en flag
en flag

The Human Context Layer for the Age of AI

Why the next frontier of enterprise intelligence may be understanding people, groups, environments and context—not simply generating more content

8/15/202628 min read

A close up of a blue eyeball in the dark
A close up of a blue eyeball in the dark

Why the next frontier of enterprise intelligence may be understanding people, groups, environments and context—not simply generating more content

Executive perspective

Artificial intelligence is entering a different phase of economic adoption. The first phase was principally about expanding machine capability: prediction, classification, search, recommendation, language generation, coding, image and video synthesis, and increasingly sophisticated multimodal reasoning. The emerging phase is about embedding those capabilities into the operating fabric of organizations. AI is moving from applications that answer questions toward systems that participate in workflows, influence decisions, interact continuously with customers and employees, coordinate with other software, retain organizational memory and, increasingly, execute actions through agents. Adoption is already broad. McKinsey's 2025 global survey found that 88 percent of respondents reported regular AI use in at least one business function and 62 percent said their organizations were at least experimenting with AI agents. Yet nearly two-thirds had not begun scaling AI across the enterprise, and only 39 percent reported enterprise-level EBIT impact from AI. Stanford's 2025 AI Index reported a similarly rapid diffusion: organizational AI use increased from 55 percent in 2023 to 78 percent in 2024, while global private investment in generative AI reached $33.9 billion in 2024, an increase of 18.7 percent from the previous year. The combination is important. AI capability is diffusing rapidly, but broad availability of capability has not translated automatically into enterprise-scale economic performance. (McKinsey & Company, 2025; Stanford HAI, 2025)

The gap between adoption and realized value suggests that access to increasingly capable models is becoming less differentiating on its own. As high-quality models become available through commercial platforms, enterprise software and open ecosystems, competitive advantage increasingly depends on what surrounds the model: proprietary organizational knowledge, workflow redesign, governance, decision rights, high-quality memory, integration with operating systems, human judgment and the ability to interpret the context in which a decision occurs. The labor-market implications reinforce the point. McKinsey has estimated that, under a midpoint adoption scenario, generative AI and other technologies could automate activities accounting for approximately 30 percent of hours currently worked in the United States and 27 percent in Europe by 2030. The World Economic Forum's Future of Jobs Report 2025 estimates that structural labor-market transformation could affect 22 percent of today's jobs by 2030, with 170 million roles created and 92 million displaced, producing a net increase of 78 million jobs even as approximately 39 percent of workers' existing skill sets are expected to be transformed or become outdated. These are projections rather than predetermined outcomes, but their direction is clear: AI transformation is increasingly an organizational and human transformation, not simply a technology deployment. (McKinsey Global Institute, 2023; World Economic Forum, 2025)

This transition exposes an important weakness in the current enterprise AI stack. Modern models are increasingly sophisticated at processing information, but enterprises still possess comparatively crude machine-readable representations of the human and social environments in which that information acquires meaning. An AI system can summarize a meeting transcript with remarkable fluency while having far less reliable means of representing whether the meeting involved productive disagreement, unresolved status conflict, cultural deference, fatigue, uncertainty, coalition formation or recovery after an earlier dispute. It can identify declining customer engagement without necessarily distinguishing dissatisfaction from seasonality, economic pressure, changed need or a temporary behavioral shift. It can observe reduced communication inside a team without knowing whether the reduction represents efficient coordination or deteriorating engagement. The problem is not simply that models require more data. It is that the same observable signal can carry materially different meanings depending on the individual, relationship, group, environment, culture, history and moment in which it occurs.

A new category of infrastructure may therefore be required as AI moves deeper into organizations: Human Context Infrastructure. The term describes a structured representational layer for human states, observable signals, interpersonal relationships, group dynamics, organizational conditions, cultural context, environmental circumstances, crises, communication patterns and other dimensions of social reality relevant to machine-assisted decisions. Its purpose is not to establish a universal psychological theory, automate mental-state diagnosis or give machines privileged access to what people are thinking. The more defensible proposition is narrower and potentially more valuable: AI systems participating in consequential human environments need explicit ways to represent multiple plausible contexts, preserve uncertainty between competing interpretations, distinguish observation from inference, and understand that behavior cannot be interpreted independently of the conditions surrounding it.

This distinction is fundamental. A context architecture should describe what may be relevant; an inference system should evaluate evidence; a predictive system should estimate possible outcomes; a decision system should assess possible actions; and governance should determine what the system is permitted to do. Collapsing those functions into a single opaque model creates the illusion that because a machine can generate a psychologically plausible interpretation, that interpretation has been established as fact. Keeping them separate creates a substantially more disciplined architecture in which observations can remain observations, hypotheses can remain hypotheses, and uncertainty can remain visible. As AI shifts from machine capability toward machine participation, that separation may become one of the most important design requirements for trustworthy enterprise intelligence.

