The Sourcing Machine: How AI Agent Pipelines Deliver 49% Reply Rates and 700 Hires in 90 Days
By Tim Kreling, Co-Founder, OVI
The Sourcing Machine: How AI Agent Pipelines Deliver 49% Reply Rates and 700 Hires in 90 Days
Celestica had roles open for more than a year across three countries. Within 90 days of deploying an AI-driven sourcing pipeline, the electronics manufacturer filled 700 positions, achieved a 49% reply rate on outreach sequences, and saw 35% of contacted candidates express active interest (Gem / Celestica case study). Roles that had languished unfilled for 12 months closed in 30 days. The TA team moved from hoping candidates would respond to predicting exactly when a role would be filled — "we'll have this role filled in 40 days" became a standard internal forecast.
This is what happens when sourcing stops being an art and starts being an engineered system.
Industry-wide, the shift is accelerating. Fifty-one percent of organizations now use AI in recruiting, up from 26% just a year earlier (HeroHunt AI adoption review, 2025). But the companies pulling ahead are not simply bolting AI onto existing workflows. They are rebuilding sourcing as a product — with funnel conversion metrics, multi-channel automated sequences, and measurable reply-rate targets that would look at home in a growth engineering standup.
Scale AI: Reverse-Engineering the Funnel From a Deadline
When Scale AI needed to fill 12 engineering roles in three weeks, the TA team did not start with job postings. They started with math. Working backward from the deadline, they mapped required pass-through rates at each funnel stage — how many candidates needed to enter the top of the pipeline to produce 12 hires at the bottom, given historical conversion data (Gem / Scale AI case study).
The result was a sourcing pipeline built like a production system. Every stage had a target conversion rate. Every conversion rate was measured in real time.
The most consequential decision was not a new tool — it was reactivating "silver medalists," strong candidates from previous searches who had narrowly missed the cut. Seventy percent of Scale AI's hires during that sprint came from the existing sourced pipeline rather than net-new outreach (Gem / Scale AI case study). The team was, in the words of a Scale AI Director, "building a machine not only for volume but also for velocity" (Gem / Scale AI case study).
The lesson is structural: most TA teams treat their sourced-but-not-hired candidates as dead inventory. Scale AI treated them as the highest-conversion segment in the pipeline — because they already had signal on fit.
This approach echoes what sourcing benchmarks consistently show. Outbound-sourced candidates are five times more likely to be hired than inbound applicants, with a 67% passthrough rate versus 8% for inbound (Pin sourcing benchmarks 2026). The math favors any team that can systematize outbound outreach and candidate reactivation.
LinkedIn Hiring Assistant: The Supervisor-Agent Architecture
LinkedIn took a different architectural approach to the same problem. Its Hiring Assistant, built on a supervisor-plus-sub-agent architecture, automates sourcing while recruiters sleep. A supervisor agent orchestrates specialized sub-agents — one to qualify candidates, another to draft InMail messages, another to manage pipeline sequencing (LinkedIn Engineering blog; InfoQ LinkedIn agent deep-dive).
The results are measurable at every stage. Recruiters review 62% fewer profiles to reach the same number of qualified candidates. InMail acceptance rates improve by 69%. Time-to-fill drops by 30%. LinkedIn estimates recruiters save more than four hours per open role (LinkedIn Engineering blog).
The product became globally available in September 2025 and contributed to $450 million in annual recurring revenue by LinkedIn's fiscal year 2026 Q3 (LeadDev — LinkedIn Hiring Assistant global rollout). That revenue figure signals something important for TA leaders evaluating the category: AI sourcing is not an experiment at the margins. It is a product line generating enterprise-grade commercial traction.
The supervisor-agent pattern also illustrates where the technology is heading. Rather than a single monolithic AI making sourcing decisions, LinkedIn decomposes the workflow into discrete, auditable steps — each handled by a specialized sub-agent. This architecture makes it easier to isolate failures, tune individual components, and maintain human oversight at critical decision points.
The Benchmark Gap: What "Good" Looks Like in AI-Powered Outreach
Industry data now makes it possible to quantify the gap between AI-powered and manual sourcing sequences.
According to Ashby's Talent Trends Report, AI-personalized outreach campaigns achieve a 35.3% reply rate compared to 24.1% for non-AI campaigns — a 46% lift (Ashby Talent Trends Report). The average campaign reply rate across all companies in Ashby's dataset sits at 19.6%, meaning AI campaigns outperform the market average by nearly 80%.
Sequence length matters, but with diminishing returns. Three-email sequences produce the best reply rates, plateauing at roughly 23%. Approximately 50% of all replies come from the first email alone (Ashby Talent Trends Report). Four-step sequences can squeeze out twice as many replies as single-touch outreach (HeroHunt AI adoption review, 2025), but beyond four touches the marginal gains collapse.
For context, the baseline cold email reply rate across industries is 3.43% (Pin sourcing benchmarks 2026). Teams running AI-personalized multi-touch sequences are operating at 10x that baseline. The gap between "we send InMails" and "we run an engineered sourcing pipeline" has never been wider.
