Sandery Insights / Decision architecture

Research note 07 · Mitch Sandery · September 2026

The Job Description Is Not the Job

AI is becoming extraordinarily good at matching candidates to job descriptions. That would be more impressive if the CV were the person and the job description were the job.

Reading time11 minutesEvidence standardResearch + interpretation

Modern recruitment is converging on an elegant proposition: take a CV, compare it with a job description, rank the candidates and assess the strongest matches. The technology is increasingly sophisticated. The information architecture underneath it is not.

We are building extraordinarily sophisticated machinery for comparing two abstractions.

The job description was never designed to model the job.

A job description performs several functions at once. It advertises an opportunity, supports internal approval and classification, communicates employment conditions, gives recruiters searchable criteria and records responsibilities. Somewhere inside all of that, it tries to explain the work.

Those functions are not identical. Advertising favours aspiration. HR favours consistency. Search requires recognisable attributes. Legal and compliance teams may favour established language. A document can satisfy each purpose while remaining surprisingly weak as a specification of the problem the organisation needs solved.

Formal job analysis begins elsewhere. The U.S. Office of Personnel Management starts with the critical tasks actually performed, the competencies required to perform them and the relationship between the two—using people with current knowledge of the work.¹ U.S. selection guidance similarly requires job-relevant procedures to be grounded in important work behaviours and their associated tasks.²

Job-analysis logic

Actual work
Critical tasks
Required capability
Relevant assessment

Recruitment shorthand

Job description
Expected profile
Perceived fit
Interview

Employers often specify the person they expect, not the work they need.

Sandery interpretation

An investment-director vacancy might request eight to twelve years of experience, prior private-equity exposure, transaction execution, board experience, sector knowledge and an elite academic or employer history.

Almost none of that describes the job. It describes the expected biography of the person imagined performing it.

The underlying demand may be to evaluate ambiguous investments, identify which assumptions drive downside risk, challenge management forecasts, negotiate with founders and advisers, recognise when diligence has disproved the thesis and make an honest recommendation under time pressure.

Private-equity experience may provide strong evidence for those requirements. Banking, corporate development or operating experience may provide different subsets. The biography is informative. It is still a proxy.

Expected biographyWhere should this person
have been before?

Actual demandWhat must this person
decide, produce or change?

A written requirement can be neither necessary nor predictive.

Empirical evidence

Harvard Business School, Accenture and Grads of Life analysed more than 26 million job postings. In 2015, 67% of production-supervisor vacancies requested a college degree, while only 16% of people already performing the occupation possessed one—a 51 percentage-point gap.³

Employers surveyed for the research often reported that experienced workers without degrees performed nearly or equally well across critical dimensions. The degree remained attractive because it was a legible proxy for the capabilities employers wanted.

This is more than a story about unnecessary credentials. It shows that organisations can state a requirement confidently, repeatedly and at market scale without that requirement being necessary to perform the work.

67%Vacancies requesting a degree

Production-supervisor postings in 2015.

HBS, Accenture and Grads of Life

16%Incumbents holding a degree

People already performing the occupation.

HBS, Accenture and Grads of Life

51ppRequirement gap

Written demand versus observed workforce.

HBS, Accenture and Grads of Life

Changing the advertisement does not necessarily change the hiring system.

Empirical evidence

Researchers from the Burning Glass Institute and Harvard Business School examined more than 11,000 roles at large U.S. employers that removed degree requirements.⁴ In affected roles, the share of workers hired without a bachelor's degree increased by about 3.5 percentage points.

Across the companies studied, however, the additional opportunity amounted to fewer than one in every 700 hires. Around 45% of firms showed no meaningful change in actual hiring behaviour after removing the requirement.

The words changed. The selection pattern often did not.

Our interpretation: the operative hiring specification also lives in recruiter pattern recognition, hiring-manager expectations, interview behaviour, organisational precedent and risk tolerance. The vacancy is not always a reliable account of how the employer will decide.

+3.5ppNon-degree hiring

Average change within affected roles.

Burning Glass Institute + HBS

45%No meaningful change

Share of firms where hiring behaviour did not shift.

Burning Glass Institute + HBS

<1 in 700Overall hires affected

Estimated additional opportunity across all hires.

Burning Glass Institute + HBS

The same title can conceal radically different work.

Jobs cannot be reduced to universal lists of attributes. Deming and Kahn found substantial variation in advertised skill requirements between employers recruiting for narrowly defined professional occupations.⁵ A study using a large corpus of executive-search specifications likewise found that C-suite demand combines cognitive, interpersonal and operational requirements in different ways.⁶

A CFO in a stable listed business is not the same requirement as a CFO facing covenant pressure in a sponsor-backed turnaround. A strategy director entering new markets is not the same requirement as one rationalising a mature portfolio.

Title matching suppresses that context. The more senior and situational the role, the more consequential the loss becomes.

A title names the position. Context determines the work.

AI can produce the form of an excellent vacancy without the organisational truth.

Ask a capable language model for a VP of Strategy job description and it will produce fluent, plausible and comprehensive prose. Strategic planning. Market analysis. M&A. Transformation. Executive communication. Sensible qualifications. The result may be immediately publishable.

It may still reveal almost nothing about why this company needs this person now. What is broken? Which decisions will transfer to the role? Where has execution failed? What is politically difficult? What would make an apparently strong hire unsuccessful?