1. AI capability is becoming abundant; reliable organizational context remains comparatively scarce

For much of the modern AI era, advanced computational capability itself was scarce. Organizations competed for machine-learning talent, proprietary models, training data, infrastructure and specialized compute. Foundation models have materially changed that economics. Language generation, document analysis, coding, image interpretation, retrieval and increasingly sophisticated reasoning are now accessible through commercial APIs, enterprise platforms and open-weight ecosystems. Performance differences between models remain economically meaningful, particularly in specialized and high-stakes domains, but the broader direction is toward increasing availability of baseline machine intelligence. The strategic bottleneck consequently begins to migrate from whether an organization can access AI toward whether it can integrate AI into the specific institutional environment where value is created.

The organizational evidence already points to this shift. Microsoft's 2025 Work Trend Index reported that 53 percent of leaders believed productivity must increase, while 80 percent of the global workforce said they lacked sufficient time or energy to perform their work. The same research anticipated that employees would increasingly be expected to work with and supervise agents, with 41 percent of leaders expecting teams to train agents and 36 percent expecting teams to manage them within five years. Deloitte's 2026 Global Human Capital Trends research provides a related signal from the organizational side: 85 percent of leaders considered organizational and workforce adaptability critical, but only 7 percent believed their organizations were leading in continuously developing that adaptability, while just 6 percent reported making significant progress in designing human–AI interactions. The individual statistics come from different studies and methodologies and should not be combined into a single causal claim, but collectively they describe a recognizable enterprise condition: machine capability is advancing faster than many organizations' ability to redesign the human systems around it. (Microsoft, 2025; Deloitte, 2026)

This imbalance matters because organizations are not simply collections of tasks waiting to be automated. They are social systems composed of people with different capabilities, incentives, histories, responsibilities and levels of authority; teams with different norms and conflict patterns; institutions with formal and informal decision structures; and environments affected by culture, regulation, economic conditions and external shocks. An agent optimizing a workflow without understanding those conditions can improve a local process while degrading the larger system. A customer-service system that maximizes handling speed can damage trust. A workforce tool that maximizes task allocation can increase overload. A meeting assistant that treats silence as agreement can misrepresent decisions in a high-hierarchy environment. A leadership system that interprets low communication as disengagement may completely misunderstand a highly autonomous team. In each case, the computational problem is not principally a lack of intelligence. It is a lack of sufficiently structured context.

Human Context Infrastructure addresses that gap by treating contextual information as a first-class enterprise resource rather than an incidental prompt variable. At its broadest, such an architecture would distinguish the state of an individual from observable behavioral signals; individual states from interpersonal dynamics; interpersonal dynamics from group structure; group structure from organizational context; and organizational context from culture, environment and exceptional conditions such as crises. The objective is not to create a perfect digital replica of human reality. Such an ambition would exceed current scientific and technical capabilities. The objective is to make enough of the relevant context explicit that downstream AI systems are less likely to interpret an ambiguous signal as though it had one universal meaning.

This becomes increasingly important as AI systems gain agency. A conventional chatbot can tolerate substantial contextual ignorance because a human remains responsible for interpreting its output and deciding what happens next. An autonomous or semi-autonomous agent has less room for that ambiguity. If it prioritizes customers, reallocates work, escalates employees, communicates externally, approves exceptions or coordinates with other agents, incorrect contextual assumptions can become operational actions. The strategic question therefore changes from whether a model can produce an intelligent answer to whether the surrounding architecture gives the model sufficient context to know when several intelligent answers remain possible.

2. Human behavior cannot be reduced to a single state variable because the same observation can support several incompatible explanations

One of the most persistent simplifications in applied AI is treating human understanding as a classification problem. Sentiment systems historically reduced language to positive, neutral and negative. Emotion-recognition systems expanded the number of categories. Behavioral analytics added engagement measures, interaction frequency, facial signals, vocal properties and other observable features. These methods can be useful for constrained applications, but they become problematic when observable signals are treated as direct measurements of hidden internal states. Human behavior is multidimensional, and the relationship between what can be observed and what a person is experiencing is rarely one-to-one.

A manager who becomes unusually quiet during a meeting may be disengaged, concentrating, disagreeing, deferring to a more senior participant, strategically withholding judgment, fatigued, uncertain, processing unfamiliar information or simply behaving according to their normal conversational style. Increased speech rate can accompany anxiety, enthusiasm, urgency, anger, excitement or a speaker's ordinary baseline. Reduced eye contact can reflect discomfort, cultural convention, concentration, neurodivergence, status relations or environmental distraction. Even when several signals occur simultaneously, interpretation remains conditional on baseline behavior and context. A robust human-context architecture must therefore represent configurations and competing explanations rather than deterministic labels.

Contemporary affective science provides substantial reason for caution. A 2024 Nature Communications study drawing on archival and experimental datasets found that isolated facial information was insufficient for robust emotion inference and that situational information could sometimes explain emotion judgments as well as, or better than, combined facial and contextual information. The researchers also observed substantial variation by emotion category and individual. Broader reviews of emotion perception have similarly challenged interpretations derived primarily from static, posed facial expressions, emphasizing the role of dynamic behavior, situation, prior knowledge and cultural context. These findings do not imply that facial or behavioral signals contain no information. They imply that the information is conditional and probabilistic rather than a reliable direct readout of internal psychological state. (Nature Communications, 2024; Nature Reviews Psychology)

This distinction has direct architectural consequences. A human-context system should be able to store an observation such as “speech rate increased relative to baseline” without automatically converting it into “anxiety detected.” It may record that the pattern is compatible with heightened activation while simultaneously retaining alternative explanations such as enthusiasm or urgency. Additional contextual information—conversation content, recent workload, environmental conditions, self-report, previous behavior and cultural norms—can then strengthen or weaken those hypotheses. The system becomes more useful precisely because it is capable of saying that the available evidence does not discriminate among several explanations.

