The Workflows: Screening, Interviewing, Onboarding, Performance

The safest genuine win in the function, and it is worth doing well because it affects who applies at all.

Drafting. A model produces a competent first draft from a role outline far faster than a person, and more usefully, it can rewrite an existing description to remove the accumulated cruft that suppresses applications: unnecessary requirements, jargon, and the long list of desirables that research consistently associates with fewer applications from underrepresented candidates.

That second use is the valuable one and it is underused. A function with two hundred stale job descriptions can improve its applicant pool substantially without touching a screening decision.

What to be careful about. A model trained on historical job postings reproduces the conventions of those postings, including gendered language and inflated requirements, so the output needs review against your own standards rather than acceptance. And a description that oversells produces applications you then have to reject, which moves cost downstream rather than removing it.

Sourcing and advertising is a different matter. Where a system decides which candidates see a vacancy, it is inside Annex III as targeted job advertising, and platform ad delivery has produced documented cases of adverts reaching skewed audiences even with neutral targeting parameters, because delivery optimisation infers who is likely to engage.

So the practical split: drafting is a low-risk capability improvement, and targeting is a high-risk decision that needs the measurement in the next lesson.

Screening is not one thing, and the position on this spectrum determines both the value and the exposure.

Extraction reads a CV and populates structured fields. Mechanical, no judgement, and it genuinely helps because it makes applications comparable. It does not profile and can rely on the derogation.

Filtering on stated requirements applies rules a human wrote: does the candidate hold the required certification. Deterministic, auditable, and it is not really AI, which is why it is safe.

Flagging surfaces candidates matching a pattern. Now the system is influencing attention, and being flagged changes treatment.

Ranking orders candidates by predicted fit. Profiling, materially influential, and squarely high-risk.

And automated rejection removes candidates without human review, which is the strongest form and attracts the most scrutiny including under data protection rules on automated decisions.

The practical judgement: the first two are worth doing and the last is worth avoiding. The middle two are where the function has to decide deliberately, with measurement.

The case for screening automation is always volume: four hundred applications, one recruiter, something has to give. It is a real problem and the reasoning deserves scrutiny.

What is true. A human cannot review four hundred applications carefully. The alternative to a model is not careful human review, it is a human skimming, applying inconsistent criteria, fatiguing through the pile, and being subject to well-documented ordering and similarity effects. The honest comparison is against that, not against an idealised process.

What is also true. A model applies its criteria consistently, which is an improvement when the criteria are good and a systematic harm when they are not. Human inconsistency is noise; model consistency is bias applied uniformly to everyone. A recruiter having a bad afternoon affects the applications reviewed that afternoon. A model that undervalues a background affects every application, forever, identically.

So consistency is not automatically better. It converts a scattered error into a systematic one, which is more measurable and more actionable, and also more damaging while it goes unmeasured.

The conclusion that follows is not that automation is wrong. It is that the volume argument justifies automation only alongside the measurement, because consistent application of unmeasured criteria is the worst configuration available. And it is worth asking whether the four hundred applications are themselves a symptom, since an overly broad advert generating unqualified volume is a sourcing problem being solved at the screening stage.

Interview support splits cleanly into uses that are genuinely helpful and uses that are prohibited, with little in between.

Helpful. Scheduling and coordination, which is ordinary automation. Transcription, which frees the interviewer to listen rather than write and produces a record. Generating structured question sets tied to the role's requirements, which improves consistency and is one of the better-evidenced interventions in selection generally. And post-interview summarisation against a defined rubric, where the interviewer's own judgement is being organised rather than replaced.

Prohibited or high-risk. Any analysis inferring emotional state, engagement, confidence or stress from video, voice or facial expression. This is the emotion inference prohibition and it applies squarely. Vendors describe these products in various ways, so the test is what the system claims to infer.

And scoring interview responses to produce a candidate ranking is profiling, so it is in the high-risk tier.

One further caution on transcription and summarisation. A summary is a compression, and what it compresses out is a choice made by the model. A summary emphasising fluency will systematically disadvantage candidates who are less fluent in the interview language for reasons unrelated to competence. So even the assistive use benefits from a rubric that says what the summary should capture, rather than a general summarise this.

And tell candidates they are being recorded and transcribed, which is both a legal requirement in many places and the minimum of decent practice.

