Five years in recruiting tech. BOSS Zhipin during hypergrowth—50 million to 200 million users. Then running the whole online platform at Liepin. I'm not going to pretend this is objective. It's not.
2018, BOSS Zhipin. Two million job seekers chatting with hiring managers every day. Insane volume. Growth was so fast that on iOS, we went months without shipping new features—just stability fixes. The product team was frustrated, but leadership made the call: don't break what's working. And it was working. BOSS Zhipin had the highest next-day retention rate of any product I've ever seen. And you know what was powering all of it? Keyword matching. Basic filters. Recruiters burning through hours doing what was essentially data entry. I remember thinking: surely someone has figured this out by now?
Nobody had. Still hasn't, really.
Now I'm building Metix AI, and honestly? Most "AI recruiting" tools I've seen are just lipstick on the same pig. This is me trying to make sense of how an entire industry got stuck.
90s and 2000s: applicant tracking systems. Taleo, SuccessFactors, Workday. Taleo got bought for $1.9 billion in 2012. Makes sense—before ATS, recruiters were literally shuffling paper resumes in filing cabinets. Digital > paper. Simple math.
Here's what nobody talks about though. These systems weren't built to find good candidates. They were built for compliance. EEO reporting. Audit trails. CYA documentation. The resume became the data model—a static document people update once a year, stuffed with keywords specifically designed to game the algorithms screening them.
I've watched recruiters at big Chinese tech companies spend 70% of their time on data entry. Not talking to candidates. Entering data. Into systems supposedly designed to help them hire.
We built bureaucracy and called it technology.
2007-2008. Glassdoor launches. Suddenly companies realize—oh shit, candidates are talking to each other.
Rich Barton (Expedia founder) supposedly got the Glassdoor idea after accidentally leaving employee survey data on a printer. That printer accident turned into a $1.2 billion acquisition. The pitch: transparency gives candidates leverage, companies will pay to manage that leverage.
This era gave us HireVue, ZipRecruiter, and a whole ecosystem of "make your company look good" tools. HR people started calling themselves "talent marketers."
Did it improve hiring? I honestly don't know. Better branding = more applicants. But more applicants ≠ better hires. Often it just meant more noise. More rejection emails. More candidates ghosting when reality didn't match the careers page.
The one good thing: people finally realized candidates are customers too. Everything else? Still broadcast mode. Spray job ads, collect resumes, screen in bulk. Same problem, shinier packaging.
2010s. Greenhouse, Lever, SmartRecruiters. Structured interviewing. Collaborative hiring. Scorecards.
Credit where it's due: Greenhouse actually pushed real research into practice. Structured interviews—same questions for everyone, consistent scoring—are genuinely more predictive. That matters.
But here's my problem with this whole era. I watched it from China with increasing frustration. The bottleneck isn't interview quality. It never was. The bottleneck is sourcing. Finding people. Getting them to respond. Getting them into your pipeline in the first place.
At BOSS Zhipin we attacked sourcing head-on. Real-time chat between candidates and hiring managers. Skip the recruiter, skip the application form, just start talking. Job seeker online, boss online, matched—conversation in seconds. That velocity? That's what actually moves the needle.
The younger the user, the more they wanted IM. This wasn't a guess—we had the data. When I moved to Liepin, they tried to catch up by adding privacy-protected phone calls. Direct voice connection, masked numbers, all the bells and whistles. It didn't matter. You can't reverse a generational shift. People under 30 don't want to call strangers. They want to text.
Meanwhile American HR tech kept layering process on top of broken sourcing. Perfect your scorecards all you want. If your best candidates never enter the funnel, you're just picking the least-bad option from a mediocre pool.
Textio was cool—NLP to make job descriptions more inclusive. T-Mobile got 17% more women applicants. Real result. But still just optimizing one node in a broken system.
2020. Suddenly every HR vendor has "AI." Rules engine? Now it's "intelligent automation." Keyword matcher? "AI-powered screening." The term became meaningless.
Some real stuff did happen. HackerRank built actual technical assessments that predict job performance. Indeed went from job board to behavioral data platform. Eightfold trained on 1.6 billion career profiles—that's real ML infrastructure.
But most of what I've seen? Garbage dressed up in buzzwords.
There's the BERT wrapper—fine-tune an off-the-shelf language model on resumes and JDs, call it "AI matching." Works about as well as keywords, burns way more compute. The problem was never the embedding. Resumes and job descriptions are both lies. You're matching lie to lie.
Then there's bias laundering. Train on historical hiring data (which encodes every bias of past human decisions), then call it "objective" because algorithm. HireVue had to kill their facial analysis in 2020. CEO said it "wasn't worth the concern." At least he was honest.
And my favorite: the feature checkbox. Add one AI feature—chatbot, resume parser, whatever—to your legacy system, market the whole thing as "AI-powered." Average enterprise ATS now has 15+ integrations, each doing their own little AI thing, none of them talking to each other.
Real talk: as of 2024, most AI recruiting is incremental polish on broken foundations. Still human-directed workflows. Still keyword matching with ML sauce. Still spray-and-pray outreach with slightly better targeting.
