On July 7, we announced two changes together. OpenJobs AI adopted the Metix AI brand, and we had raised $5.5 million across two rounds in six months.

The announcement also put a different unit at the center of the product: customers would pay for qualified interviews.

As the co-founder responsible for product, I think that line says more about the company than the new name does. Search software is relatively easy to describe. A recruiter types a prompt, gets a list, and works through it. A qualified interview is harder. The role has to be understood correctly. The candidate has to fit the important parts of the brief, respond to the outreach, want the conversation, and arrive knowing what the company is hiring for. If any link breaks, the customer did not receive the thing on the invoice.

That is the product decision behind the rebrand. The full funding and company announcement is here: https://metix.ai/blog/metix-ai-raises-5-5m.

This essay is my account of that decision, one month after the announcement. I am not reviewing Metix from the outside. I helped build it, and I have every reason to want it to work. I also know which parts of the promise are easy to put on a homepage and difficult to earn on a live role.

July 7 changed two things at once

The name change was real, but it was not a new legal entity, a new team, or a product reset. Metix AI is the customer-facing brand operated by OpenJobs AI Inc. The company remains a Delaware corporation. The login and API stayed in place. Our current company page spells out that boundary because a rebrand should not force customers to guess which contracts, accounts, or systems still apply.

OpenJobs AI was a useful name when the center of the product was access to talent data and better search. It told people the market we worked in. It did not say what we had become responsible for delivering.

Metix comes from Metis, the Greek idea of practical intelligence. I like the name because product work in recruiting has taught me to distrust the gap between knowing and doing. A hiring manager can describe an exceptional engineer. A search engine can identify people who resemble that description. Neither act puts an interested candidate on the calendar.

That missing work is where recruiting projects stall. Someone has to turn an informal request into criteria that can survive a search. Someone has to decide which requirements are firm and which are preferences. Someone has to contact candidates without flattening the role into generic copy. Someone has to interpret a reply, check genuine interest, coordinate time, and carry the context into the interview. Software can perform much of that work now. The buyer still needs someone to own whether the work produced a useful meeting.

Our rebrand made that ownership visible. It also raised the standard we could be judged against. If the product sells search, a large and relevant list can count as success. If the product sells an interview, the list is only an intermediate state.

The funding happened at the same time because the operating model needs capital on both sides. The most recent $3 million round was led by Rsquared Investment, the Singapore company behind Bossjob. An earlier $2.5 million seed was led by LongRiver Investments, with Fengshion Capital and senior technology executives participating. Those checks fund product development, but they also fund delivery capacity and expansion. We are building a system that has software economics in some places and a human quality obligation in others.

That human obligation becomes more important when the company calls itself AI.

Search left the customer with the hard work

We began with a familiar product assumption: recruiting teams needed better tools. We built faster sourcing, cleaner data, and AI-assisted search. Those improvements worked. They also left the customer operating the process after the result page loaded.

We kept building software, but software access stopped being the finished product. Metix is a platform, and repeatable product work should improve its economics over time. We also reject the convenient SaaS definition in which a login counts as delivery. If a customer pays, opens the product twice, and still spends the week sourcing, our usage chart may be active while the customer’s problem remains.

This change sounds obvious in retrospect. It was difficult in practice because it moved work onto our side of the boundary. Every step we took away from the customer became a step we had to operate, measure, and eventually make more reliable. A product team can ship a search filter and watch adoption. An outcome team has to stay with the role until another person agrees to a conversation.

Imagine a founder hiring the first infrastructure engineer. A search product can return 100 plausible profiles before breakfast. By lunch, the founder still has to decide which experience is relevant, write outreach, handle replies, explain the company, check whether compensation and location are workable, and schedule calls. The search took minutes. The job around it can take days.

The same problem appears inside a recruiting team, only at a different scale. Greenhouse’s 2026 hiring benchmark analyzed more than 6,000 companies and 640 million applications from 2022 through 2025. It found annual applications per recruiter rose 412%, from 146 to 746, while recruiters per organization fell 56%. Time to fill still increased 37%, reaching 59.67 days in 2025.

That dataset does not measure Metix or prove that one product model will fix hiring. It does show the pressure on recruiting operators: more inputs arrived, fewer recruiters handled them, and each recruiter closed more jobs while the elapsed time grew.

Adding another source of profiles can make this operating problem worse. A bigger list creates more screening. Faster outreach creates more replies to interpret. Better generated copy can raise volume while making candidate communication less distinctive. The team gains capability and inherits more work.

We saw the product boundary in customer behavior. Customers did not celebrate a list for very long. They asked who they should meet. They wanted to know whether the candidate was interested, whether the role had been explained, and whether the next step was already booked. When a customer still had to run the search after using our search product, we had improved a task without removing it from their week.

