Thesis (Alf’s read): Recruitment has quietly stopped being a conversation, and the cheapest fix available to any employer — and the only one that scales — is to put a human back at exactly two points in the funnel: the decision, and the feedback. Everything else (screening, scheduling, ranking, chasing) can be automated without harming anyone, provided the process still answers three questions honestly: can you do the job, do you want the job, and do we want you. AI can help answer the first. It cannot answer the other two, and a hiring process that cannot answer them is not selecting people — it is filtering them out politely and calling it efficiency.

Recruitment That Values People

The vision in one paragraph

A hiring process designed for the applicant as a person with three questions to answer and one question of their own: can you do the job, do you want the job, do we want you — and do I want to work here? Every stage exists to answer one of those, honestly and in writing. AI is used where it genuinely helps — screening for human review, scheduling, and above all giving every candidate an answer — and is barred from the two places where it destroys the thing being assessed: the decision, and the assessment of motivation. Applicants know where they stand at every point, including when the answer is no. Employers measure what their process does to the people in it, publish the numbers, and treat “we got back to everyone” as a standard, not a favour. The result is not a kinder version of a lottery. It is a shorter, more human funnel that produces better hires — and it is the only part of the youth-disconnection problem that any single employer can fix on their own.

What the vision is NOT

• Not anti-AI. Screening a thousand applications by hand is not virtuous; it is an excuse for not hiring. The target is where the machine sits, not whether it exists.
• Not “abolish CV screening”. Automated ranking feeding a human review is legitimate. Automated rejection is not.
• Not a quota, or an attempt to engineer outcomes. It is a set of process commitments whose result — better information on both sides — is the point.
• Not a claim that process reform fixes youth unemployment. It does not create a single job. See ## My assessment (Alf).
• Not another unpaid-internship scheme. The obvious cheap implementation — trial shifts — is unpaid work by another name and is the vision’s biggest equity trap.

The problem this addresses

• 140 applications per graduate vacancy (vs 38 in 2002–03, ISE) in a market where 73% of students and graduates now use AI somewhere in the application process, and one candidate in three has sent more than 100 applications.
• Candidates and employers do not even agree on what is happening: 75% of candidates believe AI screens CVs; 21% of recruiters say they use it there (CV-Library). 45% of employers give candidates no guidance at all on AI use, and 29% discourage it without enforcing anything (Ada Lovelace).
• Almost half of UK jobseekers have been interviewed by an AI system; 30% have walked away from a process because of one; 75% want AI use disclosed; 80% oppose AI making the final decision (Totaljobs / HR Grapevine). 53% believe they have been rejected by a machine without human review (CV-Library).
• And the stake is not in the workplace. Just under 1 million UK young people are NEET, youth unemployment is 16.2%, and the share of NEETs reporting a work-limiting health condition has risen from 26% (2015) to 44% (2025). Disconnection is sticky, welfare-funded and slow to reverse: see Youth Disconnection and NEETs - Global Problem.

The test — three questions, asked by both sides

The spine of this vision is Bolles’ frame, applied literally to the funnel. Every stage below must be able to answer the question it is supposed to serve:

Question	Who asks it	What AI can honestly do	What it destroys when automated
Can you do the job?	Employer	Keyword/qualification matching, work samples, structured scoring — at volume	Nothing fatal, unless every application is AI-polished, at which point the screen selects on noise
Do you want the job?	Employer	Nothing. Motivation is the part of the signal AI silently deletes	The entire basis of an offer — this is why “everyone sounds the same”
Do we want you? (fit)	Employer	Nothing. Resemblance to a pattern is not fit	Diversity of route and background; algorithmic monoculture favours the same candidates everywhere
Do I want to work here?	Candidate	Nothing — and currently deleted from the process entirely	The candidate’s ability to choose, and their reason to stay after joining

The design — twelve levers, by stage of the funnel

Each lever names what it does, who owns it, and the honest limit.

Stage 1 — The advert (owner: employer)

1. Say what you will do with AI, and mean it. Publish a short, specific policy: where AI is used, what it never decides, what you expect candidates to disclose, and who to ask about it. This is the single cheapest intervention in the entire vision — 45% of employers currently say nothing, which converts a tool into a grievance. Limit: a published policy that is not true is worse than silence.

