At a recent webinar, the ElevenLabs talent team walked through the exact sourcing system they use to hire designers: the framework, the prompts, the tools and the actual numbers. This is my breakdown of what they shared, step by step, with my own comments as a technical sourcer.
Disclaimer: this article is based on information shared publicly at a PromptMates Live session on YouTube, plus my own interpretation. It contains no confidential information, is not sponsored, and is not affiliated with or endorsed by ElevenLabs.
Some context. ElevenLabs is one of the leading AI voice companies, valued at $11B after its $500M Series D in February 2026. It hires in one of the most competitive talent markets in the world. When you compete for top designers, engineers and researchers against the best-funded companies in AI, inbound alone doesn't cut it. You have to find the people you want before anyone else does.
If you are hiring for high-bar roles — or you lead a recruiting team that wants to bring AI into sourcing without starting from scratch — this breakdown is for you.
The math tells you everything
Over the past year, ElevenLabs scaled its design team from 8 to 35 people. Their sourcing data:
- ~30,000 inbound applications → 1 hire
- ~5,000 sourced candidates → 16 hires
- A maintained talent pool of ~300 high-quality designers
For specialised roles, inbound produced high volume and low signal. Sourcing took more effort upfront but converted at 0.32%, versus about 0.003% for inbound — nearly 100x better.
Having hired product designers myself, I can tell you this is one of the hardest roles to fill. There's no GitHub profile to analyse and no coding test to run. You can't write a script to decide whether someone's work is good. You need to know where designers show their work — Behance, Dribbble, agency portfolios — and you need product taste and judgment to tell good work from great work.
That's exactly why this design sourcing framework is worth studying.
Why start with design
Design was a deliberate choice. If a sourcing system works for a role where evaluation is visual and subjective, it will work for engineering and research too. For roles like these, posting a job and waiting isn't enough. You build a system.
Build vs buy: the first decision
Building sourcing automation takes time, and most recruiting teams are already at capacity. According to the ElevenLabs team, a custom sourcing system makes sense in three situations:
- A niche role with a high talent bar.
- Volume hiring with five or more similar roles.
- You want to reach candidates your competitors haven't found yet.
In most other cases, your existing recruiting stack is enough. They also flagged the trade-off: more automation means more complexity. Security, compliance and access management become real issues once a tool goes beyond one person's laptop.
The framework: a 7-step sourcing flywheel
The core idea is a self-reinforcing loop. The more high-quality profiles you find, the better your signals for finding the next ones:
Signals → Enrichment → Classification → Evaluation → Human in the loop → Talent pool → Outreach
Each step feeds the next, and every cycle makes the following search smarter.
Step 1: Signals
The team doesn't start with a search. They start with a question: where do the best people show up? They use four types of signals:
- Lookalike searches based on people already identified as high-potential.
- Network discovery — great people know great people, so map their connections.
- Social engagement — followers, comments, reshares.
- Expert recognition — awards, conferences, podcasts, newsletters, credits.
Signals differ by role. For designers: Behance, Dribbble, design awards and names credited in agency portfolios. For software engineers: GitHub activity, open-source contributions, ICPC participants. For ML engineers and researchers: arXiv papers and NeurIPS talks.
Define the signal first. The more high-signal data you have about your target talent pool, the easier everything downstream becomes.
One prompt they use to brainstorm signals:
"You're a talent engineer. Do deep research. Build me a comprehensive list of ideas to gather signals online for strong designers — e.g., awards jury members, podcast speakers, newsletter contributors. Be specific about platforms. Return in a table: type, source, signal strength prediction, implementation difficulty."
One workflow I loved: pick a target design agency, pull the names credited in its portfolio pages, enrich each name with LinkedIn, social profiles and location, and export it as a TSV file. This surfaces the individual designers behind the work, not just the agency brand.
Step 2: Enrichment
Once a profile is identified, it gets enriched. For designers: portfolio, LinkedIn, Instagram, Dribbble, Behance and follower counts. For engineers: GitHub activity, personal site and tech stack. Enrichment gives you a picture of a person's real work and contributions, not just their job title.
