AI Candidate Sourcing: Scorecards and Outreach That Work

Guide · August 2026

AI candidate sourcing is how talent teams find, rank, and open conversations with people who are not already in the apply pile. Used well, it expands passive reach, speeds first contact, and keeps criteria explicit. Used poorly, it sprays generic messages, overweights pedigree proxies, and burns employer brand. This guide covers hybrid search, the Sourcing Fit Scorecard, 3-touch and 5-touch outreach, channels, tracking fields, privacy high-level notes, free vs paid stacks, passive vs active talent, failure modes, metrics, and FAQs. Pair it with Free AI Recruiting on i10X and the broader AI recruiting guide.

---

What is AI candidate sourcing?

AI candidate sourcing is the use of language models, matching systems, and search tools to identify people who may fit a role, draft personalized outreach, and keep sequences organized, while humans still decide who to contact and how hard to pursue. It covers passive talent (not actively applying), reactivation of warm past applicants, and smarter use of employee or community referrals.

Sourcing is different from AI resume screening. Screening ranks people who already applied. Sourcing builds the top of the funnel before or alongside inbound applications. Many teams mix both: AI suggests lookalikes and Boolean strings, then drafts first-touch messages that a recruiter edits and sends.

In 2026, AI candidate sourcing usually means a combination of:

• Search assistance: Boolean strings, semantic queries, and “people like this profile” suggestions.
• Fit ranking: scoring public or CRM profiles against a scorecard, not a vague job ad dump.
• Message generation: short, specific outreach and follow-ups grounded in real profile signals.
• Workflow memory: who was contacted, what they said, and when to stop (often via ATS, CRM, or an agent).

For multi-step automation that chains source, screen, and shortlist prep, see AI recruiting agents and the end-to-end AI recruiting workflow map in this series. Ethical constraints on ranking and notice live in ethical AI recruiting.

---

Why AI sourcing matters now

Adoption is rising across HR, not only screening. According to SHRM’s 2025 Talent Trends research, 43% of organizations used AI in HR tasks in 2025, up from 26% in 2024. LinkedIn’s Future of Recruiting 2025 report found that 37% of organizations were actively integrating or experimenting with generative AI in hiring, up from 27%, and that talent acquisition professionals using gen AI reported saving about 20% of their workweek on average.

That time usually lands on repetitive drafting and first-pass research, not on replacing judgment. LinkedIn has also reported that companies whose recruiters use AI-Assisted Messaging most are about 9% more likely to make a quality hire than those who use it least. Personalized, relevant outreach still matters; AI helps more people write it consistently when the inputs are real.

Speed pressure is real too. SHRM’s 2025 Recruiting Benchmarking cycle has been widely reported with a median time-to-fill around 44 days for nonexecutive roles, while SHRM’s 2026 Recruiting Executives Benchmarking reports a median time-to-fill of 39 calendar days for nonexecutive roles. Cost pressure sits beside speed: SHRM 2025 Benchmarking Report averages (as reported in press coverage) put nonexecutive cost-per-hire around $5,475 and executive around $35,879 (averages; other SHRM releases use medians that can differ). Sourcing that fills critical seats faster can move those numbers, but only if quality and fairness hold.

AI sourcing fails as a spray cannon. It works when fit is written first, proxies are banned, and contact, long sequences, and DNC stay human-owned.

---

Passive vs active candidates (and why AI treats them differently)

Active candidates already apply or update profiles; risk is volume and fair screening. Passive candidates need a reason to reply (specific problem, credible story, easy no); default shorter sequences and stricter caps. Warm reactivation (prior applicants, silver medalists) often beats pure cold; check DNC and recontact rules before AI drafts.

Segment	Primary goal	AI role	Human gate	Typical sequence
Active inbound	Fair, fast shortlist	Screen notes vs scorecard	Reject / advance	Application flow, not cold sequence
Passive cold	Earn a reply without brand damage	Find + rank + draft	Contact list + send edit	3-touch default
CRM reactivation	Restart a known conversation	Summarize past stage + draft	Eligibility check	1-2 touches then stop
Referral assist	Help employees introduce well	Suggest who + why in plain language	Employee sends intro	One intro + recruiter follow

---

Sourcing Fit Scorecard (no demographic proxies)

Primary backlink asset: a Sourcing Fit Scorecard for outbound ranking (not protected-class inference). Score 0-2 per dimension; multiply by weight. Any must-have at 0 means do not source this week unless a human logs an override. Publish weights on the requisition so sourcers score the same way.

