Recruitment is the front of the HR value chain — six discrete stages from role spec to offer, each one producing an artifact the next stage consumes: a job description, a sourced shortlist, a screening scorecard, an interview guide and notes, an assessment matrix, an offer pack. Structure each output and the whole chain becomes agent-addressable and auditable — which matters more here than almost anywhere else in HR, because South African law binds the chain from its very first step.
Recruitment isn't one event — it's a chain of six discrete stages, each transforming an input into a defined output the next stage depends on. Role spec produces a job description. The JD drives sourcing, which produces a shortlist. The shortlist feeds screening, which produces a scorecard. And so on, down to an accepted offer that hands a new employee to the rest of HR.
The leverage is in treating each stage's output as a structured artifact rather than a loose email thread or a recruiter's memory. A JD with explicit fields — title, inherent requirements, must-haves, nice-to-haves — is something an agent can draft, a sourcing query can target, and an auditor can later inspect. The same applies to a screening scorecard or an interview guide. Structure makes the chain agent-addressable and auditable: an agent can act on a well-formed artifact, and a human can later show why a decision was made — which, in the SA context, is not optional.
This leaf maps the six stages, the output each emits, how those outputs feed forward into the Frappe HR employee record at offer-accepted, and the law that binds the front of the chain. It sits under the Compliance backbone and feeds Onboarding downstream.
Six stages, six structured outputs. The output column is the point: each artifact is what the next stage consumes, what an agent can draft or reason over, and what an auditor can later inspect.
| Stage | What happens | Structured output |
|---|---|---|
| 1 · Role spec / JD | Define the role, its inherent requirements, must-haves and nice-to-haves; turn it into a job description that's lawful to advertise. | JD document — title, purpose, inherent requirements, criteria |
| 2 · Sourcing | Find candidates against the JD — adverts, referrals, talent-pool search, market research on where the skills sit. | Sourced shortlist — candidate records mapped to the JD |
| 3 · Screening | Filter the shortlist against the criteria — CV review, qualifying questions, knock-out requirements. | Screening scorecard — candidate × criterion, with notes |
| 4 · Interview | Structured conversation against a consistent guide so candidates are assessed on the same dimensions. | Interview guide + notes — questions, ratings, evidence |
| 5 · Assessment / selection | Combine screening, interview and any tests into a comparable view; make and justify the selection. | Assessment matrix — candidates × dimensions, ranked |
| 6 · Offer | Extend terms, negotiate, confirm acceptance — the boundary where a candidate becomes a hire. | Offer pack — terms, signed acceptance, start date |
Naming each stage and its output gives you a handoff contract: stage N can't start until stage N−1 has emitted its artifact. That's what lets you insert an agent at one stage without it reaching across the whole process, lets you audit a single decision in isolation, and lets you point to the exact artifact that justified a hire — or a rejection — if it's ever challenged.
The chain only works because the artifacts compose. The JD's criteria become the screening scorecard's columns; the assessment matrix becomes the justification on the offer; the accepted offer becomes the seed of an employee record.
The JD's inherent requirements and must-haves are the sourcing query. A loose JD produces a noisy shortlist; a structured one targets.
Each sourced candidate becomes a row in the screening scorecard, evaluated against the same criteria the JD declared.
Gaps and signals from screening shape what the interview probes — a consistent guide so every candidate is assessed on the same dimensions.
Interview ratings and evidence fold into the assessment matrix alongside screening and any tests — one comparable, ranked view.
The matrix is the documented basis for the selection — the artifact that explains why this candidate, attached to the offer decision.
At acceptance, the candidate crosses into the HRMS: the offer pack seeds the Frappe HR employee record and onboarding begins.
The decisive boundary is offer-accepted. Up to that point you're in the ATS — the applicant tracking system — managing candidates. At acceptance the person becomes an employee and the record moves to the HRMS. Getting that handoff clean means the data the ATS already holds (name, contact, the role they were hired into, the start date) flows once into the employee record rather than being re-keyed. The Huly leaf covers the ATS side; Frappe HR covers the receiving side.
Most HR law governs the employment relationship once it exists. Recruitment is the exception — two statutes bind every employer at the front of the chain, before a contract is ever signed, and a third constrains how AI may be used to screen.
The Employment Equity Act — prohibition of unfair discrimination. Chapter II of the Employment Equity Act 55 of 1998 (as amended by the Employment Equity Amendment Act 4 of 2022, effective 1 January 2025) prohibits unfair discrimination on listed grounds — race, gender, sex, pregnancy, disability, religion, HIV status and others — and it binds every employer regardless of size. This is distinct from the affirmative-action duties, which apply only to designated employers (50 or more employees). Discrimination may be justified only by the inherent requirements of a job — which is exactly why the role-spec stage has to name those requirements explicitly: it's the lawful basis on which you screen anyone out.
POPIA — candidate personal information. A recruitment pipeline collects a lot of personal information about people who may never become employees. The Protection of Personal Information Act 4 of 2013 applies: minimality (collect only what the role assessment actually needs), purpose limitation (don't reuse candidate data for something they didn't consent to), a lawful basis for processing, and security safeguards (s19). Unsuccessful applicants' data needs a defined retention period and deletion — you can't hold a rejected candidate's CV indefinitely without a justified purpose.
POPIA s71 — automated decision-making. This is the one that bites AI screening directly. Section 71 restricts a decision that results in legal consequences for a person, or affects them substantially, where that decision is based solely on automated processing intended to profile them — aspects like their performance, reliability or conduct. A fully automated screen that rejects candidates with no human involvement is squarely in scope. The practical rule: an agent may assist screening, but a human must remain in the loop on any decision that filters a person out.
This is orientation, not legal advice. Employment-equity and data-protection law is fact-specific and changes; the framing here is current as at mid-2026 and points you at the primary sources — the Department of Employment and Labour, the Information Regulator, and the CCMA. Validate any recruitment process, screening method, automated-decision design, or candidate-data retention policy with a qualified SA labour-law practitioner before you rely on it. An agent in this domain drafts and assists; a human and their advisor decide.
The pipeline needs a system of record for the candidate side, a clean boundary into the HRMS, and an agent that assists without ever owning the decision.
The ATS tracks the pipeline. Huly's applicant tracking gives each candidate a record and each stage a column — the six stages map onto a board, and the structured outputs hang off the candidate. The Huly leaf covers it in depth. The ATS is the home of the chain up to offer-accepted.
The HRMS receives at offer-accepted. When a candidate accepts, they stop being an applicant and become an employee. That's the boundary where the record crosses from the ATS into the Frappe HR employee record — the seed of everything downstream: onboarding, payroll, the statutory layer in Compliance.
The Fractional CHRO agent assists — it doesn't select. Grace, the fractional-CHRO agent, is genuinely useful across the chain: drafting a lawful JD from a role spec, researching where a skill sits in the market, summarising a screening scorecard, generating a structured interview guide. But Grace never makes the final selection. Every decision that filters a person out — especially one that could touch a listed ground or fall under POPIA s71 — stays with a named human. The agent is leverage on the artifacts; the accountability stays human.
Treat any agent access to the candidate store as a POPIA processing operation: scope it to the minimum, log every access, and keep the selection and rejection decisions human-approved. The agent reads and drafts; it does not autonomously reject. This is the recruiting-side application of the same access discipline the Frappe HR and Compliance leaves set out.