How to quote these findingsIf you quote a finding, quote its limit. Percentages describe the coded sample only.

Check a figure

Findings

What the evidence currently supports

Four current governed findings sit above the frozen Report 01 and Report 02 findings. Every quantitative finding retains its reviewed subset, denominator and limit.

New: 36 of 67 selective employers checked publish an official statement on AI in selection; 18 prohibit AI in at least one online test or interview, and none allows unrestricted AI in interviews. 11 disclose their own AI use in selection. AI Rules Tracker v2.0, observed 23 September 2026.

Current governed findings reviewed through 13 September 2026; foundational report findings remain preserved against their original publication snapshots.

Evidence behind them

1004current governed selection signals

383official source records in the current canonical provenance layer

4current governed comparative findings

Current governed findings

Four findings derived from defined reviewed subsets

Current canon1004 dated facts (signals)
383 official source records
4 governed findings

Machine-readable findings

01
Explicit AI rules protect candidate-owned performance.
All 8 institutions in the current reviewed official-source AI-policy subset publish an explicit boundary or control around candidate or employer AI use; 7 explicitly protect at least one assessment or interview context through an AI restriction or candidate-owned-performance requirement.
Denominator: 8 institutions in the defined AI-policy subset reviewed through 13 September 2026. Do not generalise this defined subset to all selective employers.Evidence snapshot.
02
Human judgement remains visible after digital selection stages.
All 7 institutions in the reviewed subset with explicit post-digital or hiring-decision evidence retain stated human judgement, human review, live assessor interviews or assessment-centre/Super Day decision stages.
Denominator: 7 institutions in the defined human-review subset reviewed through 13 September 2026. This describes the reviewed subset only.
03
The same employer can run materially different selection architecture by market.
Each of the three same-employer multi-market comparisons reviewed - PwC, Clifford Chance and BlackRock - contains a material geography-specific difference in published selection architecture, timing or route design.
Denominator: 3 same-employer multi-market comparisons reviewed 12 September 2026. This is not a claim that every multinational employer varies by geography. Five-hub comparison.
04
A single BlackRock programme gives London and GCC candidates different application runways.
BlackRock's 2027 Client and Product Summer Internship Programme publishes a 30 September 2026 deadline for London and a 31 January 2027 deadline for Abu Dhabi and Riyadh, while stating that applications are reviewed on a rolling basis.
Denominator: one directly comparable current EMEA programme observed 12 September 2026. Programme-specific only; do not generalise to other BlackRock programmes or employers. Governed finding WGI-FIND-004.

Frozen publication layerThe findings below belong to Report 01 and Report 02 and remain reproducible against the original 107-signal publication snapshot.

Foundational report findings

These report-level findings are preserved as historical publication outputs. They do not replace the four current governed findings above, and their denominators remain the frozen report samples stated alongside them.

Report 01 · frozen publication finding set

Ownership, judgement and the auditable claim

Sample57 signals
48 cited sources
23 institutions
Checked 30 August 2026

01
Selection includes assessment the candidate does not write at leisure.
Verified examples include a timed cognitive assessment at KPMG, an assessment centre with a written and a group exercise at PwC Middle East, a case discussion at Strategy& Middle East, essays written on the interview day at the Indian School of Business, an Admissions Written Test before interview at NUS Law, and an in-person Case Challenge at DBS. No proportion is claimed: the inclusion rule and row-level review are not complete.
This does not establish that live assessment is new, increasing or superior, and says nothing about prevalence beyond the observations coded.
02
A change of medium can make a written claim easier to test.
A claim written at leisure becomes harder to imitate when it must be explained, adapted, calculated, debated or defended under time pressure.
This does not establish that any format predicts performance on the job.
03
The AI rules observed concern ownership, not blanket prohibition.
Six observations were coded newly explicit about AI. They come from three institutions: DLA Piper, UBS and Cambridge, which permit forms of preparation while protecting independent assessment and truthful representation.
This does not establish that AI caused institutions to introduce live assessment. Fifty of 57 recorded signals were coded as existing processes at the review date.
04
Process design creates unequal translation burdens.
Thirteen signals map stage architecture or geographic variation. Candidates must transfer the same underlying evidence across different formats, sequences and markets.
This does not establish the size of any resulting advantage, or who it accrues to.

