Diversity sourcing is unusually full of both theatre and genuine practice, and the difference matters legally as well as operationally. This guide separates the tactics with a defensible mechanism from the ones that mostly generate reporting, and is explicit about where legal review is required. Nothing here is legal advice — the rules differ sharply by jurisdiction and change frequently.
The framing we find most useful: diversity sourcing is not a separate motion bolted onto sourcing. It is sourcing done with the funnel-narrowing steps examined. Most of those steps happen before anyone is contacted.
Where pipelines actually narrow
| Stage | How it narrows | Fix |
|---|---|---|
| Requirements | Degree requirements, "top-tier company" preferences, years-of-experience inflation, unnecessary credentials | Strip to what predicts performance; test each requirement against your actual successful hires |
| Target companies | A list of 15 employers that look like your current team's employers | Deliberately add adjacent industries, non-elite employers, and different geographies |
| Search query | Keywords and titles that encode one career path; AI ranking that rewards similarity to past hires | Multiple query formulations; check who appears on page 3 |
| Outreach | Messages written for one audience's motivations | Concrete detail about the work; comp range; avoid culture-coded language |
| Screening | Resume proxies (school, employer brand) standing in for capability | Structured criteria applied consistently, scored before the debrief |
The first three rows are sourcing's responsibility and they compound: a target list drawn from your own alumni network produces a candidate set that resembles your current team before a single message is sent. This is not a tool problem, and buying a "diversity sourcing" product does not address it.
Tactics with a defensible mechanism
Audit requirements against evidence. Take your last ten successful hires in a role and check which stated requirements they actually had on entry. Degree requirements and inflated experience minimums are the usual casualties, and removing them widens the qualified population immediately — with no legal exposure, because you are broadening rather than selecting.
Widen the target-company list deliberately. Add employers your team wouldn't naturally list: adjacent industries, non-coastal geographies, companies without prestige brands, public sector, contractors. This is the single highest-leverage change most teams can make, and it is free.
Use multiple query formulations. Different titles for the same job, skills instead of titles, and non-linear career paths (career changers, bootcamp and apprenticeship routes, military transitions). See Boolean vs AI search for why one query is always a narrow sample of the market.
Source from your own database. Past applicants and silver medallists are a population you already attracted, and rediscovery often surfaces people who were screened out for proxy reasons rather than capability. Cheap, and covered in our rediscovery guide.
Publish comp ranges. Ranges reduce the negotiation-driven pay gaps that later become retention problems, and they raise reply rates. Increasingly mandatory anyway.
Structure the screen. The largest measurable drop-offs in most funnels are at resume screen and first interview, not at sourcing. Structured, pre-agreed criteria scored independently before the debrief is the intervention with the strongest research support in hiring generally.
Tactics that are mostly theatre
Diverse slate requirements without pipeline work. Mandating that every shortlist include underrepresented candidates without changing requirements or targeting produces either delay or tokenism — and, in some jurisdictions, legal exposure if the slate is constructed by selecting on protected traits.
Job-board spend as a strategy. Posting to diversity-focused boards is fine and reaches actively looking candidates. It does not address the passive market, which is where most of the qualified population sits.
Blind resume screening as a whole solution. Name-blinding has mixed evidence and doesn't touch the school and employer proxies that do most of the filtering. Structure the criteria instead of hiding the name.
Vendor "bias-free AI" claims. No vendor can make a model bias-free; they can at best document what it was trained on, what it optimises, and what they audit. Ask for the audit, not the adjective.
The specific risk in AI sourcing tools
Two distinct problems, often conflated.
Invisible narrowing. Semantic ranking scores profiles by similarity to a brief — and in some implementations, similarity to people you've hired or graded positively. That is a mechanism for reproducing your existing composition at scale, and you can't see it happening: you see 25 plausible profiles, not the 300 the ranking buried. Mitigation is procedural rather than technical — inspect who appears beyond the first page, compare composition of the search output against the addressable market, and vary the query.
Regulated automated decisions. Automated screening against a hiring bar is squarely within the scope of rules such as NYC Local Law 144 (bias audits and candidate notice for automated employment decision tools) and the EU AI Act's provisions on employment-related AI. Illinois, Colorado, and others have their own regimes. Practical questions for any agent or AI-screening vendor: what does it log, what does a documented human review step look like, has it been independently audited, and who is the controller for candidate data? Have counsel review before you scale.
A blunt point on tool features: filtering or scoring candidates on inferred gender, ethnicity, or age is a different act from broadening a search, and in many jurisdictions it is unlawful in hiring. Some tools offer inferred-demographic filters. Whether you may use them, and for what, is a legal question specific to your jurisdiction — and the safe use case is aggregate measurement of your pipeline, not selection of individuals.
What to measure
Measure composition at each stage, in aggregate, and look at pass-through rates rather than absolute counts:
| Measurement | Question it answers |
|---|---|
| Composition of the sourced list vs the addressable market | Is our targeting narrower than the market? |
| Reply rate by group | Is our outreach landing differently across audiences? |
| Screen-to-interview pass-through by group | Are we losing people at the resume screen? |
| Interview-to-offer pass-through by group | Is the loop itself the constraint? |
| Offer-accept rate by group | Is comp or candidate experience the problem? |
Two disciplines make this legitimate rather than risky: use aggregate, voluntarily self-reported data separated from individual hiring decisions, and act on the stage with the biggest differential rather than adding top-of-funnel volume reflexively. A pipeline that is diverse at sourcing and homogeneous at offer is not a sourcing problem, and pouring more candidates into a leaky screen wastes both their time and yours.
Where tools help, honestly
Tool support for this work is mostly indirect, and we'd be sceptical of anything marketed primarily as a diversity product. What genuinely helps: broad, multi-source indexes that reach beyond one professional network (SeekOut and hireEZ both market diversity features; treat the aggregate reporting as the useful part and the individual filters as a legal question), skills-based search that de-emphasises employer prestige (Juicebox, Findem), and specialist indexes that surface people outside the usual pipelines (AmazingHiring for engineers with public work but non-traditional CVs).
The highest-leverage changes remain free: fix the requirements, widen the target list, structure the screen, publish the range. No product in our 2026 ranking substitutes for those, and any vendor claiming otherwise is selling you reporting. How we weigh vendor claims of this kind is set out in our methodology.