2. State the pay band and the process up front. How many stages, how long, whether there is an automated step, and the salary. It costs nothing and it converts the funnel from a mystery into a contract. Limit: none worth naming — this is free.

Stage 2 — The application (owner: employer; schools and careers services supporting)

3. Judge the work, not the artefact. Where the role allows, ask for something small and real — a short written answer to a genuine problem, a portfolio piece, a plan — rather than a fifth generation of a polished CV. Limit: unpaid effort. Keep it under an hour, and never require a completed work product.

4. Ban the keyword arms race in your own scoring. If your screen rewards containing exact advert phrases, you are selecting for prompt skill, not capability. Score on evidence of doing, and say so. Limit: requires a competent scoring design — most ATS defaults do not do this.

Stage 3 — Screening (owner: employer; regulator setting the floor)

5. AI may rank. A human must decide — and be able to explain. No automated rejection without a person having reviewed the file. This is already the direction of law (EU AI Act, high-risk, from 2 December 2027; human oversight that amounts to a rubber-stamp explicitly does not comply) and it should be a British standard regardless of what the UK legislates. Limit: at 140:1 this costs recruiter time — which is precisely why the honest version means raising the screening bar, not lowering it.

6. Screen for what you actually need and delete what you do not. Postcode, school, name, age, gap years are usually proxies. Strip them by default; if they survive, they need a documented reason. Limit: mostly a checklist, cheaply done.

7. Audit the screen for disparate impact — rejection rates by school, postcode, gender, ethnicity and age — and keep the record current. A model trained on past hires reproduces past hires. Limit: needs someone to own the number; small employers will need a shared or sector service.

Stage 4 — The human moment (owner: employer)

8. Guarantee one human contact before any automated assessment. Fifteen minutes, live, before any test or recorded interview. This is the lever that answers “do you want the job?” and it is the one the BBC’s candidates were denied. Limit: recruiter time. Mitigate with pooled hiring across employers, sector schemes, or trained non-recruiters.

9. Never open with the AI interview. If used at all, use it late, declare it plainly, and never for motivation. Recorded answers against a countdown are the single worst design in the current market. Limit: some employers use it for scale and will keep doing so — hence lever 10.

10. If you use AI interviews, cap what they decide. They may order candidates for human review; they may not rank into rejection, and emotional-state inference must be banned outright — it is already effectively prohibited in the EU and is junk science in a hiring context. Limit: enforcement is weak everywhere.

Stage 5 — Assessment and decision (owner: employer)

11. Use trials, placements and work samples — and pay for them. The employers in the six-skills film who can actually tell candidates apart are the ones who meet them: a trial shift, a factory floor, a bar. But a trial shift is unpaid work, so the vision has to say the uncomfortable part: these must be paid, short, and never a substitute for a decision the employer has already made. Limit: real cost; the equity trap; needs to be deliberate — see ## Challenges.

12. Decide on evidence, record the reason, and keep a second human in the loop. One-paragraph rationale, held for a defined retention period, reviewable on request. Limit: GDPR-defined retention, and honesty about what the true reason was.

Stage 6 — Feedback (owner: employer — this is the one nobody does)

13. Answer everyone. An acknowledgement on application, a status at each stage, and a real rejection with one sentence of reason. This is where AI should be spent: on feeding people back, not on filtering them out. It is also the intervention that directly addresses the harm in the research — silence, not rejection, is what candidates describe as “really disheartening”. Limit: needs templates and a policy on how honest a reason can be; done badly it reads as automated contempt.

14. Give feedback that is usable. “Not enough relevant experience” is not feedback. One specific observation costs a recruiter two minutes and is the difference between a candidate improving and giving up. Limit: legal caution on saying why — a real constraint, not an excuse for saying nothing.

Stage 7 — Measurement and the floor beneath it (owners: employer, sector, government)

15. Publish outcomes, not values. Applications per hire; share of candidates who received feedback; share who met a human before any automated stage; disparate-impact summary; time-to-decision. Employers measure everything else — measuring their own candidate experience is what stops “we care about people” from being marketing. Limit: publication is reputational exposure, which is exactly why it works.