That was uncomfortable because tool metrics looked good. Search speed can be measured in seconds. Database coverage can be counted in profiles and countries. Generated messages can be counted by the thousand. These numbers show system capacity, but none of them tells a founder whether Tuesday’s interview is worth forty-five minutes.

We stopped treating the result page as the finish line. Metix AI’s current homepage begins with a different promise: define the role, receive interview-ready candidates within 24 hours, and pay for the people worth meeting. Mira, our AI recruiter, still searches. It also reaches out, screens for interest, and books. A delivery lead reviews quality before a candidate reaches the customer.

The product surface now extends beyond search into work, handoffs, and judgment that were previously left to the buyer. It also gives the buyer a clearer complaint when we miss: this person was unqualified, uninterested, poorly briefed, or wrong for the calendar.

Clear complaints are useful product data. They are much harder to hide behind engagement charts.

A qualified interview can fail in four places

Outcome pricing sounds simple when it is compressed into a sentence. Our pricing page says customers pay for candidates who are qualified and ready to interview. The work underneath that unit is not simple.

The first failure can happen before the search begins. Many role briefs combine a real need with a copied job description, a founder’s memory of a former colleague, and a wish list accumulated across interviews. If the system treats every line as equally important, it can find people who match the text and miss the job.

A useful brief forces tradeoffs early. Which two skills are required on day one? Which experience can be learned? Does the company need a manager who has already led ten people, or a senior individual contributor who wants to build a team? Is the location fixed? Is the compensation range consistent with the people being requested? A search built on a vague brief does not become accurate because the model is powerful.

The second failure is retrieval and ranking. Metix says its platform can source across more than 860 million profiles in more than 190 countries. Scale matters when a role sits outside the customer’s existing network. It does not remove the need to choose the right evidence. Job titles vary by company. A five-year tenure can signal depth in one context and stagnation in another. Two engineers can list the same technology while having used it at completely different levels of responsibility.

A shortlist has to explain why a candidate cleared the bar. Without reasons, it pushes the evaluation back to the customer. We want the role criteria, evidence, and uncertainty to travel with the candidate so the hiring manager can disagree with something specific.

The third failure is interest. A candidate can be an excellent match and have no reason to take the call. Outreach has to make a truthful case for the role without pretending a generated message is a relationship. Our FAQ says customers approve every outreach message before it is sent. That approval is a product feature because the employer’s name is on the message. The employer should know what candidates are being told.

Response is only an intermediate measure. A polite reply may reflect courtesy rather than interest, and a request for salary information can end with a quick rejection. Screening has to establish that the person understands enough of the role to choose a conversation. Otherwise the calendar fills with meetings that create work for both sides and rarely produce a hire.

The fourth failure is the handoff. A candidate can fit, respond, and still arrive unprepared because scheduling stripped away the context. The hiring manager may assume a technical screen while the candidate expects an exploratory call. The company may have changed the role after outreach began. A timezone error can turn a strong match into a no-show.

These failures are why the interview is an unforgiving unit of product. Each booked meeting contains a chain of decisions that can be inspected. Did the brief reflect the actual role? Did the evidence support the match? Did the candidate express real interest? Did both parties receive the same context? We need to preserve those answers instead of reporting only that a calendar event exists.

Pricing around the interview changes incentives, but it does not solve them automatically. A vendor paid per meeting could chase easy roles, lower the qualification bar, or optimize for calendar volume. We have to counter that pressure with explicit criteria, customer feedback, and a delivery review before the handoff. A customer should reject weak work without paying for our attempt. We should be able to learn why it was weak.

I prefer a concrete unit over a broad claim about autonomy. “The agent ran” describes system activity. “The customer met a qualified and interested candidate” describes a delivered result. The second sentence can still be disputed, but at least everyone knows what to inspect.

People belong inside the delivery

The most common question about our model is why an AI hiring company employs a delivery team. My answer is that the people are part of the product specification.

Search, enrichment, drafting, follow-up, and scheduling contain enormous amounts of repetitive work. Software should take that burden. A person does not need to open hundreds of tabs to prove that judgment matters. Human attention is most useful at the points where the role is ambiguous, the model’s evidence conflicts, a candidate reply changes the search, or a hiring manager’s stated criteria do not match the people they accept.

A delivery lead reviews the shortlist before it reaches the customer. That sentence should not be read as a ceremonial sign-off. The reviewer needs enough authority to stop an attractive but weak match, question the brief, and send the search back. If the human merely approves whatever the model produced, we have added labor without adding accountability.