If an HR function wants a first deployment that delivers value without acquiring high-risk exposure, onboarding is the answer and it is consistently overlooked in favour of screening.

Why it fits. The questions are repetitive and high-volume. The answers exist in documents the organisation already maintains. The consequence of a wrong answer is usually recoverable, since a new joiner told the wrong holiday policy corrects it. And nobody is being ranked, scored or selected, so none of the Annex III machinery applies.

What works. A retrieval-based assistant over policy documents, the handbook, benefits information and process guides, answering in the joiner's own words rather than requiring them to know which document to open. Personalised checklists generated from role and location. And drafting the sequence of communications a joiner receives.

What it requires, which is the honest catch. The corpus has to be current, and HR documentation is frequently not. A confident answer from a superseded policy is worse than no answer, and the staleness and permission problems from the company brain cursus apply in full, since HR documents have genuinely different audiences and some should not be broadly retrievable.

The useful reframe: an onboarding assistant is a forcing function for fixing the documentation. Most functions discover, on attempting it, that they have three versions of several policies and no owner for any of them, which is a finding worth having.

Performance work divides the same way screening does, and the label problem returns in a sharper form.

Assistive and useful. Summarising feedback collected from several colleagues. Drafting a review from a manager's notes, which addresses a task managers genuinely dislike and defer. Checking a draft review for vague or non-specific language. Surfacing what a manager wrote previously so the review is consistent over time.

High-risk. Scoring or ranking employees, or generating a rating that feeds compensation or promotion. This is monitoring and evaluating performance under Annex III.

The label problem is worse here than in hiring. A model trained on historical performance ratings learns the rating behaviour of past managers, and performance ratings are among the most studied and most consistently biased artefacts in organisational research, with well-documented differences in how the same behaviour is described across groups. A model reproducing those ratings faithfully is reproducing that.

And there is a specific trap in drafting. A model drafting a review from sparse notes will fill gaps with plausible language, and the plausible language for a given role and seniority carries the conventions of the corpus it learned from. That is fabrication about a named individual, in a document that affects their pay.

So the rule for drafting: the model organises what the manager wrote and does not add substance. If the notes are thin, the answer is better notes, not a longer draft.

HR holds the most sensitive personal data most organisations process, and putting it near an AI system raises issues beyond the AI Act.

What sits in HR systems: health information including sickness absence and accommodations, trade union membership, disciplinary records, salary, performance history, family circumstances, and in some jurisdictions demographic data. Several of these are special category data under the GDPR with heightened protection.

Three specific cautions.

The consent problem. Consent from an employee is generally a weak legal basis because of the power imbalance, so processing usually rests on another basis, which has to be identified rather than assumed. An HR AI deployment resting on employee consent is on shaky ground.

The secondary use problem. Data collected for payroll being used to train an attrition model is a new purpose, and purpose limitation applies. This is the most common quiet breach in people analytics.

And the retrieval permission problem. An HR assistant over internal documents must be permission-aware, because HR documents include material about individuals that other individuals must not see. The oversharing risk from the company brain cursus is acute here, since the corpus contains exactly the material that must not surface.

Also worth knowing: employee representative bodies have consultation rights over monitoring and workplace technology in many European jurisdictions, so the works council conversation is a real step rather than a courtesy.

Pulling the workflows into a sequence, ordered by value delivered per unit of risk acquired.

First tier, do these. Job description drafting and rewriting. An onboarding assistant over current policy documents. Interview scheduling and transcription with consent. Structured question generation. Feedback summarisation for managers. All of these improve the function's capacity, none of them decides anything about a person, and together they absorb a meaningful share of the administrative load that stops HR doing its actual work.

Second tier, with care. CV extraction into structured fields, which genuinely helps and stays inside the derogation provided it does not score. Review drafting from manager notes, with the rule that the model organises rather than adds.

Third tier, only with the full apparatus. Anything that ranks, scores, flags or filters a person. That means the diligence from the procurement cursus, the disaggregated measurement from the bias cursus, the oversight design from the regulated-industries cursus, and a documented position on the Article 6(3) analysis.

And not at all: emotion inference and biometric categorisation, in any form, on candidates or staff.

The observation worth ending on is that the first tier is where most of the realisable value sits, and it is the tier that gets least attention, because screening is the problem HR functions feel most acutely and the vendors sell hardest against.