Narrative Artifacts and the Illusion of Precision
In the first article in this series, I suggested that AI in talent has crossed a threshold. It is no longer merely accelerating workflow. It is increasingly shaping decisions that determine access to work, advancement, and performance consequence. When systems begin to intermediate judgment, the standard of rigor required of the underlying inputs changes.
That raises a necessary question.
If AI systems are increasingly operationalizing talent decisions, what exactly are they optimizing against?
In most organizations, the answer remains consistent: narrative artifacts.
Job descriptions.
Resumes.
Performance summaries.
Succession profiles.
These artifacts were not designed to function as deterministic decision engines. Yet they are increasingly being treated as though they were.
Compression Was a Feature, Not a Flaw
To understand the structural weakness, we need to understand the original design intent.
Job descriptions and resumes are compression devices. They condense complex realities into formats that can be reviewed quickly and communicated broadly. A job description distills a business need into a postable format. A resume distills years of experience into a manageable narrative. A performance review summarizes a year of activity into evaluative prose.
Compression was necessary in a human-limited system.
Andrew Gadomski (HR and AI governance expert, Managing Director, Aspen Analytics) captured this succinctly when he noted that a three-page job description is a poor proxy for 2,000 hours of annual work, just as a two-page resume is a poor proxy for a career. These artifacts existed because proxies were required.
The issue is not that these artifacts are inherently flawed. The issue is that the constraints under which they were designed have changed.
We now possess the computational capacity to ingest structured dialogue, transcripts, task simulations, and performance signals at scale. Yet in many cases, we continue to treat compressed narrative artifacts as if they represent the full dimensionality of role performance.
Compression was efficient for communication. It is insufficient for constraint engineering.
Role Story Versus Role Structure
To make this distinction clearer, it is helpful to move beyond the language of “good” or “bad” job descriptions and instead examine structural design.
Most job descriptions tell a Role Story. They describe responsibilities, qualifications, and generalized expectations. They often blend tasks with aspirations. They frequently emphasize credentials and activities more than measurable outcomes.
What they rarely provide is Role Structure.
Role Structure includes:
- Defined time-bound outcomes
- Observable success criteria
- Leading indicators of performance
- Resource constraints and tradeoffs
- Stakeholder dependencies
- Decision rights and autonomy
Role Story is narrative. Role Structure is architecture.
When AI systems are asked to rank candidates, generate interview guides, or assess fit based primarily on Role Story, they are optimizing against narrative abstraction rather than structured performance intent.
Ravi Subramanian observed that the richest hiring context frequently resides in the mind of the hiring manager rather than in the formal job description. That gap between implicit knowledge and explicit structure becomes consequential when automation formalizes only what is written.
The model executes faithfully. The misalignment emerges because the artifact does not encode the full architecture of the role.
Expansion Beyond Hiring
While hiring provides the most visible case study, the structural issue is not confined to it.
Promotion decisions are often justified through narrative summaries that vary widely in rigor and structure. Performance reviews combine measurable outcomes with subjective commentary. Succession planning artifacts frequently describe readiness aspirationally rather than through calibrated performance thresholds.
Abhishek Shah’s observation that we are automating decisions off narrative artifacts rather than observable capability data is particularly salient here.
When AI is introduced into promotion modeling, succession analytics, or performance flagging, it operates on whatever structure exists. If the underlying artifacts are unsegmented blends of narrative and signal, the system learns from that blend.
The result is not necessarily bias. It is ambiguity formalized.
The distinction matters. Bias implies distortion. Ambiguity implies insufficient structure.
Automation scales both.
The Illusion of Increased Precision
One of the more subtle consequences of automating narrative artifacts is the emergence of perceived objectivity.
A ranked list generated by an AI model appears more systematic than a manual review. A structured interview guide produced algorithmically appears more consistent than ad hoc questioning. A predictive promotion score appears more analytical than managerial intuition.
However, precision in output does not guarantee precision in constraint definition.
Gary Lear (Expert on the Leadership & Dynamics of High Performing Organizations) pointed out that most job descriptions are not grounded in validated job task analysis. If the constraints embedded in the artifact have not been empirically validated, then the model’s precision operates within unvalidated boundaries.
Similarly, Jonathan Duarte raised the possibility that AI-generated resumes increasingly optimized to match AI-assisted job descriptions could create a self-referential matching loop (I bring this up everytime I hear a TA leader mention the overwhelming volume of applicants they receive.) In such a system, alignment between two artifacts may increase, while alignment with actual performance requirements remains untested.
This is what I mean by the illusion of precision.
The system becomes better at matching text to text. That does not automatically mean it becomes better at matching capability to outcome.
Structural Segmentation as a Counterpoint
The alternative is not abandoning automation. It is segmenting architecture intentionally.
Adam Hopewell (Global HR System Architect) described an approach in which hiring architecture is separated into calibrated streams, such as knowledge and skills alignment, applied task simulation, and values alignment, each correlated against defined achievement targets.
That approach reflects an important structural shift.
Instead of treating the job description as a monolithic narrative input, it decomposes performance into measurable components and correlates those components with observable outcomes.
In that environment, automation operates within defined, testable boundaries.
Without segmentation, automation operates within blended narrative abstractions.
The distinction is architectural, not technological.
Velocity Changes the Stakes
It would be possible to tolerate artifact ambiguity in a slow-moving system where human judgment dominates and scale is limited. That is not the environment we are entering.
AI compresses cycle time. It reduces friction. It increases the number of decisions processed within a given period.
Ravi Subramanian drew a parallel to earlier compliance inflection points in which organizations discovered that velocity without structural clarity generated liability. While the regulatory landscape differs today, the structural lesson holds: when decision volume increases, the integrity of constraint definition becomes more consequential.
If automated systems influence who is screened out, promoted, or flagged as underperforming, the defensibility of those decisions depends on the structural soundness of the underlying criteria.
If those criteria originate in artifacts designed primarily for communication rather than governance, exposure increases.
From Artifact Reliance to Architectural Intent
The solution is not to perfect every job description. It is to recognize that narrative artifacts and decision architecture serve distinct functions.
Narrative artifacts communicate intent externally and internally.
Decision architecture defines constraints, calibrates expectations, and anchors accountability.
If AI systems are increasingly acting within the domain of governance rather than mere coordination, then talent leaders must examine whether the architecture beneath them is sufficiently explicit.
That examination includes:
- Separating activities from outcomes
- Defining measurable success criteria
- Establishing time-bound achievement targets
- Calibrating expectations across managers
- Differentiating observable capability from narrative commentary
When those elements are structured, automation can reinforce alignment.
When they are absent, automation reinforces abstraction.
What’s next? Governance responsibility, specifically, we will examine what auditability, drift monitoring, and explicit decision ownership require in an environment where AI intermediates talent decisions.