Dimension	Weight	0 (fail)	1 (partial)	2 (strong)	Banned proxies (never score)
Core craft match	5	Wrong function or no relevant craft	Adjacent craft, weak evidence	Direct craft with recent evidence	School name prestige as a substitute for craft
Level and scope	4	Clearly under or over level with no path	Close level, unclear ownership	Scope matches role (IC/lead/manager)	Age, graduation year as age proxy
Domain or product context	3	No transferable domain	Partial domain overlap	Relevant product, industry, or user type	Employer “brand coolness” as a proxy for skill
Stack or method evidence	4	Missing required tools or methods	Related tools only	Required stack or methods shown in work	Hobby lists or social interests
Location / work model fit	3	Hard conflict with true constraint	Unclear or needs confirmation	Meets stated location or remote rules	Neighborhood prestige or commute stereotypes
Work authorization (if truly required)	5 (gate)	Known hard fail for the role	Unknown (flag, do not invent)	Clear match to posted requirement	Name, accent, or nationality guesses
Signal of openness (optional)	2	Public “not open” and no warm path	Neutral / unknown	Open to work, recent job change, or mutual path	Family status, photos, marital signals
Evidence quality	3	Claims only, no artifacts	Titles only, light detail	Projects, outcomes, or public work you can cite	Photo quality, “energy,” or writing polish alone

Operator notes: cite artifacts for craft; map level to ownership (not title inflation); lower domain weight when transferability is real; prove stack from work not headlines; use location/authorization only when truly required. Paste the scorecard before Boolean generation ( free tools 2026, Free AI Recruiting).

---

Boolean vs semantic vs hybrid search (deep decision table)

Sourcers still need Boolean. Models still invent confident matches. Hybrid is the default for most professional roles in 2026.

Dimension	Boolean	Semantic / LLM assist	Hybrid (recommended default)
What it optimizes	Precision, reproducibility, audit strings	Recall across titles, narrative fit, ranking notes	Precision gate + narrative rank inside a capped set
Best when	Exact skill, license, language, cert, stack token	Titles vary, career paths are non-linear, lookalikes	Almost all mid and senior professional searches
Weak when	Sparse profiles, synonym-heavy crafts	Compliance needs string replay; thin text over-inference	Team refuses to log either string or prompt version
Must-have skill is exact	Primary filter	Only expand synonyms after Boolean	Boolean gate, then semantic ranking inside the set
Role language is messy	Secondary (title OR clusters)	Primary for title expansion and lookalikes	LLM proposes title OR list; human locks it; Boolean runs it
Compliance or audit trail	Strong (reproducible strings)	Weaker unless you log prompts and outputs	Store both: string version + model notes
Passive talent, sparse profiles	Can over-filter empty profiles	Can over-infer from thin text	Loose Boolean, strict human sample of top N
High-volume commodity role	Fast volume filter	Helpful for ranking, risky for auto-contact	Boolean volume, scorecard rank, human send
Niche or senior search	Precision strings for stack	Narrative fit and career-path reasoning	Small Boolean set, deep semantic notes, human shortlist
Typical failure	False negatives from synonym miss	Confident wrong matches; pedigree flavor	Hybrid without human gate becomes automated spam
When NOT to use	As only method when titles are chaotic	As sole auto-contact engine	When no scorecard exists yet (fix intake first)

Practical hybrid workflow (15 minutes per req).

1. Lock the Sourcing Fit Scorecard with the hiring manager (must-haves first).
2. Ask the model for three Boolean variants: tight, medium, wide. Edit for accuracy. Never paste unedited strings into a mass action.
3. Run Boolean on your licensed sources. Export a capped list (for example top 50 to 100).
4. Ask the model to rank that list against the scorecard with evidence bullets only. Ban demographic inference in the system prompt.
5. Human selects the contact list for the week. AI drafts messages. Human edits and sends.