16. A floor, not a favour. The policy version: disclosure of AI use, no solely automated rejection, a right to human review on request, and a duty to respond. The EU has already legislated the first two and deferred them once; the UK has guidance and no statute, which makes British candidates the test population. This is the lever the state owns, and it is the only one that reaches the employer who does none of this voluntarily.

The candidate’s own playbook — moved 2026-10-13

The candidate-facing playbook (Bolles’ three questions, the doom-loop market evidence, and levers 17–21 as they then stood) has been consolidated with the wellbeing evidence into one note: The Candidate’s Path — Getting In, and What It Is For.

This note keeps the systemic half — what employers, platforms and regulators should do about the funnel. The two are one question asked twice: how to enter the market, and what to want from it. They stay cross-linked rather than merged because the owners are different people: employers own this one, the candidate owns that one.

Both sides need the competence — and neither should need to be technical (added 2026-10-11)

The owner’s steer, verbatim: “Both demand and supply side of recruitment needs to adopt proficiency in AI and their personal Hermes agents, but not everyone is an IT geek like me (being AI native).”

Read as a design constraint rather than a slogan, it says three things:

1. The competence requirement is symmetric, and the gap is on both sides. Employers already run AI screening without understanding it — the Australian evidence is that they switch on bias-creating features unknowingly and could not audit what they bought; candidates are then told to work around the same systems. A market where one side deploys a tool it does not understand and the other must out-guess it is not a candidate-side skills gap. It is a competence gap on both.
2. The remedy is a personal agent, not a training course. If “adopt proficiency in AI” means everyone must learn to prompt, the burden lands on the people with the least time — the 53-year-old with 240 applications, not the graduate with a free afternoon. The capability has to arrive pre-built, which is what turns “be AI native” from an aspiration into an access question.
3. Access is the constraint that keeps the rest honest. Any recruitment reform that assumes technical fluency rebuilds the same exclusion one layer up: it stops being about age and becomes about who had the confidence to learn the tool. Design for the person who will never open a terminal — the agent does the work; the human answers the three questions.

This connects to Being AI Native - Competency Ladder with a caveat worth keeping: a competency ladder only the already-technical can climb is not a ladder, it is a filter.

Who does what — owners and honest limits

• Large employers: all of it. They have the volume to justify the design and the reputational exposure to be moved by it. This is where the standard should be set.
• Small and medium employers: levers 1, 2, 8, 13, 14 (and paid trials where the role allows). They have neither margin nor HR function, so anything requiring universal virtue will fail — the mitigation is shared services, sector pooling and template policies, not exhortation.
• Platforms and ATS vendors: build rank-not-reject as the default, make the audit data exportable, and stop shipping emotional-state scoring. Vendor defaults are, in practice, the industry’s policy.
• Government / regulators: the floor (lever 16) and the measurement that makes it enforceable. The UK’s current position — guidance and no law — means the market sets its own standard, and the market currently sets it at 21%.
• Schools, colleges, careers services: teach the three questions explicitly, and treat the application as evidence-building rather than form-filling.
• Candidates: the playbook in The Candidate’s Path — Getting In, and What It Is For. Not obligations — survival tactics in a market that is not currently designed for them.

What it would cost — and what it saves

• Costs: recruiter time (the scarce thing, and the real constraint), paid trials, feedback templates, audit capability. None of it is capital expenditure; all of it competes with hiring volume.
• Saves: screening waste. 83% of recruiters use AI to speed hiring and only 36% think it actually delivers — the industry is already paying for a promise it does not believe. Better screening, fewer wasted interviews, less re-hiring after bad fits, and a candidate pipeline that does not poison your reputation in the year you need it.
• The externality argument, which is the real one: the costs of a dehumanised process do not appear on any employer’s P&L. They appear as youth disconnection, welfare spending, NHS mental-health costs and lost tax — Eurofound’s modelled European cost of NEETs is around €142 billion a year. An employer that runs a machine-only funnel is exporting a cost it never pays. That is the definition of something that needs a floor beneath it.