This creates an uncomfortable operating question for an AI startup. Human review costs money each time a role runs. If delivery effort rises at the same rate as customer volume, we have built a service business with software inside it. If we remove reviewers before the system can hold the quality bar, we have built software that exports its errors to customers. Neither outcome is acceptable.

The product work sits between those two failures. We need to record which decisions repeat, which exceptions recur, and which judgments still depend on context that the system does not capture. Automation should take a verified pattern out of the delivery queue. Removing a person merely to make a financial model look cleaner would export the work, and perhaps the error, to the customer. The test is whether the next customer receives equal or better candidates with less operational effort from both teams.

The customer also remains inside the system. Metix does not make the hiring decision. The company defines the role, approves outreach, interviews candidates, chooses whom to hire, signs the offer, and employs the person. Our job is to remove the search work without pretending we own decisions that belong to the employer.

This division matters for candidates. They should know who is contacting them, what role is real, and which company will employ them. They should not have to decode whether an AI message made promises the hiring manager has never seen. Buyer approval and human review create places to catch those failures before candidates experience them.

Human review is not a magic answer to bias. A person can import an unstated preference as easily as a model can reproduce a pattern in historical data. The safeguard begins with job-related criteria that the customer can see and challenge. When a hiring manager rejects a candidate for a reason that never appeared in the brief, the team should decide whether the role changed, the evidence was weak, or an irrelevant preference entered the process. Silently teaching the search to copy every acceptance decision would make the next shortlist less accountable, not more accurate.

Candidates also need a boundary around automation. Our system can help find and contact people, but a candidate should not be reduced to an inferred score that neither side can explain. The point of the shortlist is to create a serious human conversation. It is not to automate the final judgment about a person’s career.

There is also a product reason to keep delivery close. A model can learn from a rejection label, but recruiting feedback is often messy. “Not senior enough” may mean the candidate has not managed people, has not worked at comparable scale, explained the experience poorly, or simply did not match an unstated preference. A delivery lead can push for the missing explanation and translate it into a better next search.

That loop is slower than pretending every rejection is clean training data. It is faster than repeating the same mistake across twenty candidates.

Over time, software should handle more of the routine path. The funding announcement says we plan to deepen the platform so it can run more of the search with less prompting. I want that progress. I do not want the product to erase the checkpoint merely to claim a higher percentage of automation. The useful measure is how much customer work disappears while shortlist quality holds or improves.

The destination is a hiring process where people spend their attention on decisions that deserve it, rather than a dashboard designed to look unattended.

Early numbers need their denominators

Our July 7 announcement reported several results from early enterprise customers. Roles were filled in about a week against a cited industry average of 45 days. Ninety-five percent of roles received a first interview within 24 hours. Outreach drew five times the response rate of cold outreach. The post described hiring cost as about 65% lower. It also said the company was approaching 100 customers, revenue had grown more than 35% month over month, and the platform searched across more than 860 million profiles in more than 190 countries.

I am proud of those numbers. I also want readers to read them correctly.

They are company-reported early operating results, not independent benchmarks across every role and market. The announcement does not publish the size and composition of each metric cohort. A 5x response rate depends on the baseline campaign, candidate segment, sender, role, and message. A one-week fill can mean different things depending on when the clock starts and what counts as filled. A cost comparison depends on whether the alternative is an agency, an internal recruiter, a sourcing subscription, or a founder doing the work after midnight.

Until those cohort definitions are available, readers should treat the figures as early company-reported results rather than a benchmark for every role.

Even our public wording shows why definitions matter. The funding post reported about 65% lower hiring cost from early enterprise customers. The current Metix homepage presents 65% less manual work across roles run on the platform and tells visitors that results vary by role and market. Cost and manual work are different measures. We should not slide between them because the number happens to be the same.

For a customer, the right benchmark starts with the role in front of them. How long does the current process take to produce a candidate the hiring manager accepts? How many recruiter and hiring-manager hours sit inside it? What response rate does the company’s own outreach achieve? How often do scheduled candidates attend? How many qualified interviews are needed to produce an offer?

The measurement record should also keep stages separate. A shortlist precedes an accepted interview, scheduling precedes attendance, and an attended call qualifies only when the candidate matched the agreed brief. An offer becomes a hire after acceptance. When a vendor compresses these events into one funnel number, fast activity can conceal weak delivery.

The same discipline belongs in a cost comparison. Agency fees, software subscriptions, recruiter salaries, founder hours, hiring-manager time, and vacancy delay are different costs. A buyer does not need a perfect economic model before running a role, but they should decide which costs count before seeing the result. Otherwise every vendor can choose the denominator that makes its product look cheapest.