A 2026 Personnel Review study ran three experiments comparing AI-generated, expert-written and real job advertisements. Applicants perceived the AI-generated material as broadly as effective and informative as human-written advertisements.⁷

That demonstrates the technology's command of the form. Perceived informativeness is not the same as fidelity to the underlying work.

The danger is not bad AI writing. It is excellent AI writing applied before the job has been understood.

Skills without context risk becoming the new keywords.

The movement toward skills-based hiring is directionally useful. But broad labels can reproduce the same abstraction under more modern language.

Stakeholder management might mean negotiating with a regulator, managing a hostile lender through restructuring, persuading a founder to sell, navigating an executive committee or leading a works-council consultation. The label is identical. The work is not.

The same applies to AI. Indeed found that U.S. job titles referencing AI more than tripled between early 2022 and early 2026, spreading well beyond technology occupations.⁸ Some of that reflects genuine changes in work. Some may reflect signalling. “AI strategy” still does not tell us whether the person must redesign workflows, govern models, implement systems, evaluate vendors or identify economic use cases.

A taxonomy is useful. It is not evidence. Serious analysis connects a capability to the decisions and behaviours through which it matters.

The useful unit of analysis is closer to the decision than the title.

Sandery interpretation

“Strategic thinking” tells us very little about a strategy director. The demand becomes more intelligible when it is expressed as decisions and outcomes.

Which markets should the company exit? Should it acquire a competitor? Which categories deserve investment? Where can margins improve structurally? How should capital be allocated between organic and inorganic growth? What must change across six countries to make the decision executable?

Those questions make evidence examinable. Has the person evaluated markets, built investment cases, challenged operating assumptions, influenced executives and implemented decisions after the analysis ended? Have they done so under comparable constraints?

The title becomes context. The capability label becomes a hypothesis. The underlying experience becomes evidence.

Low-resolution demandTitle · tenure · sector
qualification · skill label

High-resolution demandDecisions · outcomes · context
constraints · uncertainty

A requirement should disclose what kind of requirement it is.

Vacancies frequently collapse necessity, preference and precedent into one undifferentiated list. Automated systems can then treat a hiring manager's preference as an objective feature of the job.

Some criteria are genuinely necessary. Others are strong but substitutable evidence. Some provide useful context. Others simply describe the profile historically selected.

Years of experience illustrates the problem. Two people can each possess ten years in the same occupation while encountering radically different responsibility, complexity, repetition and development. The variable is easy to process. The evidence underneath it is not.

Requirement types

Necessary on entry
Strong evidence preferred
Useful context
Historical pattern

The more measurable a proxy is, the easier it becomes to mistake it for the thing being measured.

The job needs a system of record beneath the advertisement.

The Sandery view

Better matching begins before the candidate enters the system. A perfect representation of a person compared with a weak representation of the opportunity still produces a weak match.

The alternative is not a longer job description. It is a richer demand specification beneath it: why the role exists, the outcomes that matter, the decisions it owns, relevant context and constraints, what must be known on entry, what can be learned, where evidence could transfer and where the organisation remains uncertain.

From that underlying intelligence, an organisation could generate a candidate-facing advertisement, recruiter brief, sourcing specification, interview scorecard and assessment plan. The documents could differ without pretending any one of them is the job.

This is a direction for the wider architecture—not a claim that Sandery has already solved it. Sandery's current starting point is the individual: creating a persistent evidence base beneath the CV. The same logic points, over time, toward better intelligence beneath the opportunity.

Candidate side

Career evidence
Professional intelligence
Relevant representation
Targeted validation

Opportunity side

Actual work
Demand intelligence
Relevant specification
Targeted assessment

The mirror-image problem

The CV compresses the person. The job description compresses the opportunity.

The existing market asks candidates to decide which fragments of a career will matter, employers to decide which fragments of a job will predict success, and recruiters and algorithms to infer everything lost between them.

That architecture was rational when information was expensive and interpretation was slow. AI changes the constraint. It can process more evidence, retrieve semantic relationships and generate the appropriate representation for a specific decision.

Much of the market is using that capability to optimise the compressed documents instead.

The provocation

An AI writes the job description.
Another rewrites the CV.
A third scores their similarity.

The technology may be modern. The architecture is not. The opportunity is to compare career evidence with actual work demand—not to automate the approximation.

Start beneath the document

Build the evidence before the representation.

Explore the Career Bank →Explore Job Decompiler →

References

Evidence used in this note.

The distinction between a job advertisement and a demand specification, the requirement taxonomy and the proposed evidence-matching architecture are Sandery interpretation. The cited research demonstrates that stated requirements can diverge from incumbent profiles, changing vacancy language may not change hiring behaviour, role requirements vary materially by employer and AI can reproduce the form of an informative advertisement. It does not establish that every job description is inaccurate or every AI matching system is ineffective.

The job description was never designed to model the job.

Employers often specify the person they expect, not the work they need.

A written requirement can be neither necessary nor predictive.

Changing the advertisement does not necessarily change the hiring system.

The same title can conceal radically different work.

AI can produce the form of an excellent vacancy without the organisational truth.

Skills without context risk becoming the new keywords.

The useful unit of analysis is closer to the decision than the title.

A requirement should disclose what kind of requirement it is.

The job needs a system of record beneath the advertisement.

The CV compresses the person. The job description compresses the opportunity.

An AI writes the job description.Another rewrites the CV.A third scores their similarity.

Build the evidence before the representation.

Evidence used in this note.

Research & institutional evidence

An AI writes the job description.
Another rewrites the CV.
A third scores their similarity.