Step 3: Classification
Each enriched profile is classified across four dimensions:
- Department — product, design, engineering, marketing.
- Relevance — does this person map to an open role?
- Seniority — junior, senior, staff, director.
- Category — the specific type of designer or engineer.
This is a perfect job for an LLM: structured, repetitive and easy to check.
Step 4: Evaluation
Three levels of candidate evaluation, depending on role complexity and resources:
- Level 1 — simple rules. Individual contributor? Staff level or above? From a target company?
- Level 2 — combined signals. Online presence, portfolio quality, engagement patterns.
- Level 3 — a model trained on past hiring decisions that finds correlations in historical outcomes.
Automation can handle signals, enrichment, classification and scoring. But the moment you need to look at someone's work and decide whether it's actually good, that's still on you. For design roles, portfolio review stays firmly human. No model today reliably tells technically correct work from great work.
The rule is simple: automate everything that leads to the judgment call. Never automate the judgment itself.
Step 5: Human in the loop
Automation narrows the pool; people make the decisions. The talent team reviews the shortlist and about 20% pass. Hiring managers then approve 10–20% of those. The interface? A Google Sheet — intentionally simple.
Step 6: The talent pool — your gold mine
Every strong candidate who isn't a fit for a current role is saved and revisited every three months.
At the start of Q4, ElevenLabs had about 300 top designers in its pool and 17 new openings. They reached out to all 300 in roughly 8 hours. The result: 25 interviews booked, with 5–7 hires expected from that cycle.
Approved candidates also become seed data for the next lookalike search. That's how the pipeline compounds — the flywheel effect.
My note here: great sourcing isn't about blasting candidates with messages. It's about long-term relationships. Someone who isn't looking today may be your next hire in six months, so stay on their radar. ElevenLabs clearly does this very well.
Step 7: Outreach
The channels are standard: LinkedIn Recruiter, email sequences and warm introductions.
What stands out is how they use their own product. Candidates get a link to an ElevenLabs voice agent that describes the role — so the outreach doubles as a live product demo. The recruiting tool is the product. Few companies think of that kind of alignment.
Where to start with AI sourcing
The hardest part isn't building the system. It's adoption.
Most recruiting teams already run full req loads, manage stakeholders and close roles. An AI sourcing layer needs time to build, test and maintain — and most teams don't have that bandwidth.
The second challenge is knowing where to start. There's no shortage of AI recruiting tools, but a tool is not a system. Turning them into a repeatable sourcing workflow — consistent outputs, one source of truth, a feedback loop that improves over time — is the real work. What ElevenLabs built isn't a set of prompts. It's compounding infrastructure.
Start small, think big:
- Pick two or three signals for your role that go beyond LinkedIn keywords.
- Build a shortlist of 20–50 high-signal people.
- Use an LLM to classify them.
- Save the strong ones who aren't a fit yet, and revisit them in three months.
Automate one step, then the next. Building a sourcing system isn't about budget. It's about deciding that your next search will be better than the last one.
FAQ: AI sourcing and designer hiring
How do you source product designers?
Start from where designers show their work — Behance, Dribbble, design awards, agency portfolio credits, conference talks and newsletters — then enrich each profile with LinkedIn and portfolio links and review the work yourself.
Is sourcing better than inbound for specialised roles?
In ElevenLabs' case, yes by a wide margin: about 30,000 inbound applications led to 1 design hire, while about 5,000 sourced candidates led to 16.
What is a sourcing flywheel?
A self-reinforcing sourcing loop — signals, enrichment, classification, evaluation, human review, talent pool and outreach — where every strong profile you find improves the next search.
What should AI automate in recruiting, and what shouldn't it?
Automate signal gathering, enrichment, classification and scoring. Keep the judgment — portfolio review, final evaluation and relationships — human.
How often should you re-engage a talent pool?
ElevenLabs revisits strong candidates every three months, which let them reach 300 designers in about 8 hours and book 25 interviews.
The takeaway
ElevenLabs didn't win design hiring with more job posts. They won it with a system: clear signals, automated research, human judgment and a talent pool that gets better every quarter. You don't need their budget to start — just one step you automate this month.