The design principle is straightforward but consequential: signals should be treated as evidence, not verdicts. The difference separates contextual intelligence from automated mind reading. It also creates a more scientifically defensible foundation for enterprise applications because the system can expose what it observed, what it inferred, what alternatives remain and how confident the inference should be. In human systems, a well-calibrated “unknown” can be more valuable than a confident but unsupported psychological label.

3. Context is not metadata surrounding human behavior; it is part of what gives behavior meaning

The limitations of single-signal interpretation become more pronounced when behavior moves across relationships, organizations and cultures. A direct disagreement between two colleagues can represent healthy intellectual challenge in one organization and a serious breach of status expectations in another. Silence during a negotiation can indicate uncertainty, resistance, strategic patience or respectful attention. A manager's brief response can be interpreted as efficient communication by one employee and as disapproval by another. High emotional expressiveness may be considered authentic and engaged in one cultural environment and inappropriate in another. The behavioral event has not changed; the contextual system that determines its meaning has.

This is why human context cannot be treated as metadata appended after an inference has already been made. Context participates in the inference itself. A sufficiently rich architecture therefore needs to represent at least several interacting layers: the individual and their relevant baseline; the observable signal; the relationship between participants; the group or organizational structure; the immediate environment; relevant cultural norms; the temporal situation; and exceptional circumstances capable of changing normal behavioral expectations. Not every decision requires all of these layers, and collecting unnecessary personal information would itself create serious governance problems. The point is architectural rather than maximalist: systems should be able to retrieve the contextual dimensions that materially change interpretation rather than assume that meaning is contained entirely within the immediate signal.

Cross-cultural emotion research reinforces the importance of this design. Contemporary psychology increasingly treats emotion as emerging through interactions among biological processes, individual experience, social learning and cultural concepts rather than as a set of perfectly universal labels that map identically onto observable expressions in every population. A 2024 Scientific Reports study found cultural differences in how facial expressions were described, while broader work reviewed in Nature Reviews Psychology has emphasized cultural variation in emotion concepts, interpretation and display. These findings do not imply that every emotional phenomenon is culturally unique. They demonstrate why a global AI system should not assume that one expressive vocabulary can be applied without qualification across populations. (Scientific Reports, 2024; Nature Reviews Psychology)

The enterprise implications are substantial. Multinational organizations increasingly deploy the same collaboration, HR, customer-service and AI platforms across countries with very different norms around hierarchy, directness, emotional display, conflict, privacy and formality. A leadership assistant operating in New York, Hanoi, Riyadh, Singapore and São Paulo should not assume that identical conversational behavior carries identical social meaning. Equally, it should not convert population-level cultural tendencies into deterministic claims about an individual. Culture can appropriately modify an interpretation prior; it should not determine the person. That distinction is critical because a system intended to become culturally aware can otherwise become an automated stereotyping mechanism.

A mature contextual architecture therefore needs both population-level context and individual-level correction. Cultural and organizational knowledge can inform the range of plausible interpretations, but observed individual history, explicit preference and situational evidence should be capable of overriding those priors. This makes context adaptive rather than categorical and helps prevent the architecture from confusing statistical tendencies with personal truth.

4. The economic case begins with a large mismatch between the sophistication of enterprise transactional systems and the weakness of enterprise representations of human conditions

Organizations have spent decades creating increasingly precise digital representations of financial and operational reality. An ERP system can represent inventory quantities, purchase orders, production states and accounting relationships with high precision. CRM platforms can represent pipeline stages, customer accounts and commercial interactions. HR information systems can represent reporting relationships, compensation, tenure and employment status. Observability platforms can record system events at millisecond resolution. Modern enterprises consequently possess sophisticated systems of record for transactions, resources, customers, software and increasingly knowledge.

Their representations of human context are much less mature. Concepts such as deteriorating trust, unresolved conflict, social withdrawal, role ambiguity, decision paralysis, team overload, status tension, psychological safety, collaboration quality and recovery after disruption are often captured through periodic surveys, manager judgment or not represented structurally at all. This is not necessarily a technology failure. Many of these constructs are intrinsically difficult to measure, ethically sensitive and dependent on interpretation. Nevertheless, the asymmetry becomes economically important as AI systems are asked to make decisions inside the human environment while possessing far more structured information about transactions than about the people and relationships that determine whether those transactions succeed.