This hybrid keeps reproducibility (Boolean) and narrative understanding (semantic) without handing the send button to a black box.

---

Channels playbook: LinkedIn, GitHub, portfolios, boards, CRM, referrals

AI multiplies what licensed channels allow. Choose channels by craft and evidence density, not tool fashion.

Channel	Best for	AI assist	Human gate	Watch-outs
LinkedIn	Most professional roles; title clustering	Boolean variants, lookalikes, InMail drafts	Weekly contact list; edit T1	Platform terms; InMail fatigue; prestige bias in headlines
GitHub / code hosts	Engineering, data, infra with public code	Summarize repos vs stack must-haves; rank by evidence	Verify license and real ownership of code	Activity ≠ job readiness; ignore stars as prestige proxy
Portfolios / personal sites	Design, product, content, research	Extract case outcomes into scorecard rows	Confirm role and level from case depth	Scraping ethics; do not paste private contact data into consumer chat
Job boards / talent pools	Active seekers; niche boards	Rank applicants and saved searches	Fair screen same as inbound	Do not double-message people already in ATS mid-process
CRM reactivation	Silver medalists; past finalists	Summarize last stage, reasons, reopen draft	Check DNC and recontact rules	Stale notes; prior “no” still counts
Employee referral AI assist	Warm intros with context	Suggest who might fit and why in plain language for the employee	Employee owns the ask; recruiter follows	Do not pressure employees to spam their network

GitHub: Prefer ownership signals (sustained commits, design docs) over stars. Cite real modules in T1; humans verify authorship. CRM reactivation: sort by last positive stage; open with what changed; ask permission; one thoughtful note beats a long sequence.

---

3-Touch Outreach Sequence (template with example copy)

The 3-Touch Outreach Sequence is the cold-passive default: three value-bearing touches, then stop or switch to a warm path (referral, event, mutual intro).

Touch	Timing	Goal	Must include	Must avoid
T1: Specific open	Day 0	Earn a reply or soft no	One real profile signal, one role outcome, clear ask (15-min chat or “not now” is fine)	Flattery about prestige, fake “I came across your profile” with zero detail, salary bait if you cannot stand behind it
T2: Value follow-up	Day 4-6	Add new information	One new fact (team problem, stack, impact, interview process honesty) plus easy exit	Guilt language, “just bumping this,” repeating T1 word for word
T3: Graceful close	Day 10-12	Close loop, leave door open	Permission to re-contact later, optional referral ask, thanks	Fake deadline pressure, third-party spam CC

T1 skeleton (edit heavily):
“Hi [Name], I noticed [specific project, talk, open-source, or product outcome]. We are hiring a [role] to [one outcome in plain language]. Your [skill/evidence] maps to [scorecard line]. Open to a short conversation, or a simple no if timing is wrong?”

T2 skeleton:
“Quick add: the team’s main constraint right now is [problem]. Interview process is [stages]. If useful, happy to share the scorecard we use so you can self-select. Either way, appreciate your time.”

T3 skeleton:
“Last note from me on this role. I’ll close the loop on my side. If you ever want a look when [condition], or know someone who thrives on [problem], I’m easy to reach. Thanks again.”

---

5-Touch Outreach Sequence (when longer is justified)

Use 5 touches only for hard-to-fill roles when each touch adds value. Not a default. Frequency caps still apply.

Touch	Timing	Content job	Example angle
T1	Day 0	Specific open	Cite one real project; one outcome; soft ask
T2	Day 4	Problem depth	One technical or business constraint the hire will own
T3	Day 9	Process honesty	Stages, who they meet, decision speed you can keep
T4	Day 16	Social proof without prestige worship	Team craft story, not “top company” flattery
T5	Day 24	Close + referral	Stop cold sequence; invite future ping or referral

When NOT to run 5 touches: fragile brand; hard no already; spam complaints; legal/frequency limits; weak scorecard fit (volume is not a substitute for fit).

---

Copy-paste template: Sourcing Prompt Pack

Approved workspace only. Keep the ban list intact.