NYC Local Law 144, 2023. Bias audits for automated hiring tools. EU AI Act classifies HR AI as "high-risk." EEOC tightening guidance on algorithmic discrimination.
These regulations exist because the industry fucked up. Algorithmic hiring tools discriminate. That's documented. Audit requirements make sense.
But think about what mandatory audits actually mean: the technology never earned trust. Companies shipped systems they didn't understand, made decisions they couldn't explain. Now governments are forcing transparency that should have been there from day one.
Regulation will accelerate. "Move fast and break things" in HR AI is over. Probably good for candidates. For anyone building in this space—including me—it means documentation and explainability aren't optional anymore.
Okay, full disclosure: I'm building an agentic recruiting platform. So take this section with that bias in mind.
Traditional HR software: human uses tool. Search for candidates. Send email. Schedule interview. Human directs every step.
Agentic flips it. You say "find me a senior backend engineer with Kubernetes experience in Denver, get them scheduled for a technical screen." System figures out how. Multiple agents coordinate—one searches, one evaluates, one reaches out, one schedules.
Paradox is the obvious example. Their chatbot Olivia does screening, scheduling, engagement in 100+ languages. GM claims $2 million saved annually. Chipotle went from 12 days time-to-hire to 4. I've verified these numbers independently—they're real.
Eightfold and Beamery are building similar architectures. AI that can source, nurture, manage pipelines with decreasing human involvement.
Now the caveats, because I'm trying to be honest:
Most "agentic" tools are still mostly rules. Marketing says autonomous. Reality is scripted workflows with AI filling gaps. Truly autonomous hiring decisions are rare. Probably should be—stakes are too high.
Integration is still a disaster. Average enterprise: 15+ HR tools. Data standardization doesn't exist. Everyone's building point solutions claiming to be platforms.
And candidate experience? Mixed bag. Some people like chatbots. Others feel dehumanized. Research shows AI interviews increase anxiety. We're trading recruiter efficiency for candidate experience and we don't really know the tradeoffs yet.
Same pattern everywhere I looked. BOSS Zhipin. Liepin. Enterprise teams. The bottleneck is never screening. It's reachability.
Best AI matching in the world doesn't matter if your emails hit spam. If LinkedIn InMails go unseen. If outreach never connects. You're optimizing in a vacuum.
So Metix AI is built around what I call the Reachability Graph. Not just who to contact—how, when, which channel. Email deliverability. LinkedIn viewability. SMS timing. Phone pickup rates. Every node has measurable conversion. We optimize the path, not just the destination.
Architecture: multi-agent. Intake Planner translates job requirements into search specs. Search Agent finds candidates. Engage Agent handles multi-channel outreach. Monitor Agent tracks what's working. They all coordinate around one metric: Qualified Interviews Booked (QIB) per seat per week.
Why QIB? Because that's what recruiting teams actually care about. Not "candidates sourced" (vanity metric). Not "applications received" (volume theater). Actual interviews with qualified people, on calendars. That's the output.
The recruiter role is going to split. Executive search stays human—those relationships are too valuable to automate. High-volume hourly hiring goes mostly automated. Chipotle doesn't need humans screening cashier applications. The interesting fight is the middle: skilled roles that need judgment but have patterns. That's where everyone's competing.
Resumes die slowly. I've been saying this for years and it still hasn't happened, so maybe I'm wrong again. But the signals that actually predict performance—demonstrated abilities, project history, how fast someone learns—those are increasingly capturable without relying on self-reported BS.
Consolidation is coming. HR tech is absurdly fragmented—thousands of point solutions. The platforms that actually integrate full workflows will swallow the pieces. M&A incoming. Some of us get bought. Some of us die. That's just how it works.
And compliance becomes a moat. Explainable, auditable AI is table stakes now. Companies that built it in from day one can sell to regulated industries. Companies that didn't? Good luck retrofitting. (This is one reason I'm actually optimistic about Metix AI—we started with explainability because I'd seen what happens when you don't.)
Video interviews. Thought they'd go universal post-COVID. They didn't. Most companies retreated to phone screens. Tech was ready, people weren't.
Legacy ATS stickiness. I left BOSS Zhipin thinking enterprises would migrate to modern platforms within a few years. Nope. Oracle Taleo—ancient, clunky, hated by everyone who uses it—still dominates. Switching costs are just too brutal.
Candidate sophistication. I expected job seekers to adopt tools that gave them leverage. AI agents that apply for you, negotiate salaries, optimize your search. That market exists but it's way smaller than I thought. Most people still apply manually, one job at a time.
Lesson: this market moves slower than the technology. Cultural change in HR departments, candidate habits, enterprise procurement cycles—that's the rate limiter. Not the AI.
Does better recruiting tech actually produce better hires?
I don't know. And that bothers me.
Most HR metrics are efficiency metrics. Time-to-fill. Cost-per-hire. But the thing that actually matters—did we hire the right person? Are they performing? Are they staying?—we barely measure.
Whoever cracks this wins. Connect recruiting inputs to performance outputs. Close the loop. Then you can actually optimize the right thing instead of just making the funnel faster.
Everything else is theater.