The scale of the underlying workforce challenge is substantial. WHO estimates that approximately 15 percent of working-age adults had a mental disorder in 2019 and that depression and anxiety result in roughly 12 billion working days lost globally each year, at an estimated productivity cost of about US$1 trillion annually. WHO also identifies excessive workloads, low job control, discrimination, inequality, job insecurity and other organizational conditions as material workplace mental-health risks. Gallup's 2026 global workplace reporting, based on 2025 measurement, found that only 20 percent of employees worldwide were engaged at work and approximately one-third were thriving in their lives; measures of negative daily experience remained elevated relative to prepandemic levels. These statistics do not demonstrate that AI-based contextual analysis can solve workplace mental health, nor would they justify automated psychological monitoring. They establish something more basic: the human operating environment is economically consequential, and organizations continue to struggle to understand and manage it. (World Health Organization; Gallup, 2026)

This distinction matters because the strongest business case for Human Context Infrastructure is not “emotion detection.” It is reducing contextual error in decisions involving people. A workforce planning system that understands workload, role structure and organizational conditions may allocate resources more intelligently than one optimizing task volume alone. A customer-service system that represents conversational trajectory and prior interactions may determine that escalation is appropriate without claiming to know a customer's internal emotional state. A leadership system can identify repeated decision bottlenecks from observable workflow evidence without diagnosing the psychology of the executives involved. A coaching application can structure self-reported experiences without inferring hidden clinical conditions. In each case, the system becomes more contextually capable without crossing into unsupported claims of psychological certainty.

The economic opportunity should therefore be measured through decision outcomes rather than the accuracy of human-state labels. Relevant measures include reduced escalation failure, improved customer resolution, lower workflow friction, faster conflict resolution, improved intervention timing, reduced employee overload, higher adoption of AI tools and fewer consequential decisions based on missing contextual information. If contextual representation does not improve such outcomes relative to simpler baselines, the additional complexity is not justified.

5. The more important transition is from sentiment analysis toward multimodal, multi-level representation

The history of computational human understanding has been characterized by useful reduction. Sentiment analysis made large volumes of language tractable by compressing expression into a small number of categories. Emotion classifiers expanded the vocabulary. Speech and computer-vision systems added prosody, facial configuration and physical behavior. Organizational analytics incorporated communication frequency, collaboration networks and workflow patterns. Each development increased machine visibility into human systems, but each also created the temptation to mistake a measurable proxy for the underlying phenomenon.

Human Context Infrastructure represents a different architectural direction. Instead of searching for one increasingly sophisticated classifier capable of declaring a person's state, it separates different classes of information. Observable behavior is one layer. Possible internal state is another. Action is another. Relationship structure, group dynamics, environment, culture and temporal context occupy additional layers. The distinction is more than semantic. An observation that a person's speech rate has increased is fundamentally different from an inference that they may be anxious; an inference of anxiety is different from a prediction that they may disengage; a prediction is different from a recommendation to intervene; and a recommendation is different again from organizational authorization to act.

Keeping these layers typed and separable is one of the strongest safeguards against epistemic overreach. It allows a downstream reasoning system to combine observations and context while preserving the provenance of each conclusion. It also makes correction possible. If a behavioral observation proves wrong, only conclusions dependent on that observation need to change. If a cultural assumption is inappropriate for a particular individual, that assumption can be replaced without invalidating unrelated evidence. If a prediction performs poorly in a particular environment, its authority can be narrowed without deleting the underlying observations. This architecture is more inspectable than an end-to-end system that directly converts multimodal input into a consequential human classification.

The commercial category is therefore better described as multimodal contextual intelligence than emotion AI. The objective is not to build machines that claim to “read” people. It is to give machines structured access to the relevant dimensions of a situation so they can reason with greater humility and precision about what remains uncertain. That may involve language, conversational timing, self-report, workflow events, environmental data, relationship history and, in carefully governed contexts, selected multimodal signals. The architecture should use the smallest amount of human data necessary for the decision rather than maximizing surveillance simply because more signals are technically collectible.

This last point will determine whether the category becomes broadly acceptable. The same contextual capability that can help a user recognize their own overload can be repurposed to rank employees by inferred emotional stability. The same conversational analysis that can improve customer support can be used for covert psychological profiling. The same multimodal signals that may improve accessibility can be used to infer sensitive traits or make employment decisions from scientifically weak proxies. The value of Human Context Infrastructure therefore cannot be separated from the governance architecture controlling collection, inference, retention and use.

6. Group intelligence may become more strategically important than individual intelligence as AI agents enter teams

Most enterprise AI applications are still designed around an individual user even when the relevant unit of economic performance is a team. Yet teams possess emergent characteristics that cannot be reconstructed by adding together individual states. Information can become concentrated around one person, informal coalitions can form, decision rights can remain ambiguous, disagreement can disappear from formal meetings while continuing privately, a team can become dependent on one technical expert, and an organization can appear operationally functional while its members progressively disengage. These are properties of interaction and structure rather than attributes of any single participant.

The introduction of AI agents makes this distinction more important. Future teams are likely to contain humans with different expertise and authority alongside agents with different models, tools, memory, permissions and objectives. The performance of such teams will depend not simply on whether each component is individually capable but on whether information moves appropriately between them, whether decision rights are understood, whether disagreement is surfaced, whether humans can challenge agent outputs, whether agents know when to escalate, and whether shared dependencies create hidden common-mode errors. A highly capable collection of individuals and agents can still form a poorly functioning decision system.

