users, recommending roles and rewriting resumes.

- **AI mock interviews** are open to students and workers with under three years of experience; the questions adapt to the person's background.

- During one spring recruitment drive, the AI recommended **more than 100,000** roles to students.

- With permission, it builds a **customized job report** — fit analysis, optimization tips, and a communication strategy — from the user's resume and goals.

The point is not a single trick but coverage of the whole funnel, from "I need someone" to "I think I'm ready."

## The model behind it

BOSS直聘 (BOSS直聘) runs the agent on a stack of DeepSeek-R1 plus its own in-house "Nanbeige" (南北阁) model. The research team behind Nanbeige (南北阁) had work accepted at top conferences including KDD, ACL, and ICLR in 2025, the company said — a sign the deployment rests on a real modeling effort, not a thin wrapper.

The design choice is deliberate: a general reasoning model for language, a specialized model for the labor market's messy, multilingual, resume-shaped reality.

Crucially, the employer agent is conservative by design. Management has said publicly it placed no ceiling on the robot's capability but remained "extremely cautious" about when, how widely, and under what conditions to switch it on — a rare admission that deployment speed is a governance choice, not just an engineering one.

## The safety tax

An open hiring market is also a fraud market. BOSS直聘 (BOSS直聘) says AI lifted its content-review efficiency by **30%**, and in the first quarter of 2025 it pushed safety warnings to users more than **50 million** times. In the same period it worked with police on cases that led to **26 busted operations** and **205 arrests**.

That is the unglamorous half of the story: the same agent speed that helps a recruiter also helps a scammer, so the platform spends real compute on catching both. Beyond catching fraud after the fact, the platform now requires certain categories of employers to disclose any fees upfront, a rule designed to starve the scam economy of its most common hook.

## The business backdrop

The AI push lands on a recovering market. BOSS直聘 (BOSS直聘) reported third-quarter 2025 revenue of **2.16 billion RMB (about USD 303 million)**, up **13.2%** year on year, with paid enterprise customers reaching **6.8 million** over the trailing twelve months. In the first quarter the company posted revenue of **1.92 billion RMB (about USD 270 million)** and net profit of **512 million RMB (about USD 72 million)**.

The platform was also named to Morgan Stanley's "China AI 60" list, a recognition less of hype than of a working product with real usage.

The agent does not operate in a vacuum. It sits on a two-sided marketplace where, by the company's own figures, paid enterprise customers reached 6.8 million over the trailing twelve months and jobseekers number in the hundreds of millions. That scale is exactly why a 25% efficiency lift matters: multiplied across millions of postings, even a small per-role saving compounds into a large labor-market effect. It is also why the bias question weighs heavily — the same model shapes who gets seen by whom at national scale.

## Honest limitations

- **The 25% is self-measured and selection-biased.** It covers recruiters who opted into the agent — likely the more motivated ones — not a random sample. Treat it as directional, not audited.

- **The 30% content-review efficiency gain is likewise self-measured** by the company and should be read as directional, not audited.

- **Matching can harden bias.** Any model trained on historical hiring can quietly reproduce it: who got hired before becomes who gets recommended next. The company has not published bias-audit details.

- **Auto-generated "advantages" can backfire.** Jobseekers get AI-written highlights pushed to recruiters; done poorly, that floods inboxes with polished but generic profiles.

- **Mock interviews rehearse, they don't place.** Practice helps, but no tool closes the offer.

- **Fraud is down, not gone.** 205 arrests in a quarter is also proof the problem persists at scale.

## What readers can do now

- **Jobseekers:** use the mock-interview tool to rehearse, but own your story — review the auto-generated "advantages" before they reach an employer, and don't let the model speak for you.

- **Recruiters:** pilot the agent on a single role and measure time-to-hire yourself; watch whether your candidate pool narrows toward the familiar.

- **Everyone:** ask platforms how AI matching is monitored for bias and what appeal path exists when the agent gets it wrong.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-boss-zhipin-ai-recruiter
Free to quote with attribution and a link to the original.
