Cue Extraction
Organizations are saturated with available cues — far more data than anyone could possibly process. Sensemaking proceeds by extraction: the selective noticing of certain signals and the corresponding unnoticing of others. The extracted cues become the raw material of interpretation. What gets extracted depends on identity, on prior frameworks, on institutional priorities, on the affective tone of the moment, and on the sheer chance of where attention happens to land. Cue extraction is the mechanism through which the same situation produces radically different interpretations for different observers, and through which weak signals — the small, ambiguous, inconsistent cues that precede catastrophic failure — either get noticed and incorporated into organizational action or get filtered out by the prevailing narrative. In AI-augmented organizations, cue extraction is reshaped by the tool's own extraction patterns, which amplify certain signals and suppress others in ways that are structurally invisible to the users whose attention the tool is now mediating.
The Extraction Hierarchy Problem — Contrarian ^ Opus
There is a parallel reading that begins from how cue extraction has always been a site of organizational power, not merely cognitive selection. The question is not whether AI changes what gets noticed, but who controls the extraction apparatus and to what end.
The expert clinician's superior extraction is not just trained perception—it is credentialed authority to define which signals count as relevant. The Bristol case is not simply about cues being invisible to a dominant framework; it is about whose extraction patterns have institutional standing. The nurse's private tally, the anesthesiologist's concerns—these were extracted cues that lacked the power to become organizational cues. When we frame this as a sensemaking problem, we obscure that it was fundamentally a power problem: certain observers' extractions were structurally inadmissible.
AI does not merely mediate extraction invisibly; it industrializes extraction hierarchies. The training data encodes whose historical extractions mattered enough to become recorded knowledge. The deployment decisions encode whose current extractions will be amplified by computational scale. The interface design encodes which extraction patterns users can even perform. What appears as the tool's 'own biases' are actually congealed social choices about whose attention patterns deserve algorithmic reproduction. The weak signal that gets filtered out is weak not because it is small or ambiguous, but because the people who would extract it lack the institutional position to have their attention patterns trained into the model.
In the AI Story
The concept operationalizes what Weick meant by attention in organizations. Cues are not sense-data; they are the small subset of sense-data that an observer with a particular identity, history, and framework actually registers as relevant. The extraction happens below the level of conscious deliberation — which is why cue extraction can be systematically trained (the expert clinician who sees what the novice misses) and systematically distorted (the manager whose prior commitments make the disconfirming signal invisible).
The organizational catastrophes Weick studied are cue-extraction failures. At Bristol Royal Infirmary, the elevated mortality data existed but was not extracted as a cue about surgical competence; it was extracted instead as a cue about case complexity. The nurse who kept a private tally, the anesthesiologist who raised concerns, the pathologist who saw patterns — each of these was a parallel extraction that did not penetrate the organizational sensemaking because the dominant framework had no room for their cues.
AI changes cue extraction in two directions. It expands what can be extracted — language models surface patterns in data that humans could not process unaided. It also narrows what does get extracted — the tool's own training biases and output formats channel attention toward certain cues (the ones the tool is good at extracting) and away from others (the anomalies that do not match any pattern in the training data). The weak signal that a practitioner with decades of embodied experience would extract as meaningful — the one that registers as feeling rather than as data — is precisely the kind of cue that AI-mediated attention systematically filters out.
Segal's observation that his senior engineer in Trivandrum realized judgment was the twenty percent that mattered is, in Weick's terms, a claim about cue extraction. Judgment is the capacity to extract cues that the specification does not mention. It is built through the friction of years of implementation work, during which weak signals repeatedly force reinterpretation. When the implementation work is done by AI, the developmental pathway through which cue extraction is trained is bypassed, and the next generation of practitioners inherits the capacity for output without the capacity for extraction.
Origin
Weick developed the concept across his corpus, most explicitly in Sensemaking in Organizations (1995) and in the Mann Gulch paper (1993). It draws on Herbert Simon's bounded rationality and on the ecological psychology of J. J. Gibson, whose concept of affordances Weick generalized into his account of how organizations perceive their environments.
Key Ideas
Selection is the mechanism. Organizations do not perceive situations whole; they perceive the cues their extraction patterns surface.
Identity shapes extraction. Who you understand yourself to be determines what you notice; the KLM captain at Tenerife extracted cues consistent with his identity as a senior commander.
Weak signals are systematically disadvantaged. Small, ambiguous, inconsistent cues rarely survive the extraction process unless a practitioner with domain expertise insists on their relevance.
AI mediates extraction invisibly. The tool's own biases shape what users notice without announcing themselves as biases.
Expertise is trained extraction. Domain experts see what novices miss because they have learned, through friction, to extract cues the novice filters out.
Appears in the Orange Pill Cycle
Cognitive and Political Extraction — Arbitrator ^ Opus
The cognitive account is fully right (100%) that extraction is selective, identity-dependent, and trainable through friction—these are empirical facts about how human attention works in complex environments. The power account is also right (100%) that extraction capacity has always been institutionally distributed—credentialing, hierarchy, and documentation norms determine whose extractions become organizational knowledge. The question is which lens clarifies what.
For understanding individual expertise development, the cognitive frame dominates (80/20). The senior engineer's judgment capacity is genuinely built through years of encountering weak signals that force reinterpretation; this is a learning process, not just a credentialing process. For understanding organizational catastrophes, the weighting shifts toward 50/50. The Bristol mortality data was both cognitively filtered (existing frameworks made it hard to see) and politically inadmissible (certain observers lacked standing to make their extractions count). Both mechanisms operated simultaneously.
For understanding AI's impact, we need a synthetic frame: AI as extraction infrastructure. The technology does mediate attention through trained biases (cognitive), and it does encode institutional choices about whose patterns matter (political). The developmental concern is real—bypassing implementation work removes the friction that trains extraction. The power concern is equally real—deploying extraction at scale amplifies whatever hierarchies the training data contains. The concept itself benefits from recognizing that cue extraction has always been both a perceptual capacity and an institutionally situated practice, and AI transforms both dimensions while making their entanglement harder to distinguish.
Further reading
- Weick, K. E. (1995). Sensemaking in Organizations, ch. 3.
- Weick, K. E. & Sutcliffe, K. M. (2001). Managing the Unexpected.
- Simon, H. A. (1971). Designing organizations for an information-rich world.