System intent:
“You assist with candidate sourcing. Score only against the provided Sourcing Fit Scorecard. Quote public evidence. If evidence is missing, mark unknown. Do not infer race, gender, age, disability, religion, family status, or nationality from names or photos. Do not use school prestige or employer brand coolness as skill. Do not invent contact history. Output structured notes only.”

User packet:
“Role: [title]. Scorecard v[N]: [paste]. Must-haves: [list]. Non-criteria: [list]. Channel: [LinkedIn/GitHub/CRM]. Task: (1) propose tight/medium/wide Boolean, (2) rank the pasted profiles 1-N with scorecard rows and evidence quotes, (3) draft T1 under 120 words citing one real signal. Flag any profile with must-have zero.”

Recruiter checklist before send:
Signal is real; claims about company are true; no sensitive data in prompt that was not needed; person is not mid-process elsewhere; DNC checked; message sounds like a human on your team would send it.

---

Reply-rate tracking spreadsheet fields

Track a small field set per touch and channel so you fix scorecards and copy, not vanity views.

Field	Type	Why it matters
req_id / role	Text	Attribute outcomes to a scorecard version
scorecard_version	Text	Detect criteria drift
channel	Enum	LinkedIn, email, GitHub, referral, CRM
segment	Enum	Passive cold, active, reactivation, referral
fit_score_total	Number	Compare reply quality by fit band
must_have_fail	Yes/No	Should be No for contacted people
touch_number	1-5	Attribute reply to sequence step
sent_at	Date	Cadence and frequency caps
replied	Yes/No	Raw reply rate
reply_type	Enum	Positive, neutral, negative, OOO, unsubscribe
positive_reply	Yes/No	Primary quality metric
interview_booked	Yes/No	Conversion past vanity reply
complaint_or_spam_flag	Yes/No	Stop-the-line signal
human_edited_message	Yes/No	Audit AI vs human quality
notes	Text	Why this person, what signal was used

Definitions: positive = interest, referral, or “tell me more”; neutral = maybe later; negative = hard no. Unsubscribe/complaint ends the sequence and updates DNC. Report positive reply rate and interview conversion weekly.

---

Privacy, frequency caps, GDPR and CAN-SPAM high-level

Operational orientation only, not legal advice. Laws and platform terms vary. Use counsel for your facts.

• Licensed sources only: platform terms for LinkedIn, GitHub, boards, CRM. Violating terms is not innovation.
• CRM before cold: prior apps, DNC flags, referrals first.
• Minimize prompt data: strip contacts when unneeded; use approved tools, not personal consumer accounts.
• Retention and frequency: define note life for non-applicants; default 3 cold touches; no re-open for 90 days without new reason and human approval.
• GDPR-oriented (high-level, not legal advice): know legal basis; honor access/deletion; no secondary training use unless contract and policy allow.
• CAN-SPAM-oriented email (high-level, US context): accurate sender, truthful subjects, opt-out where required, honor opt-outs. Platform messages still need platform rules.
• Truth in advertising: no invented headcount, funding, or promotion claims.

For high-level AI governance context (not legal advice): under the EU AI Act, Annex III treats AI used for recruitment and selection (including targeted job ads and systems that analyse or filter applications or evaluate candidates) as a high-risk employment use case. High-risk obligations continue to phase in through 2026-2027 depending on system type and role. Outbound ranking and automated filtering that affect access to work deserve documentation, human oversight, and counsel review if you are in scope. See ethical AI recruiting for a practical checklist.

---

Diversity sourcing without illegal proxies (careful framing)

Widening who you consider is legitimate. Using protected-class proxies as scoring features is not. Do: widen skill synonyms; drop false must-haves; multi-channel search; score evidence not pedigree; sample exclusions; structured scorecards. Do not: score names, photos, or schools as diversity signals; prompt demographic preference; use zip/hobby proxies; invent group-specific thresholds without counsel. Score the job, expand surface, audit exclusions (not legal advice).

