Sourcing and outreach at volume: the discipline has to come from you, because it no longer comes from effort

Volume metrics are seductive because they move. Messages sent went up 400%. Something must be happening. What is actually happening is that you have divided the same amount of relevance across five times as many people, and relevance per message is the thing the recipient experiences.

• Three questions before any message goes out. Why this person specifically? What in the scorecard's outcomes would they recognise as their own work? What would make them reply even if they are happy where they are? If you cannot answer all three, the message is not ready and the answer is not a better template.
• Set the cap before you need it. Nadia S. and you agree a weekly ceiling per role now, while it is abstract, and write it in the workspace. Kestrel's is 40 a week per open role. Deciding it in week 9 with 90 roles unfilled is not deciding, it is negotiating with your own panic.
• Track replies and conversations, not sends. Sends measure your afternoon. Replies measure whether the market thinks you are worth answering.
• The cap is what protects the 115 roles after this one. Your reputation in a talent market is a shared, slow-moving asset. It is spent quickly and rebuilt over years.

"I was really impressed by your work at [Company]" is not a compliment. It is a form field, and everybody receiving it now knows what a form field looks like. Three separate costs, and the third one is the expensive one.

• It is detectable. A senior engineer has received forty of these this year. The tell is that the praise contains no specific claim - nothing that could be checked, nothing that could be wrong.
• It wastes the candidate's time. They read a paragraph designed to look like it took thought and discover it took none. That is not neutral, it is a small insult delivered at scale.
• It teaches your market to ignore you. This is the real cost. Each generic message trains the recipient's filter, and the filter does not distinguish between the message you sent in a hurry and the one you will need them to read next year.
• The honest alternative is short. "We are hiring a Senior Support Engineer at Kestrel. Two enterprise accounts, a real escalation queue, band published in the ad. I am contacting people who have owned a queue like it. No pitch beyond that." That is fine. It claims nothing false and it respects the reader's time.

• A number, per role, per week. Kestrel: 40. Yours might be 15 or 100. What matters is that it exists before the pressure does.
• A relevance gate that a colleague could apply. Written as a rule, not as an instinct: "one line naming what in this person's actual experience matches an outcome on the scorecard, or it does not send."
• A stop rule. No third message. Ever. Two contacts and then silence is a real answer you have received.
• A named owner. Nadia S. owns the cap for this wave. If it moves, it moves in writing with a reason, because a cap that quietly drifts was never a cap.

Every sourcing habit is a filter wearing comfortable clothes. None of these are decisions anyone made; they are grooves, and grooves get deeper.

• The alumni groove. Search filters that keep returning the same institutions, because that is what the last five hires had and the filter was built from them.
• The competitor groove. Sourcing exclusively from two or three named firms. Efficient, legible to the hiring manager, and it imports another company's hiring bias wholesale into yours.
• The referral groove. Referrals are your best-converting channel and your least diverse one, because people refer people like themselves. Use them, cap them, and never let them be the only channel for a role.
• The unbroken-timeline groove. Filtering out gaps removes carers, people who were ill, people who retrained. None of that is a capability signal.
• The fix is boring and it works. Source from the capability, not from the biography. Kestrel's outcome is "has owned a live escalation queue", and that person exists in managed service providers, in-house IT, telco NOCs and two-person startups, none of which are on the current list.

If an ad or a sourcing campaign is shown to a narrowed audience, that narrowing has already made a selection decision - before any application exists, and without producing any record you can audit later.

• It is inside the regulated scope. Targeted job advertising falls within the EU AI Act's high-risk employment category (Annex III point 4). Who an ad is shown to is treated as part of selection, not as adjacent marketing.
• Delivery optimisation narrows without being told to. A system optimising for engagement can learn to show a role to whoever historically engaged with that kind of role. Nobody wrote a rule. The pattern came from the past, which is precisely the failure mode b4 puts under the microscope.
• So ask the same three questions you would ask of a screen. Where did it run, who was it shown to, and on what parameters. Keep the answers with the role file.
• Not legal advice. This is education for HR professionals; obligations vary by jurisdiction and by what your platforms actually do. Leader session a2 holds the legal map and your DPO holds the position.

Sourcing generates the exact material rail 3 exists for. Public does not mean unregulated, and a spreadsheet of 400 profiles is a candidate database whether or not anyone calls it one.

• Lawful basis. You need one to hold and use candidate data under GDPR and PDPA, including profile data gathered from public sources. "It was on the internet" is not it.
• Minimisation. Collect what the role decision needs. Not photos, not age signals, not personal life, not inferred characteristics, not the whole profile because it was easy to copy.
• Transparency. People are generally entitled to know you hold their data and where it came from. Your first message is often the honest place to say it.
• Retention. A sourcing list with no deletion date is a permanent one. Set the date when you create the list, not when someone finally asks.
• And the workspace rule still holds. Profiles, CVs and candidate contact details do not go into an AI workspace. The outreach kit is drafted from the scorecard, which contains no personal data at all.

• The outreach message, built from real outcomes. Feed it the scorecard's month-6 outcomes and ask for a message that describes the actual work. The interesting thing about Kestrel's role is not "exciting opportunity", it is two enterprise accounts and a queue that nobody has owned end to end.
• The follow-up sequence. One follow-up, sent once, that adds something rather than repeating. Then silence.
• The role FAQ. Band, level, on-call, remote pattern, interview stages, timeline. Every question you are about to answer 24 times, answered once and linked.
• The keep-warm templates. The "you are still in process, next update Thursday" note. Unglamorous, and the single biggest driver of whether candidates think you are a serious company.

• It does not decide who to contact based on inferred characteristics. Not from a name, a photo, a school, a location, a graduation year, or anything a profile lets you guess about a person. The contact list comes from capability signals, and a human decides. This is rail 1 arriving earlier than most people expect it to.
• It does not touch a rejection. Not the decision, and not the message that carries it. A rejection is the most consequential thing a candidate receives from you and it is the one place where a fluent, warm, entirely generated paragraph is worst. A person writes it, or at minimum a person reads every word and sends it under their own name with a real reason in it.