Microsoft's 2025 workplace research anticipates a substantial increase in human supervision of agents, while Deloitte's 2026 findings indicate that organizations remain early in designing human–AI interaction. These trends point toward an emerging organizational-design discipline in which interaction architecture becomes as important as model architecture. (Microsoft, 2025; Deloitte, 2026) Enterprises will need to understand not only whether an agent performs its assigned task but how its presence changes communication, authority, workload, trust, skill development and decision quality across the team.

Human Context Infrastructure could provide part of the representational foundation for that discipline. At the group level, relevant dimensions could include information distribution, decision concentration, conflict expression, coordination load, dependency structure, escalation patterns and recovery after failure. Importantly, these should be represented through observable organizational evidence wherever possible rather than speculative psychological attribution. A system can establish that decisions are repeatedly bottlenecked around one approval point without claiming that the approver has a particular personality. It can identify that information is not crossing functional boundaries without diagnosing why individual employees are withholding it. This distinction keeps group intelligence focused on systems that organizations can actually change.

The economic opportunity may be significant because AI adoption itself is likely to alter these patterns. Automating tasks changes who holds information. Agents can reduce coordination cost but can also create new bottlenecks around verification. They can democratize expertise while increasing dependence on shared models. They can give junior employees greater capability while simultaneously reducing the apprenticeship experiences through which expertise historically developed. The relevant management question is therefore not simply whether AI increases individual productivity. It is whether the combined human–AI system becomes a better organization.

7. Human-context systems become most valuable—and most difficult—when normal operating assumptions break

Contextual systems are easiest to design when environments are stable. Behavioral baselines can be estimated, organizational norms remain relatively consistent, and historical relationships provide useful priors. The challenge becomes substantially harder during discontinuity. Mergers, restructurings, leadership changes, layoffs, cyber incidents, economic shocks, pandemics, wars, natural disasters, rapid growth and institutional crises can change the meaning of behavior so quickly that a system calibrated under normal conditions becomes misleading.

The COVID-19 pandemic demonstrated this problem at extraordinary scale. Behaviors that would have been anomalous under 2019 operating conditions—mass remote work, social withdrawal, disrupted mobility, unusually high uncertainty and rapidly changing institutional rules—became understandable responses to a radically different environment. Organizations experience smaller versions of the same phenomenon continuously. Communication may decline during restructuring because employees are uncertain about information boundaries rather than because collaboration has deteriorated. Customer contact may surge because of an external disruption rather than declining product quality. Employee behavior after a cyber incident may reflect temporary procedural controls rather than a durable change in organizational culture.

A mature human-context architecture therefore needs regime awareness. The system should represent the environmental conditions under which an observation occurred and recognize when those conditions differ materially from those supporting previous interpretations. Historical data should remain useful, but historical similarity should not be mistaken for proof that the current causal environment is unchanged. In practical terms, contextual models should lose confidence when the surrounding environment changes faster than they can be revalidated.

This principle has broad relevance beyond behavioral interpretation. NIST's AI Risk Management Framework treats AI as a socio-technical phenomenon and emphasizes ongoing governance, mapping, measurement and management across the system lifecycle rather than one-time technical validation. Human-context systems intensify the need for that approach because the subject being modeled—human and organizational behavior—is itself adaptive. People change in response to technology, incentives, observation and institutional conditions. A system deployed into an organization can therefore alter the environment it is attempting to measure. (NIST)

The implication is that contextual knowledge requires expiration and revalidation. A behavioral interpretation valid for one team, culture, leadership structure or crisis regime should not silently become a universal rule. Human Context Infrastructure should preserve where an interpretation came from, under what conditions it was supported and what changes would require reconsideration. Without that discipline, organizational memory becomes a mechanism for institutionalizing outdated assumptions.

8. The largest strategic opportunity may emerge as agents require context before they can safely exercise greater autonomy

The move from AI assistance to AI agency changes the economic importance of context. A human using an assistant can recognize that a recommendation is socially inappropriate, notice that a colleague is under unusual pressure, understand that a customer has a history not captured in the immediate ticket, or know that a technically correct action should not be taken because of an unwritten organizational constraint. Agents increasingly need machine-readable access to at least some of that context if they are expected to operate without continuous human interpretation.

The issue can be illustrated through ordinary enterprise ambiguity. A delayed response from a colleague may indicate low priority, overload, disagreement, absence, illness or deliberate avoidance. A customer's declining engagement can reflect dissatisfaction, seasonality, changed financial circumstances or successful completion of their original need. A reduction in internal communication can represent improved efficiency or deteriorating coordination. An agent optimized to act on the raw metric must choose among explanations that the metric itself cannot resolve. More model intelligence does not eliminate this ambiguity because the missing variable is contextual evidence.

Human Context Infrastructure would not solve this problem by telling the agent which interpretation is true. Its more important function would be to expose the hypothesis space. The system could represent several plausible explanations, identify the contextual evidence supporting each, expose missing information and instruct the agent to seek clarification when the uncertainty materially affects the action. In some cases, the most intelligent behavior would therefore be not greater inference but greater restraint.