New evidence — age, the other end of the same funnel (added 2026-10-11)

Filed against lever 6 (strip the proxies) and lever 7 (audit for disparate impact), after the owner sent 60 Minutes Australia’s Computer Says No — see 261011 Computer Says No — AI Screening and the Ageism Nobody Switches Off.

The finding this vision was missing: the same funnel excludes at both ends. The problem half above is built on youth disconnection — young people screened out for having no experience. The Australian programme supplies the mirror image: the middle-aged screened out for having too much, with the average duration of unemployment at 86 weeks for the 55+ group against 37 weeks for 15–54-year-olds. One process, two excluded generations, and the design answer is identical in both directions — judge the work, not the artefact; strip age along with the other proxies; and audit the outcome rather than trusting the intent.

What the new evidence does to lever 7. Disparate-impact auditing is the lever with the weakest enforcement behind it, and this is why: New York City’s Local Law 144 requires bias audits for race and sex but not age, and the EEOC’s AI guidance covers disability and race but not the ADEA. Age keeps falling through the regulatory gap. That makes the employer’s own audit — lever 7 — not a compliance exercise but the only place the number currently exists at all.

Three corrections this evidence forces on the framing.

1. Do not attribute the age penalty to AI. Australia’s duration gap is in ABS data from February 1978 — decades before automated screening. AI is at most an amplifier, and the honest claim is narrower and stronger: a model trained on past hires reproduces past hires, without leaving a witness.
2. The silence is the new harm, not the bias. An auto-rejected applicant triggers no notification to the hiring team, so the human moment that used to catch an explicit instruction — “don’t shortlist anyone that’s not under 40” — has been deleted from the process. Lever 5 (“a human must decide”) is therefore not a courtesy; it is the only remaining check on conduct that is otherwise unobservable.
3. The candidate-side workaround is documented practice and belongs in the honest section, not the advice. “Botoxing” a CV — removing age and trimming history — is mainstream guidance with no evidence behind it, and a tax paid only by those with something to hide. The candidate’s playbook should say so rather than recommend it.

One thing the new evidence supports, and one it undercuts. Supports: structured, standardised assessment has both the highest predictive validity and the lowest adverse impact — the empirical spine under levers 3, 5 and 11. Undercuts: blind recruitment, the most obvious remedy for age bias, is mixed to negative in the trials (Canada’s 2017 name-blind pilot found no effect; the French experiment backfired). Lever 6 should therefore be argued as proxy removal with a reason, not as a proven fix.

My assessment (Alf)

Strongest part: the diagnosis and the cheap lever, into which this vision adds something the films do not say out loud — that feedback is the most under-used application of AI in existence. “Answer everyone with one useful sentence” is the only intervention in this note that is simultaneously good for candidates, cheap, automatable, and directly opposite to what the research says is happening (“silence, not rejection, is the complaint”). If one thing from this note survives contact with the world, it should be that.

Weakest part: the arithmetic of what is being distributed. At 140 applications per vacancy, this vision distributes a scarce good more humanely; it does not create the good. That is why I have been careful to file it as a Problems-level intervention upstream of a demand-side crisis (Economic Instability) rather than as an answer to youth disconnection. A rival could fairly say the honest title is “recruitment that wastes less of everyone’s time”, and it would still be worth doing.

The uncomfortable equity trap, stated plainly: paid trials are the right answer and the least likely to happen. Unpaid trials exclude exactly the people this vision is meant to help — those who cannot afford a day of unpaid work, which the six-skills film itself admits before moving on. Any employer adopting lever 11 without paying for it has made the problem worse while feeling virtuous. I would rather this note said “pay for it or don’t do it” than pretend there is a free version.

Most realistic single reform: transparency. Disclosure of where AI is used, and no solely automated rejection. It is cheap, it is already EU law (twice deferred, but law), it requires no employer virtue, and it directly attacks the measurement gap in the research (75% assume screening; 21% report it). The gap between those two numbers is where most of the felt harm lives — not in algorithmic rejection, which is real but smaller than candidates believe.