These numbers matter for forecasting. Amplify Partners' recruiting funnel framework identifies three metrics that determine pipeline health: volume in, conversion at each stage, and velocity through each stage (Amplify Partners — debugging your recruiting funnel). AI sourcing agents improve all three simultaneously — higher volume through automated outreach, higher conversion through personalization, and higher velocity through automated follow-up sequencing.
Making Sourcing-as-Product Accessible
The Scale AI and LinkedIn examples share a common thread: both invested heavily in internal TA engineering infrastructure. Scale AI built a bespoke funnel-modeling practice. LinkedIn built a multi-agent orchestration layer with dedicated engineering teams.
Most mid-market companies cannot replicate that investment. But the sourcing-as-product philosophy — systematic multi-channel outreach, automated follow-up sequences, reply-rate tracking as a core KPI — is now available as off-the-shelf infrastructure.
OVI's Sora agent, for example, operationalizes this pipeline for teams that lack a dedicated TA engineering function. Sora runs systematic outreach across LinkedIn and email, automates follow-up sequencing until a candidate replies, and surfaces reply-rate data so recruiters can tune campaigns the same way a growth team optimizes conversion funnels — starting at $99/month on the Starter plan (ovi-me.com). It is the same sourcing-as-product pattern, packaged for companies hiring 10 roles, not 1,000.
Frequently Asked Questions
What reply rate should we target for AI-powered sourcing campaigns?
Aim for 25–35%. Ashby's data shows AI-personalized campaigns average 35.3%, but even reaching the mid-20s puts your team well above the industry average of 19.6% and far beyond the 3.43% cold email baseline. Track reply rates by role type and seniority, since technical and executive roles typically run lower than operational roles.
How many emails should be in a sourcing sequence?
Three is the sweet spot. Ashby's data shows three-email sequences produce peak reply rates around 23%, and roughly half of all replies arrive after the first email. Going to four touches can improve total volume, but beyond that the marginal returns diminish sharply. Focus energy on making the first email strong rather than adding more follow-ups.
Should we prioritize re-engaging past candidates over sourcing new ones?
Yes — for roles where you have a historical pipeline. Scale AI sourced 70% of hires from previously identified candidates ("silver medalists") rather than new outreach. Outbound-sourced candidates convert at 5x the rate of inbound applicants. Re-engagement campaigns are typically cheaper, faster, and higher-converting than net-new sourcing.
What does a supervisor-agent architecture mean for AI sourcing?
It means the AI system is not a single black box. A supervisor agent coordinates specialized sub-agents — one for candidate qualification, one for message drafting, one for pipeline sequencing. This makes each step auditable and tunable independently. LinkedIn's implementation of this pattern reduced profiles reviewed by 62% while improving InMail acceptance by 69%.
How do we measure ROI on an AI sourcing pipeline?
Track three metrics across your funnel: volume entering the pipeline, conversion rate at each stage, and velocity through each stage. Compare pre- and post-deployment baselines for time-to-fill, reply rate, and cost per qualified candidate. Celestica's benchmark — roles open 12+ months filled in 30 days — illustrates the velocity impact. For cost, compare your per-candidate outreach cost against the 3–10x reply-rate improvements AI sequencing delivers over manual outreach.
What reply rate should we target for AI-powered sourcing campaigns?
Aim for 25–35%. Ashby's data shows AI-personalized campaigns average 35.3%, but even reaching the mid-20s puts your team well above the industry average of 19.6% and far beyond the 3.43% cold email baseline. Track reply rates by role type and seniority, since technical and executive roles typically run lower than operational roles.
How many emails should be in a sourcing sequence?
Three is the sweet spot. Ashby's data shows three-email sequences produce peak reply rates around 23%, and roughly half of all replies arrive after the first email. Going to four touches can improve total volume, but beyond that the marginal returns diminish sharply. Focus energy on making the first email strong rather than adding more follow-ups.
Should we prioritize re-engaging past candidates over sourcing new ones?
Yes — for roles where you have a historical pipeline. Scale AI sourced 70% of hires from previously identified candidates ("silver medalists") rather than new outreach. Outbound-sourced candidates convert at 5x the rate of inbound applicants. Re-engagement campaigns are typically cheaper, faster, and higher-converting than net-new sourcing.
What does a supervisor-agent architecture mean for AI sourcing?
It means the AI system is not a single black box. A supervisor agent coordinates specialized sub-agents — one for candidate qualification, one for message drafting, one for pipeline sequencing. This makes each step auditable and tunable independently. LinkedIn's implementation of this pattern reduced profiles reviewed by 62% while improving InMail acceptance by 69%.
How do we measure ROI on an AI sourcing pipeline?
Track three metrics across your funnel: volume entering the pipeline, conversion rate at each stage, and velocity through each stage. Compare pre- and post-deployment baselines for time-to-fill, reply rate, and cost per qualified candidate. Celestica's benchmark — roles open 12+ months filled in 30 days — illustrates the velocity impact. For cost, compare your per-candidate outreach cost against the 3–10x reply-rate improvements AI sequencing delivers over manual outreach.