Workable's 0.1% Counted Applicants. Passive Engineers Never Applied

If your job is getting a senior engineer who never applied to answer a cold message, Workable's new 0.1% figure tells you almost nothing. Act as if it does and you will quietly burn your shortlist.

At HR Tech 2026 in Las Vegas, Workable reported that only 0.1% of candidates asked to speak with a human when given the option during its AI recruiting conversations. The sample covered 1.2 million applications. Every person in it had already applied.

What did Workable actually measure?

Workable counted how many people clicked to ask for a human while they were inside an AI chat. Every one of them had already applied for a job and been picked for follow-up. The answer was 222 people.

CRO Sarfraz Ali walked through the funnel:

• 1.2 million applications across about 12,000 jobs.
• Workable Agent evaluated every application as it came in and identified more than 350,000 candidates for further engagement.
• Those candidates got outreach, typically within minutes.
• 63% opened the conversation, and 77% of those who opened it finished.
• 222 asked to be handed to a human.

This is selection bias: the measured group was filtered in a way that makes it unlike the group you care about. Three filters stack up. The person applied, the agent picked them, and they opened the chat. A passive engineer gets through none of those filters.

Ali hedged his own number. He warned against reading it as proof that candidates prefer AI to people, and said: "What we were offering before wasn't a human conversation either." That's the honest version of the finding. For applicants, the comparison was never AI versus a recruiter. It was AI versus an applicant tracking system that never wrote back.

The 0.1% leaves out roughly 180,000 quiet exits

The 0.1% only counts people who asked for a human out loud. It leaves out the much bigger group who simply walked away. Using Workable's own percentages, about 180,000 of the 350,000 people contacted either never opened the chat or dropped out partway.

Here is the math, worked from the published rates:

• About 350,000 candidates contacted.
• About 129,500 never opened the conversation (37%).
• About 220,500 opened it.
• About 50,700 opened it but didn't finish (23% of openers).
• About 169,800 finished.
• 222 asked for a human.

The 0.1% only works if you divide 222 by the roughly 220,000 who opened the chat. Divide by everyone contacted and you get closer to 0.06%. Workable's summary figures and these percentages don't line up perfectly, so treat the counts as approximations. The ratio holds either way.

~800 to 1

Asking for a human takes effort. You have to find the option, decide the bot is the problem, and say so. Leaving takes no effort at all. People who dislike a process rarely complain. They close the tab. Any metric that only counts complaints will make a process look better than it is, whether the sender is a bot or a person.

Why don't applicants and passive engineers behave the same way?

An applicant has already picked the employer, so an instant AI screen just saves them waiting. A passive engineer hasn't picked anything, so the first message is a test of whether the sender is credible. Speed doesn't help with that.

The inbox numbers show the gap. Pin's passive candidate saturation report found that roughly 68% of New York software engineers had heard from more than one employer. New York engineers replied at 6.3%, against 16.5% in Dallas-Fort Worth. HeroHunt's 2026 benchmarks put engineer reply rates at 4.64%, against 9.94% for product managers. Engineers reply at less than half the product manager rate, and in a crowded market like New York they reply at well under half the rate of a quieter one.

An applicant wants speed. A passive engineer wants to know whether you are real.

Small pools make every generic message expensive

In a small pool, a generic AI message doesn't lose you one candidate out of a hundred. It burns a real share of everyone you could ever hire for the role. Applicant funnels have plenty of spares. Senior engineering pools don't.

In Refolk's index of professional profiles, 215,025 people in the US currently hold Senior or Staff Software Engineer titles. In Germany the figure is 16,478, mostly in Berlin. Narrow the US group to a specific skill and it shrinks fast.

The US pool is about 13x the size of Germany's. Rust engineers make up 0.24% of the US Senior/Staff group.

522

Those 522 people work at companies like Google, Shopify, Uniswap, Saronic Technologies, Figure and Helius. Once one of them files your company under "bot spam," your hiring manager's personal follow-up starts at a discount. You can't re-run the experiment on the same person.

That's why hard-to-fill roles drag. You can't send in bulk, you can't keep retrying, and every wasted first message comes out of a pool that won't grow back this quarter.

"Double the response rate" compared to what?

Workable says its agent's personalized outreach gets double the response of standard templates. But doubling a generic 1 to 3% engineer reply rate only gets you to 2 to 6%, while signal-based messages reach 20 to 30%. The 0.1% announcement gave no outbound numbers at all.