This has implications for the design of agentic enterprises. Permission systems currently focus largely on what tools and data an agent can access and what actions it can execute. Future governance may also need to consider epistemic permission: whether the evidence available to the agent is strong enough to justify a particular interpretation or action. An agent may have technical permission to send a message, reassign work or escalate a case while lacking sufficient contextual evidence to know that doing so is appropriate.

As autonomous systems proliferate, this distinction between capability and contextual authority could become strategically important. The most valuable agent may not be the one that acts most frequently. It may be the one that can distinguish when the environment is sufficiently understood for autonomous action from when ambiguity requires another observation, human confirmation or a reversible experiment.

9. The principal risk is not simply incorrect classification; it is persuasive psychological certainty unsupported by evidence

Generative AI creates a distinctive risk in human-context applications because it can transform weak evidence into highly coherent explanations. A model can take a small number of behavioral observations and produce a psychologically plausible narrative about motivation, emotion, conflict or intent. The narrative may be fluent, nuanced and internally consistent while remaining only one of several possible explanations. Human readers are particularly vulnerable to this failure because psychologically coherent stories are intuitively satisfying.

The central governance problem is therefore epistemic overreach. A system should not transform “the speaker paused more frequently than usual” into “the speaker lacks confidence,” or “team communication volume declined” into “the team is disengaged,” unless additional evidence actually supports those conclusions. Even when probabilistic associations exist at a population level, they may be inappropriate for the individual, environment or moment under consideration. Correlation should not be silently converted into mechanism, and mechanism should not be silently converted into certainty about a particular person.

This requirement becomes even stronger when the ontology contains concepts adjacent to clinical psychology or psychiatry. Terms associated with depression, anxiety, trauma, mania or other mental-health conditions may be legitimate components of a broad human knowledge architecture, but their presence does not license an enterprise AI system to diagnose those conditions from workplace behavior. Clinical diagnosis requires validated methods, appropriate professional expertise and contextual information far beyond the signals available to most enterprise systems. Using clinical-sounding labels without those safeguards would create both scientific and ethical risk.

A robust architecture should therefore make uncertainty linguistically and computationally visible. “Observed rapid speech relative to baseline” is a stronger statement than “the person is anxious.” “The pattern is compatible with heightened activation” is more defensible than “stress detected.” “The available evidence supports several interpretations” may be more valuable than selecting the most probable interpretation when the difference would influence a consequential decision. The system should also expose what evidence would discriminate among alternatives.

This approach is not a limitation on intelligence. It is a higher standard of intelligence. In complex human environments, the ability to preserve ambiguity until evidence resolves it is often more useful than producing immediate certainty.

10. Governance will determine whether Human Context Infrastructure becomes augmentative technology or a new surveillance layer

The closer AI moves toward representing human psychological and behavioral context, the more consequential governance becomes. Algorithmic management already demonstrates the tension. The International Labour Organization describes algorithmic management as the use of tracked data and algorithmic systems to organize, assign, monitor, supervise and evaluate work, with applications across sectors including transport, logistics, customer service, healthcare, banking and digital labor platforms. Adding richer contextual inference to these systems could improve coordination and support. It could also materially expand the capacity of employers to monitor and classify workers. (International Labour Organization)

The distinction between augmentation and surveillance therefore needs to be architectural rather than aspirational. High-value applications are likely to include systems that help individuals understand their own self-reported patterns, enable teams voluntarily to identify workflow problems, improve accessibility, adapt interfaces to user needs, surface aggregate organizational conditions and prompt human conversations where intervention may be appropriate. These applications use contextual information to increase agency or improve the environment.

High-risk applications include covert emotional surveillance, hidden psychological scoring, employment decisions based on inferred internal states, continuous behavioral monitoring without a proportionate purpose, automated mental-health classification from weak proxies and systems in which individuals cannot challenge or correct contextual interpretations about themselves. Such applications can be harmful even when technically sophisticated because the inference itself may be scientifically uncertain while the institutional consequence is very real.

NIST's trustworthiness framework is directly relevant. Validity and reliability, safety, resilience, accountability, transparency, explainability, privacy and fairness are not optional additions to human-context systems; they determine whether the systems should be deployed at all. (NIST) The greater the consequence of an inference, the stronger the evidence, transparency and human review required. A contextual cue used to adjust the tone of an optional coaching interface requires a very different evidentiary standard from an inferred state used in promotion, insurance, lending, healthcare or disciplinary decisions.

The governance principle can therefore be expressed simply: the weaker the evidence about the person and the greater the consequence to the person, the less authority the system should possess. This principle should shape data collection, inference, retention, access and action. It also implies that some technically feasible uses of contextual AI should remain prohibited because no realistic improvement in model performance can make the institutional power imbalance acceptable.

11. The strongest business model is contextual augmentation, not behavioral omniscience

Commercially, the category becomes more credible when it abandons the promise that AI can “understand people” in an unrestricted sense. The stronger proposition is that enterprise AI can become less context-blind. This is both more scientifically defensible and more directly connected to measurable business outcomes.