Where I would resist the framing altogether: “best practice” lists of this kind have a habit of becoming compliance theatre — a code of conduct that the top 5% sign and the other 95% never read. The mechanism that actually changes behaviour in the evidence here is not exhortation but reputational measurement (lever 15) and a legal floor (lever 16). Rules beat appeals — the same conclusion the Direct Vision Standard reached in the freight industry: a mandate moved behaviour that a decade of voluntary good practice did not.

And the honest uncertainty. Every headline figure behind this vision is a survey, most are one year deep, several are vendor-adjacent, and the one genuinely hard number (140:1) predates generative AI. The direction — more automation at the door, more candidate scepticism, more silence — is well-evidenced. The magnitudes are not, and a vision built on magnitudes would be irresponsible. That is why the levers here are designed to be worth doing even if the numbers are half-wrong.

Challenges (the vision’s own, plus mine)

• Employer self-interest is the only durable engine. Anything requiring virtue fails. Every lever here is therefore pitched at a benefit employers already want: better hires, less wasted screening, legal safety, reputation.
• SME capacity. Half the UK workforce is employed by firms with no HR function. Without shared services, levers 7, 11 and 15 are large-employer-only, which shrinks their impact precisely where youth hiring happens.
• The paid-trial cost. Real money, for a population that may not stay. Mitigation: short (half a day), pooled across employers, or replaced with paid work samples where the trade allows.
• Feedback risk. Honest reasons carry legal caution. Solved by templating, not by silence.
• The counterfactual is untestable. Nobody can prove a human-in-the-loop process hires better. The case rests on candidate harm reduction and legal exposure, not on a hiring-quality guarantee — and I would not pretend otherwise.
• AI-only applicants. Nothing here prevents a candidate from automating 500 applications. If process reform raises the human cost per application, the rational response may be more automation upstream, not less.

How this connects to the bigger thread

This is the third step of the loop the owner set out in Purpose of Alf: intentional news (the two BBC films) → polycrisis understanding (Youth Disconnection and NEETs - Global Problem, and the crises it sits inside) → a vision of a world that addresses it — here, a hiring process whose design answers the three questions Bolles said every interview really asks. It is deliberately modest in scope: one process, in one domain, that a single employer can change on Monday morning without legislation. The demand-side answer — more entry-level roles, and work that covers its own commute — belongs to Economic Instability and the missions in Gen Z Rising - Skilling Path and Worker Cooperative. This note is the part that does not need to wait for the economy to change.

Open questions

• Does the owner want this at the level of a playbook (as written) or a manifesto? The e-bike precedent was a playbook; this one is halfway.
• Test the “answer everyone” lever against a real process — what would it cost a mid-sized UK employer to reply to all applicants with one usable sentence? Worth a worked estimate.
• Has any UK employer published candidate-experience metrics? If one exists, this vision has a proof case instead of an argument.
• Watch the EU deadline — high-risk duties now bite 2 December 2027 (deferred once). Does the UK legislate anything before then, or stay on guidance?
• Verify the four load-bearing figures against their primary releases (ISE 140:1, Prospects at Jisc 73%, CV-Library 75%/21%, Ada Lovelace 45%) — currently cited from secondary reporting.
• Is “recruitment that values people” the right title, or does the honest one need the demand side in the name?

Sources

All figures are as reported by the sources below; those marked [secondary] were reached through reporting rather than the primary release, and have not been traced further. See the open questions.

• Richard Nelson Bolles, What Color Is Your Parachute? Guide to Rethinking Interviews (Ten Speed Press, 2014) — the three questions (“Do we like you? Do we want you? Do you have the skills…? Do you have the right attitude…?”) and the candidate’s mirror question (“Do I want to work here?”); the interview-as-conversation framing with two phases on each side.