Metix should be evaluated against that baseline. A 24-hour first interview is meaningful only if the candidate clears the customer’s bar. A higher response rate matters only if the replies come from the right people. A lower cost matters only if the calculation includes the work the customer still has to do.

Billing around a qualified interview invites scrutiny. A customer can inspect the candidate, the criteria, the interest, and the meeting. We can still disagree, but the dispute stays close to the work rather than disappearing into a software utilization report.

The early results tell us the model deserves more investment. They do not excuse us from measuring the next cohort more carefully.

The $5.5 million has three jobs

Fundraising is a strange company milestone because it is public proof of investor conviction and no proof at all that the next customer will be satisfied. The money gives us more time and more capacity. It also raises the cost of confusing motion with progress.

We named three uses in the announcement.

First, we will deepen the platform so it can handle more of the search with less prompting. More of the expert work around role interpretation, search refinement, outreach, and screening should become product behavior. The employer will continue to make the final hiring decision while we reduce the manual coordination required to produce a good shortlist.

Second, we will grow the delivery team. This may look unusual beside the first goal. To me, the two belong together. Delivery shows us where automation is reliable and where it fails. It catches weak output today and produces the judgment we need to improve the system tomorrow. The team also gives a customer one accountable owner when a live role does not behave like a demo.

Third, we will expand in North America, where the announcement said most of our new customers already are. Expansion involves much more than translating a landing page. Each market changes the candidate pools, compensation expectations, employer brands, working arrangements, and legal handoffs surrounding a role. We have to learn those differences without turning every exception into permanent manual work.

Rsquared Investment brings direct recruiting context through Bossjob. LongRiver Investments led the earlier seed, and Fengshion Capital and senior technology executives joined. An investor who knows the category can challenge a common product mistake: mistaking more database activity for a better hiring outcome.

Our own team also spans the two halves of the model. Kin Fu, our founder and CEO, has spent more than two decades in talent acquisition. Dr. Zhilin Wang, our co-founder and CTO, built the AI-agent and retrieval systems behind the platform. I have spent my career building recruiting products. None of those backgrounds guarantees the result. Together, they let us argue about the product at the level where it breaks: recruiter practice, model behavior, customer workflow, and candidate experience.

The next stage should make those arguments more demanding. More roles create more edge cases. A larger delivery team creates management and consistency problems. More automation creates new ways to fail at scale. North American growth creates a higher bar for trust from employers who already own a crowded recruiting stack.

The round pays for us to take on that work, and it makes skipping it more expensive.

North America will test the promise

The qualified-interview model is most attractive when a company has a real open role and too little recruiting capacity. A founder needs an early engineer but has no recruiter. A talent team can run interviews well but is buried in sourcing. A company is hiring in a market where its existing network is thin. In each case, the buyer wants work removed, not another place to log in.

North America has many teams in that position. It also has mature recruiting operations, large professional networks, specialized agencies, and buyers who have seen years of AI recruiting claims. Those customers can compare Metix with an internal sourcer, LinkedIn Recruiter, an AI search tool, an agency, or a combination of all four. We have to earn a place in that comparison with delivered work.

The model will not fit every role. Executive search may depend on relationships built over years. A confidential replacement can require controls that do not belong in a broad workflow. Some large recruiting teams already have the people, process, and systems to turn search results into interviews efficiently. For them, a narrower tool may be the better purchase.

We should be specific about where Metix helps. The strongest starting point is an active role with a clear business need, a hiring manager willing to define the bar, and a measurable cost of delay. That gives the platform a brief it can execute and gives the customer a result it can judge.

Global hiring adds another handoff after the candidate says yes. On July 24, we announced a partnership with Deel that connects Metix users to employer-of-record and payroll infrastructure across more than 130 countries. Deel remains the employment and payroll provider, and the customer keeps the final employment decision. The connection is meant to reduce the gap between choosing a candidate and employing that person compliantly in a market where the customer may not have an entity.

That partnership expands the workflow, but I still judge the product at the calendar. Did we understand the role? Did we bring someone qualified and genuinely interested? Did both sides arrive with the same expectations? If the answer is no, a larger network and a global payroll handoff do not rescue the result.

One month after the rebrand, that is the promise I want attached to the name Metix. Practical intelligence should show up as completed work: the customer defines the role, our platform runs the search, our delivery team stands behind the shortlist, and the customer meets people worth considering while keeping the final decision.

If your company has an open role, visit Metix AI and define it. Write down the requirements that cannot move, the points that can, and the evidence you will accept. Then judge us by the people who arrive on the calendar.

Gene Dai is co-founder and chief product officer at Metix AI. He has built recruiting products at BOSS Zhipin, Liepin, and Zuoyebang. This essay reflects his product perspective on the company’s July 2026 funding and rebrand.