In customer service, contextual infrastructure could help a system distinguish a routine transactional request from a repeatedly unresolved case, adapt communication to conversational history and recognize when uncertainty warrants human escalation. In leadership and organizational applications, it could surface repeated decision bottlenecks, information silos or unresolved workflow conflict based on observable organizational evidence rather than inferred personality. In coaching, it could structure self-reported experiences and identify patterns over time. In multinational organizations, it could remind systems that directness, hierarchy and disagreement are culturally contingent. In crisis operations, it could alter assumptions about normal behavior when the environment has materially changed. In human–AI teams, it could help agents determine when to ask, when to escalate and when several interpretations remain plausible.

The common economic mechanism across these applications is reduction of contextual error. Model-level intelligence may increasingly become commoditized, but the cost of applying an intelligent model to the wrong interpretation of the situation can remain high. If two organizations use similarly capable foundation models but one possesses better contextual knowledge, stronger organizational memory, clearer decision rights and better-designed human–AI workflows, the second organization can create superior outcomes without owning a fundamentally superior base model.

This suggests a broader shift in enterprise software. Systems of record digitized transactions. ERP digitized resources. CRM digitized customer relationships. Business intelligence digitized performance measurement. Knowledge-management systems digitized institutional information. Foundation models created a new interface to unstructured knowledge. The next layer may increasingly consist of systems of context: infrastructure that represents not only what happened but under what organizational, relational, environmental and temporal conditions it happened, how certain that interpretation is, and when the context is no longer valid.

Such systems would not replace existing enterprise applications. They would provide contextual services to them. A customer record could carry relevant interaction history without asserting a psychological profile. A workflow could contain information about current organizational load. An agent could query whether the operating environment is normal or exceptional. A recommendation could include competing contextual interpretations. Organizational memory could distinguish durable facts from temporary states. The resulting architecture would make context machine-readable while retaining boundaries around what machines are entitled to infer.

12. The category will require substantial validation before it can become trusted enterprise infrastructure

The strategic case for Human Context Infrastructure is stronger than the empirical case for any single comprehensive ontology today. That distinction should remain explicit. A taxonomy can be coherent, useful and architecturally sophisticated without every category, relationship or inferred transition having been scientifically validated. Enterprise deployment therefore requires a staged research agenda rather than treating conceptual completeness as evidence of predictive validity.

Construct validity is the first requirement. Categories intended to represent psychological, behavioral or organizational phenomena need comparison with established research and validated instruments where those exist. Cross-cultural validity is equally important because constructs developed within one linguistic or cultural environment may not transfer directly to another. Multimodal mappings require independent validation because observable signals such as facial movement, vocal change, gaze or posture are ambiguous and can be affected by individual baseline, disability, neurodivergence, environment and culture. Where the system represents probabilistic relationships between state and behavior, those relationships should remain hypotheses until demonstrated in relevant populations.

Temporal validity represents another underappreciated requirement. Different contextual variables decay at different rates. An immediate state such as fatigue may change within hours. A project role can change within weeks. A team relationship can evolve over months. Organizational culture may change over years but can also shift rapidly during crisis or leadership transition. A mature infrastructure therefore needs different freshness rules for different classes of context rather than treating all human knowledge as equally persistent.

Privacy, consent and purpose limitation must be structural. The architecture should minimize collection, distinguish information provided explicitly by an individual from information inferred about them, allow correction where appropriate, restrict sensitive inference, and prevent contextual data gathered for supportive purposes from silently migrating into consequential evaluation. Clinical terminology requires particularly strict separation from ordinary workplace inference.

Finally, contextual systems need evidence that their additional complexity improves outcomes. Rich ontologies can become intellectually impressive but operationally unnecessary. Every major application should therefore be tested against simpler baselines. If ordinary workflow data and straightforward rules perform equally well, collecting sensitive multimodal human data is not justified. The burden of proof should increase with both privacy intrusion and decision consequence.

13. Human Context Infrastructure could become a strategic layer of the agentic enterprise

The broader opportunity can now be stated more precisely. The AI industry has built increasingly powerful layers for computation, language, retrieval, memory, reasoning and action. It has invested comparatively less in explicit, governed representations of the human, organizational and environmental context within which those capabilities operate. That imbalance was manageable when AI principally answered questions. It becomes more consequential as AI participates in work, coordinates tasks, communicates on behalf of organizations and gains authority to act.

Human Context Infrastructure would address that gap by providing machine-readable representations of relevant individual, relational, organizational, cultural, environmental and temporal conditions while refusing the assumption that representation equals psychological truth. Its purpose would not be to tell machines exactly what people are thinking. It would be to prevent machines from behaving as though people and organizations exist without context.

The distinction is strategically important because the economics of AI are shifting. As baseline model capabilities become more widely accessible, durable advantage is likely to migrate toward proprietary data, institutional knowledge, workflow integration, trusted memory, contextual understanding, governance and organizational capability. McKinsey's adoption data already suggest that widespread AI use and enterprise-level financial impact remain different states. Deloitte's findings similarly indicate that organizations recognize the need for substantial cultural and workforce adaptation while remaining early in the redesign of human–AI interactions. (McKinsey & Company, 2025; Deloitte, 2026) The bottleneck is increasingly not whether the machine can perform a cognitive task but whether the organization can place that capability inside a functioning human system.