Workable Agent also searches a database of more than 400 million profiles and runs outreach on its own. That's useful reach. Still, the benchmarks say what the reply rate depends on:

• Generic cold email to engineers: 1 to 3% response.
• Personalized, signal-based outreach (citing a recent commit, talk, or post): 20 to 30%.
• Codility reports a 30% response rate when reaching out to developers on GitHub.

The lift doesn't come from automation. It comes from proof that the sender looked at the person's work. An engineer can tell within one sentence whether you read their code. "Saw your PR adding backpressure to the ingest worker" is something they can check. "Your impressive background caught my eye" is something they can't.

Outbound is worth getting right. Gem's 2026 data shows direct sourcing delivers 11% of hires from just 2.6% of applications, about 4x the yield of inbound. That edge only lasts if the first message earns the reply.

The hard part is research, not sending. Finding the 522 and pulling the one commit worth mentioning for each of them is where the hours go. Refolk is built for that step. You describe the person in plain English and get back a ranked shortlist from GitHub, LinkedIn, and the open web, along with the signals you'd actually quote in line one.

Do candidates prefer AI interviews?

Applicants put up with AI interviews when the alternative is silence, and they often do well in them. Still, a large share say they distrust AI enough to walk away. Neither finding is about cold outbound, because both studies looked at people already inside a hiring process.

The evidence points in two directions:

• Behavior: Brian Jabarian (Chicago Booth) and Luca Henkel (Erasmus) found that candidates interviewed by a voice-AI agent instead of a human were 12% more likely to get a job offer and 18% more likely to start the job.
• Stated trust: Greenhouse's 2026 Candidate AI Interview Report, a survey of 2,950 job seekers, found 38% had abandoned a hiring process because AI was involved. Another 12% said they'd be willing to.

What candidates say and what they do don't match, and that's the real story. Inside a process they chose, structured AI conversations can be consistent and fast. But the 38% is a warning about first impressions. When AI shows up before the candidate has committed, a lot of people leave. A passive engineer hasn't committed to anything.

So the honest answer to "do candidates prefer AI interviews" is: applicants, sometimes, in practice. Passive candidates, nobody has measured. Neither study above looked at them.

Where should AI sit in outbound to passive engineers?

Let AI do the research and the drafting, and have a named person the engineer can look up sign and send the first message. That copies what actually kept Workable's opt-outs low: speed, plus a human available at any time.

A workflow that keeps the leverage and drops the risk:

1. Let AI build the list. Search GitHub, LinkedIn, and the open web by what people have built, not just their titles.
2. Let AI pull the signal. A recent commit, a conference talk, a post about a hard bug.
3. Draft with AI, edit as a human. One specific line about their work, one line about the role, one clear ask.
4. Send from a named person. The hiring manager or a recruiter with a real, checkable profile.
5. Make it checkable. Link the team page, the public repo, or the posted job.
6. Hand off fast. Workable's outreach went out within minutes. Match that speed once someone replies, and let AI handle scheduling.
7. Measure the silence. Track non-replies and unsubscribes, not just complaints. Workable's ~800-to-1 ratio is the reason.

Where AI-run conversations do work well:

• Screening inbound applicants who already picked you.
• Scheduling and logistics after a passive candidate replies.
• Answering process questions about stages, location, or comp bands.

Where a human-signed first message still decides the outcome:

• First contact with senior engineers in small pools, like the 522 US Rust engineers.
• Crowded markets like New York, where most engineers already have competing pitches.
• Any market as thin as Germany's 16,478 Senior/Staff engineers, where retries run out fast.

FAQ

Does Workable's 0.1% stat mean candidates prefer AI recruiters?

No, and Workable says so itself. CRO Sarfraz Ali warned against reading the result as a preference for AI, noting that "what we were offering before wasn't a human conversation either." The 222 handoff requests came from applicants who had already opened an AI chat. About 180,000 others never opened it or didn't finish, and the stat doesn't count them.

Can an AI recruiting agent handle outbound to passive engineers?

It can handle the research and drafting well. The first message works better coming from a named person. Generic cold outreach to engineers gets 1 to 3% response, and signal-based messages that cite real work reach 20 to 30%. The jump comes from specificity and a sender the engineer can check, not from automation, so let AI build the research and let a human send.

What reply rate should I expect from engineers on cold outreach?

Expect less than for other roles. HeroHunt's 2026 benchmarks put engineers at 4.64%, against 9.94% for product managers. Location matters too: Pin found New York engineers reply at 6.3% versus 16.5% in Dallas-Fort Worth. Personalized messages tied to a recent commit or talk can lift that to 20 to 30%, and Codility reports 30% on GitHub outreach.