In 2007, a Nokia engineer demonstrated a prototype of a full touchscreen smartphone in an internal meeting. Management’s response was dismissive. The device was technically impressive but commercially implausible — or so the story went inside Nokia’s walls. Three years later, the iPhone had rewritten the industry, and Nokia’s decline was underway.
This wasn’t an intelligence failure. Nokia had the signal. The problem was what they did with it — how it was interpreted, assigned meaning, and ultimately suppressed. That is a sensemaking failure, and it is far more common, and more consequential, than organisations like to admit.
Karl Weick, the organisational psychologist who developed the theory of sensemaking in the 1990s, argued that organisations don’t simply discover reality — they construct it. Sensemaking is the process by which people convert ambiguous streams of experience into coherent accounts of what is happening and what to do next. Weick identified seven properties of this process, and three are especially relevant to how organisations fail to act on weak signals.
First: retrospect. Sensemaking is inherently backward-looking. People make sense of events after they happen, projecting a plausible narrative onto what has already occurred. This means that organisations are structurally inclined to interpret new signals through existing stories, not to let new signals revise the story.
Second: plausibility over accuracy. Weick was explicit about this. Sensemaking doesn’t seek the truth; it seeks a sufficiently workable account that allows action to continue. The Nokia story was plausible. Touch interfaces were a novelty. The company was strong. The existing mental model held. Accuracy came later, at significant cost.
Third: enactment. Organisations don’t passively observe their environment — they help produce it. The decisions companies make, the structures they build, the metrics they track — all of these shape what signals become visible and which ones stay in the noise. Nokia’s organisational structure, siloed between hardware and software, literally made it harder to see what was happening in software-driven consumer behaviour.
Weak signals are the precursors to major disruptions: faint, ambiguous, easily dismissed. They appear in customer complaint patterns, in supplier behaviour shifts, in forum threads, in the margins of quarterly reports. Individually, they are unremarkable. Together, they constitute a picture — but only if someone is assembling it.
This is precisely where AI is changing the calculus.
A well-designed AI system can hold dimensions of data in parallel that no human analyst can. It can identify correlations across a supplier’s invoicing patterns, a logistics provider’s delivery variance, a weather model, and a port congestion alert — and surface them as a single coherent risk signal before any individual analyst would notice each piece separately. Adidas reportedly used AI to map risk across 500+ suppliers and 300+ logistics providers, identifying that 47% of its athletic footwear depended on materials routed through just three high-risk ports. Acting on that signal, it redistributed its supply chain and avoided an estimated $135 million in potential disruption costs.
This is AI as an instrument of cue extraction — finding the signal before it becomes a crisis. It works by operating at a scale and speed that no human sensemaking process can match.
But here is the problem: extracting the signal is only half the work. The signal still has to enter an organisational sensemaking process. And that process has the properties Weick described — retrospective, plausibility-biased, resistant to uncomfortable information. An AI that surfaces a correct weak signal does not guarantee that anything changes if the human system around it is configured to suppress it.
Deloitte’s 2026 Human Capital Trends report makes the scale of this gap concrete. Just 5% of organisations consider themselves leaders in AI-enabled decision-making. Fifty-seven percent operate at low decision-making maturity. Seventy-two percent of executives say data volume and lack of trust prevent them from making decisions at all. This last number is striking: AI is producing more signals, at greater speed, and the result for most organisations is not better decisions — it is paralysis.
McKinsey’s 2025 State of AI report found that only 33% of senior leaders have at least a working understanding of how AI creates value in their business. Fifty-one percent of organisations using AI reported at least one significant negative consequence in the past year.
These are not technology problems. They are sensemaking problems. The AI is generating outputs. Leaders are not equipped to interpret them, weigh them against other signals, or integrate them into a coherent narrative that drives action. The data is there; the story is not.
The deeper disruption is harder to see, and Weick’s concept of enactment is the right lens for it.
When an AI system becomes part of an organisation’s decision-making infrastructure, it does not just process signals — it participates in producing the environment those signals come from. When a retailer’s AI recommends inventory redistribution, and the retailer acts on it, that action changes customer behaviour, which changes the data the AI receives, which shapes the next recommendation. The organisation and its AI are co-producing reality in a feedback loop that no single human decision-maker controls or fully understands.
This is genuinely new. Pre-AI, human analysts were slow enough, and few enough, that their interpretations rarely created fast feedback loops in external markets. AI systems operate at a different tempo and scale. In financial markets, this is already visible: algorithmic systems responding to signals they themselves partly created can produce oscillations that bear no relationship to underlying fundamentals. The same dynamic is beginning to appear in supply chains, hiring markets, and pricing systems.
The practical question for organisations is not just “can our AI detect weak signals?” It is “how does our AI change what signals exist, and are we governing that loop?”
The gap between detection and action has a name in some emerging practitioner writing: nudgment — a blend of nudge and judgment — defined as the organisational capacity to recognise, prioritise, and act on weak signals before they become obvious data. The term may be new, but the problem it names is not. A 2026 research paper on frontline safety systems (Codourey and Gonzalez, arXiv) makes the same observation from a different direction: weak signals only become organisational signals through structured group interpretation. The AI can identify that something anomalous is happening; the team has to construct what it means.
The distinction matters. Detection is the easy part now; AI handles detection well. Discernment — knowing which signals matter, in this context, for this organisation, at this moment — remains a human function. And the organisations that will navigate this well are not the ones with the most sophisticated models. They are the ones that have invested in the human judgment infrastructure that sits around the model: clear decision rights, escalation paths that do not filter out uncomfortable signals, leadership that understands AI outputs well enough to push back on them.
Liberty Mutual Insurance, for instance, explicitly designed its AI-enabled claims process so that adjusters can override AI suggestions — and tracked who overrides, when, and why. That is not a workaround for a weak model. It is a designed feedback mechanism. It creates a continuous signal about where AI judgment diverges from experienced human judgment, and it routes that signal back into model improvement. The AI learns; so does the organisation.
Weick argued that what collapses an organisation is often not the absence of signals, but the presence of a story so compelling that signals which contradict it get suppressed, reinterpreted, or simply not seen.
AI does not fix this. In fact, it can amplify it. An AI trained on historical data embeds the stories the organisation has already told about itself. If those stories are wrong — if they encode the assumption that film photography is the future, or that streaming is a niche market — then the AI will extract cues that confirm those stories and route anomalies to the noise bucket.
The organisations that use AI well will be the ones that deliberately configure their sensemaking systems to surface disconfirming signals — not just patterns that validate what leaders already believe. That means diverse data inputs, interpretive processes that involve people who hold different organisational identities, and leadership that has learned to treat an uncomfortable AI output as a prompt for inquiry rather than a system error to be explained away.
The signal is not the hard part anymore. The story is.
Sources: Karl Weick, Sensemaking in Organizations (Sage, 1995); McKinsey State of AI 2025; Deloitte 2026 Global Human Capital Trends; Everstream Analytics on AI in Supply Chain Risk; SCMR, “How AI helped a retailer prevent stockouts”; Codourey and Gonzalez, “The Weak Signal Cultivation Model” (arXiv, April 2026); B2BNN, “Nudgment: Why Signal Discernment Is the New Organisational Capability in the AI Era” (April 2026); Stanford HAI, 2025 AI Index Report.