---

Free stack options and when to pay

Choose free / freemium when	Choose paid when
You are piloting one role and can afford human review of every send	Volume requires CRM sequencing, deliverability tools, or ATS-native rankers with logging
Two model UIs + spreadsheet scorecard are enough for ranking notes	You need SSO, DPA, training opt-out, and exportable audit logs
Brand risk of mistakes is high and you want slower, careful outreach	Multiple recruiters need shared memory of who was contacted
Engineering can self-source via GitHub without a sales-style sequencer	You must integrate calendar, ATS stages, and compliance fields

Start free via free AI recruiting tools for 2026 and Free AI Recruiting on i10X. Pay when process discipline is ready. Features without human gates become shadow automation.

---

Worked example: backend hire (before / after)

Context (anonymized). 40-person B2B SaaS, remote EU backend engineer, distributed systems + Go/Java. TTF pressure vs SHRM nonexec medians (~44 days 2025 reports; 39 days SHRM 2026 exec benchmarking). Cost context SHRM 2025 ~$5,475 nonexec average, not a spam justification.

Before: JD paste into a chatbot for “50 ideal profiles,” five near-identical touches, prestige flattery. Weak positive replies; two interviews failed stack must-haves; HM lost trust.

After: scorecard v1 with systems + Go/Java must-haves; prestige banned. Tight/medium/wide Boolean; cap 80; evidence-only rank; human list of 22; 3-touch with real signals.

Illustrative outcomes (scenario, not a published study statistic). Positive replies improved with real-work signals. Prestigious-but-weak-stack profiles were not contacted. One strong non-linear Go candidate advanced. Criteria matched later screening. Style-risk awareness from the i10X AI CV bias study (up to 42 pp hire-rate gap; 1,576 points; 100 profiles; 29 pt evaluator gap) kept notes evidence-first.

---

When NOT to use AI candidate sourcing (or not yet)

• No scorecard exists and the hiring manager cannot name must-haves in 30 minutes.
• You plan to auto-send hundreds of messages without human edit or frequency caps.
• The only “signal” you have is school or employer brand.
• Legal, works council, or platform constraints on outbound are unclear for your market and you have not checked.
• The role is confidential C-level with a closed search firm already running a sensitive process.
• Your CRM is a mess of DNC flags and you would re-contact people who asked to stop.

Fix intake and data hygiene first. AI multiplies quality and garbage equally.

---

Metrics for AI sourcing (baseline, 30 days, 90 days)

Metric	Baseline (week 0)	30 days	90 days
Scorecard signed before first contact	% of reqs with v1 locked	100% on pilot req type	Standard for all sourced roles
Positive reply rate	Measure last 20 cold sends	Improve via signal-rich T1	Segment by channel and fit band
Interview show rate from sourced	Current	Track vs inbound	Stable HM satisfaction
Complaint / unsubscribe rate	Near zero target	Any spike pauses sequences	Quarterly review with brand
Time-to-first-shortlist from source	Current	Where gen AI time savings should appear (LinkedIn: ~20% workweek for TA using gen AI)	Do not trade fairness for speed
Time-to-fill context	Your median	Compare carefully to SHRM ~44 days (2025 cycle reports) / 39 days nonexec (2026 exec benchmarking)	Lagging; never sole KPI
Cost-per-hire context	Your fully loaded cost	SHRM 2025 averages ~$5,475 nonexec / ~$35,879 exec as external reference only	Rework from bad sourced shortlists counts as cost

Avoid vanity metrics (“AI messages generated”) without quality checks.

---

Failure modes and anti-patterns

Failure	What it looks like	Fix
Scorecard by JD paste	Model ranks on buzzwords	Rewrite must-haves as evidence tests
Pedigree overweight	Big logos rank above craft	Ban employer prestige as a dimension
Semantic over-inference	Thin profiles get “confident” scores	Require quotes; score “unknown” as 1 max
Sequence spam	Five identical bumps for every role	3-touch default; 5-touch only when justified
No human gate	Auto-send to hundreds	Human approve contact lists weekly
Style bias bleed-in	Polish beats substance in notes	Evidence-only ranking; read CV bias research
Channel mess	Same person messaged twice in two tools	Single system of record for contact history
Proxy “diversity” scoring	Model told to prefer names or schools	Expand channels and remove false must-haves; no demographic proxies

---

Frequently asked questions

---

---