Report 02 · frozen publication finding set

Convergent verification, divergent architecture

Cohort50 signals
8 or 9 per market
21 official sources
Checked 31 August 2026

01
Verification is universal in the cohort.
All six markets include a published mechanism that checks more than bespoke prose.
This does not establish prevalence in any country or sector. Two to three institutions were reviewed per market.
02
The verification instrument is local.
Supervised retesting, on-the-spot writing, authenticated school work, case challenges and language tests perform different versions of the same function.
This does not establish that one instrument is more predictive than another.
03
Transparency varies materially.
Indian institutions in the sample publish staged formulas and weights; other markets publish sequences but not the final decision calculus.
This does not establish that undisclosed calculus is absent, only that it is unpublished.
04
Access to the live gate is itself designed.
Shortlisting, programme preference, prior tests and invitation-only events determine who is allowed to demonstrate judgement.
This does not establish intent, fairness or outcome effects.
05
AI changes ownership controls, not the need for judgement.
The clearest new design choice is separating AI-mediated presentation from human scoring or independent performance.
This does not establish that these controls work, or that they are widely adopted.

Why this is hereA research programme that only publishes what it can prove has to publish the boundary too.

Boundary

It does not claim that AI caused institutions to introduce live, interactive or whole-person assessment. Elite selection already combined written applications with live assessment before generative AI became widely available; what AI has made more explicit is the question of who owns the thinking in an application.

It does not rank markets or institutions. It does not estimate selectivity, acceptance odds or candidate outcomes. It does not infer fairness or institutional intent beyond the stated published process. It cannot establish prevalence, and it makes no market-wide claim that is not triangulated across institutions and sectors.

ThresholdReport 03 will emerge from evidence, not from a publishing schedule.

How an observation becomes a published finding

An observation enters the Monitor as one narrowly worded row against one official primary source, with a source ID, URL, confidence grade and review date. It becomes a published finding only when it is supported across enough institutions and sectors to survive the stated limits, and it is published with those limits attached.

Market-wide claims require triangulation. Historical trend claims require longitudinal evidence. Anything that does not clear the threshold stays a hypothesis in the register and is not published.

Open the register

Selection architecture

How the coded stages fit together

Conceptual, not statisticalA taxonomy of the stages evidenced across both coded samples. It orders published architecture; it does not measure a candidate pathway.

Exhibit 1Access to the live gate is itself designedShortlisting, programme preference, prior tests and invitation-only events determine who is permitted to demonstrate judgement at all. The gate sits before the judgement, not after it.

Candidate claim

What the candidate asserts

Written application
Personal statement
Curriculum vitae
Self-reported experience

Access gate

Who is permitted to proceed

Shortlisting
Programme preference
Prior tests
Invitation-only events

Verification

What is checked against record

Prior record checks
Test or structured assessment
Authenticated school work
Language test

Live judgement

What is produced under observation

Live writing
Case or task
Interview
Supervised retesting

Institutional decision

Who decides, and on what

Panel assessment or committee review
Published outcome

Boundary

This does not establish intent, fairness or outcome effects. The tension between designed access and demonstrated judgement is identified here, not measured.

A measured candidate pathway. The ordering describes published architecture. Not every institution runs every stage, and stages are not necessarily sequential in practice.

Source Global Elite Selection Monitor. A conceptual taxonomy, not a statistical claim. Stages list only categories evidenced across Report 01 (57 signals) and the Report 02 cohort (50 signals).

How to citeKeep the scope and evidence dates with the finding.

How to cite Who Gets In?

Suggested citation: Akram, H., Who Gets In? (Elite Careers Strategy), [finding, tracker or report title], [date stated on the finding], [page URL]

Cite the specific finding or register rows where possible. Percentages and denominators describe the defined sample only, not all elite employers or universities.

Permanent citation objects

Each finding has a stable URL, denominator, source chain and machine record.

This page: revised 13 September 2026. Corpus: 1004 dated facts (signals) from 383 official source records, last canonical update 27 September 2026. Release notes