• BBC News, “Everyone’s using AI to get hired. Now what?” (YouTube, 18 Sep 2026) — https://www.youtube.com/watch?v=JulfdGi0qMw → filed as 260918 Everyone Using AI To Get Hired Now What

• BBC News, “The six skills employers actually want” (YouTube, 26 Aug 2026) — https://www.youtube.com/watch?v=4ih0t8wlKIg → 260826 The Six Skills Employers Actually Want; written version by Branwen Jeffreys and Hope Rhodes, 27 Aug 2026 — https://www.bbc.co.uk/news/articles/cn9n9n5p9eqo (NFER five-year study; six skills; up to 3m low/mid-level jobs at risk by 2035; just under 1m young people NEET)

• Institute of Student Employers (ISE), Student Recruitment Survey — 140 applications per graduate vacancy vs 38 in 2002–03; graduate hires −8% (2024/25) with −7% forecast [secondary]

• Prospects at Jisc, Early Careers Survey 2026 (~5,000 respondents), reported by People Management — 73% AI use (up from 55%); 57% CV/cover-letter editing; 89%/83%/62% unfairness asymmetry; 20% of recent graduates applied to 100+ roles — https://peoplemanagement.co.uk/article/1960804/three-quarters-students-graduates-use-ai-during-job-applications-study-finds [secondary]

• CV-Library, 2026 AI in Recruitment Survey (~500 recruiters, 1,000+ candidates) — 75% vs 21% CV-screening belief gap; 83% recruiter adoption vs 36% believing it works; 53% believe machine-rejected; 28% less likely to apply; 40% considered abandoning — https://cv-library.co.uk/recruitment-insight/ai-in-recruitment-survey [secondary]

• Ada Lovelace Institute — 45% of UK employers give no guidance on AI use in applications; 29% discourage without enforcing [secondary]

• Totaljobs / HR Grapevine (May 2026) — 47% of UK jobseekers interviewed by an AI system; 30% walked away; 75% want disclosure; 80% oppose AI making the final decision [secondary]

• EU AI Act (Regulation (EU) 2024/1689), Annex III 4(a): recruitment/selection AI is high-risk; Digital Omnibus (Regulation (EU) 2026/1744, in force 27 Jul 2026) defers Annex III high-risk duties to 2 Dec 2027; Article 50 transparency applies from 2 Aug 2026; Article 14 human oversight (rubber-stamp review insufficient); Article 26(11) duty to inform candidates [secondary]

• Eurofound — NEETs cost European economies ~€142bn a year (as cited in the Economics Explained documentary) [secondary]

• Economics Explained, “Why Youth Unemployment Is Worse Than It Looks” (15 Aug 2026) → 260815 NEETs The Hidden Youth Unemployment Crisis — 262m young people NEET globally / 20.4% NEET rate vs ~6% unemployment; one in five NEETs voluntarily disengaged; UK NEETs with work-limiting conditions 26% (2015) → 44% (2025); Japan’s hikikomori and the “80/50 problem”; flexicurity [secondary]

• 261012 Why Applying to 100s of Jobs Online Doesn’t Work Anymore — the Forbes Future of Work Panel — the 2026 market evidence, now consolidated into The Candidate’s Path — Getting In, and What It Is For (added 2026-10-12)

• The Candidate’s Path — Getting In, and What It Is For — the other half of one question: this note is how to enter the hiring market; the other is what to want from a life that includes it (2026-10-13)

Related

• Youth Disconnection and NEETs - Global Problem — the problem half, seeded the same day
• 260918 Everyone Using AI To Get Hired Now What — the triggering news note and the Parachute analysis
• 260826 The Six Skills Employers Actually Want
• 260815 NEETs The Hidden Youth Unemployment Crisis
• 260808 EU AI Act Enforcement Begins — the regulatory floor being built elsewhere
• Economic Instability — the demand side this vision does not fix
• Digital and Information Chaos — the opaque layer this vision makes answerable
• Gen Z Rising - The Global Problem
• Gen Z Rising - Skilling Path — the skills route into work
• Being AI Native - Competency Ladder — what candidates should be able to show instead of keywords
• Worker Cooperative — the higher-quality-employment half
• Humanist AI Manifesto — human-in-the-loop as a principle
• Diffusing the E-Bike Crisis - A Governance and Market Playbook — the sibling playbook pattern: levers, owners, honest limits
• Education Cradle to Grave - AI Personalised Learning
• Purpose of Alf — the thread this sits in
• Intentional News Consumption and Polycrisis