This creates a potentially significant infrastructure category. Organizations may ultimately need contextual services in much the same way they now require identity, security, data governance and observability. Those services could maintain contextual provenance, separate observations from inferences, preserve competing explanations, track cultural and environmental applicability, manage freshness, define which systems can access sensitive context and govern when an agent must defer to human judgment.

The architecture should remain deliberately modular. Content and ontology describe the possible contextual space. Evidence systems populate observations. Inference systems assess competing interpretations. Prediction systems estimate possible outcomes. Decision systems evaluate options. Governance systems determine authority. This separation allows organizations to improve one layer without silently changing the others and makes it possible to challenge an inference without erasing the evidence from which it was produced.

Conclusion: the next enterprise AI advantage may come from understanding the system around the model

The economics of artificial intelligence are changing rapidly. Organizational adoption has become mainstream, agent experimentation is expanding, investment remains substantial and automation is expected to reshape a meaningful share of working activity. At the same time, organizations face persistent workforce pressure, low global engagement, skills disruption and an incomplete operating model for human–AI collaboration. McKinsey reports broad AI adoption but substantially narrower enterprise-level EBIT impact; Stanford documents the rapid increase in organizational AI use and investment; the World Economic Forum expects major job and skill transformation through 2030; Microsoft anticipates a growing workforce role in training and managing agents; and Deloitte finds that organizations remain early in designing the human–AI interactions needed to make those systems work. (McKinsey & Company, 2025; Stanford HAI, 2025; World Economic Forum, 2025; Microsoft, 2025; Deloitte, 2026)

The human environment into which AI is being introduced is itself under significant pressure. WHO estimates that depression and anxiety contribute to approximately 12 billion lost working days and around US$1 trillion in lost productivity annually, while Gallup reports global employee engagement at approximately 20 percent. Those statistics do not create a mandate for automated psychological surveillance. They demonstrate the opposite: the human system is economically important enough that organizations cannot afford to represent it carelessly. (World Health Organization; Gallup, 2026)

The strategic gap is therefore becoming clearer. Machines are increasingly capable of reasoning about work at precisely the moment organizations need better ways of representing the human context of work. A credible Human Context Infrastructure would approach that challenge by modeling multiple interacting levels—observable signals, individual state, relationships, groups, organizational conditions, culture, environment, crisis and time—while maintaining a strict boundary between representation and truth. Its classifications would remain hypotheses where evidence is incomplete. Cultural context would inform rather than determine individual interpretation. Sensitive inference would be constrained by purpose and consequence. Clinical constructs would remain outside ordinary enterprise classification unless appropriate validated methods and professional governance existed.

This is a materially different proposition from emotion recognition or automated mind reading. It does not require believing that machines can reliably infer every human state. It requires recognizing that machines already make decisions affecting humans and that those decisions become less reliable when context is absent, oversimplified or silently assumed.

The first generation of enterprise AI was primarily about machine capability. The emerging generation is about machine participation. Participation requires systems to understand not only tasks and information but the conditions surrounding them. As agents become more autonomous, context becomes more important because the human who previously supplied that context implicitly is increasingly removed from each intermediate decision.

The most successful systems are therefore unlikely to be those that claim the greatest psychological certainty. They may be those capable of representing several plausible interpretations, identifying the evidence supporting each, understanding cultural and environmental boundaries, recognizing when historical assumptions have become stale, and knowing when the available evidence does not justify inference at all.

That is the stronger case for Human Context Infrastructure.

It is not a machine-readable theory of human nature. It is not a universal behavioral prediction engine. It is not a license for organizations to infer hidden psychological states from every available signal.

It is an architectural response to a much more practical problem: intelligence without context can be highly capable and still be wrong about the situation in which it is acting.

As foundation-model capability becomes increasingly accessible, that distinction could become commercially significant. Competitive advantage may progressively migrate from access to the most fluent model toward the quality of the system surrounding the model: proprietary organizational knowledge, contextual data, trustworthy memory, human–AI workflow design, governance, provenance, calibrated uncertainty and high-quality representations of the environments in which decisions actually occur.

The next frontier of enterprise intelligence may therefore not be making AI generate more.

It may be building the contextual infrastructure that allows increasingly capable AI to understand where it is, who it is interacting with, what it does not know, and when it should not infer at all.

References and source notes

McKinsey & Company. The State of AI: How Organizations Are Rewiring to Capture Value, 2025.

McKinsey Global Institute. Generative AI and the Future of Work in America, and related research on automation and occupational transitions.

Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2025, Stanford University, 2025.

World Economic Forum. Future of Jobs Report 2025, Geneva, 2025.

Microsoft. 2025 Work Trend Index Annual Report, 2025.

Deloitte. 2026 Global Human Capital Trends, 2026.

World Health Organization. Mental Health at Work, WHO.

Gallup. State of the Global Workplace 2026, Gallup, 2026.

National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework and Generative Artificial Intelligence Profile, NIST.

International Labour Organization. Research and policy materials on algorithmic management and digitalized work.

Nature Communications. Research on contextual information and emotion inference, 2024.

Nature Reviews Psychology. Reviews of emotion perception, dynamic expression and cultural context.

Scientific Reports. Cross-cultural research on descriptions and interpretation of facial expressions, 2024.