Introduction
On February 28, 2026, joint US-Israeli airstrikes hit Iranian military infrastructure. Within weeks, Iran closed the Strait of Hormuz, choking off roughly one-fifth of global oil and liquified natural gas supply (IEA 2026a). Brent crude surged from $72 to $120 a barrel in three weeks (EIA 2026). Iranian drone strikes hit three Amazon Web Services data centers in the UAE and Bahrain, the first direct military targeting of commercial cloud infrastructure in history (CNBC 2026). Gulf states, which import upwards of 85 percent of their food (FAO 2026), faced severe supply disruption as shipping through the strait collapsed. The International Energy Agency described the crisis as worse than the oil shocks of 1973, 1979, and 2002 combined (IEA 2026b).
For many firms, the crisis exposed a structural shortcoming in their AI-driven risk and strategy frameworks: predictive platforms rated Gulf energy exposure as manageable. Their ESG scoring algorithms, calibrated to a post-Cold War integration consensus, assigned moderate risk to Middle East operations as LLM-generated country assessments drew on decades of settled assumptions about Strait of Hormuz transit security. Most of these models pointed to a region and a supply chain that, within seventy-two hours, no longer functioned as described. The increasing corporate reliance on AI tools diminished the very variation in strategic insight and complexity that allows smart firms to succeed during crisis.
In 2022, a similar pattern played out with Russia’s invasion of Ukraine. ESG platforms rated Russian-market exposure as moderate as predictive models assigned low probability to full-scale invasion. Afterwards, BP wrote down $25 billion in Rosneft assets. Shell took a $4 billion impairment on its Russian operations. The ‘partisan CSR’ that followed, where corporate political positioning supplanted traditional stakeholder-balancing logics, rendered ESG measurement frameworks inoperative (Bamiatzi et al. 2025). That was supposed to be a wake-up call for business to think more in the geopolitical present, the kind of shift that new AI tools claimed to offer.
Four years and billions of dollars in AI-driven risk infrastructure later, the Iran crisis exposed the same structural failure at a larger scale. In many ways, these tools performed as designed: they synthesized the best available information, identified patterns, and generated recommendations grounded in evidence. The question is whether that design is adequate for high-risk environments. AI-driven decision tools have become central to how firms assess risk, identify stakeholders, evaluate market opportunities, and navigate regulatory complexity in volatile environments. The geopolitical risk analytics platform sector was valued at $3.5 billion in 2025, with AI-driven predictive analytics projected to reach $15 billion by 2035 (SNS Insider 2026). Palantir Technologies, Verisk Maplecroft, EY’s Geostrategic Business Group, McKinsey’s geopolitical risk practice, Dataminr, and Marsh McLennan’s political risk advisory have all integrated AI-driven scenario tools into their core client offerings. These platforms constitute the analytical infrastructure through which multinational firms assess where to invest, when to exit, and how to comply with an expanding web of sanctions, due diligence mandates, and regulatory regimes. The structural limitations of these tools under conditions of crisis and instability remain poorly understood, and largely unexamined in the business practitioner literature (Lumineau & Keller 2025).
The academic landscape on AI and business decision-making has grown rapidly, but along two tracks that rarely intersect. One celebrates AI’s capacity to enhance managerial cognition and optimize resource allocation (Davenport & Ronanki 2018; Agrawal et al. 2022). The other interrogates what AI does to the nature of knowledge itself: how it reshapes what counts as evidence, whose expertise is valued, and what forms of understanding get systematically excluded (Zuboff 2019; Bommasani et al. 2021). Neither has adequately addressed the conditions under which AI’s analytical frameworks may fail, such as environments defined by instability, rapid institutional change, and normative fragmentation. We are living in a world of polycrisis, defined here per Miklian and Katsos (2024) as the condition in which overlapping crises interact and compound in ways that exceed the sum of their individual effects, constituting a permanent operating environment of uncertainty. AI tools were designed for a more stable world, but the real forward issue is rapidly becoming how much they can tell us about this one.
I engage this gap through stakeholder theory to address a core question for managers and practitioners: if the tools we increasingly rely on for business intelligence are structurally weakest in the moments when we need them most, what does responsible decision-making actually look like under crisis? The growing literature on human-AI collaboration in organizations has examined how algorithmic tools reshape managerial cognition, decision authority, and the division of analytical labour between humans and machines (Jarrahi 2018; Raisch & Krakowski 2021). But this work has largely focused on stable operating contexts. I argue that AI tools can undermine the three capacities stakeholder engagement requires under instability: the ability to identify who matters, the ability to track how salience shifts, and the ability to read signals from affected communities. This argument draws on case experiences of firms navigating crisis and instability, illustrative cases from algorithmic governance, and the broader literature on AI’s epistemic constraints (Bommasani et al. 2021; Bender et al. 2021).
The article proceeds in five sections. Section 2 lays out AI’s knowledge problem under instability, disaggregating it into three distinct epistemic modes with particular attention to the failure of AI to read relational risk. Section 3 examines how these limitations distort stakeholder identification and engagement, introducing the concept of algorithmic invisibility and drawing on both corporate and public-sector cases. Section 4 specifies what AI does well. Section 5 explores why AI’s structural limitations make local knowledge networks indispensable, and proposes a phased implementation pathway for firms. Section 6 concludes with separate implications for practice and research.
2. The Nature of AI’s Knowledge Problem
AI-driven business tools operate in three temporal modes, each with a distinctive relationship to the environment it purports to describe. Under stable conditions, all three function reasonably well. Large language models (LLMs) train on vast corpora of published text, including academic papers, news archives, policy reports, corporate filings, and web content, with data cutoffs typically running three to twelve months behind the present. ESG scoring platforms aggregate historical disclosure data, analyst assessments, and regulatory filings on similar timescales. Predictive risk platforms like Verisk Maplecroft compile country-level indicators from institutional sources that are themselves backward-looking. The result is a body of evidence calibrated to the recent past.
In stable environments, the recent past is a serviceable guide to the present. Swiss banking regulation six months ago tells you something useful about Swiss banking regulation today. But instability is defined by the rupture between past and present. The conventional wisdom about Myanmar’s business environment in January 2021 bore little relationship to the operating reality of February 2021, after the military coup upended the democratic transition that every risk model had priced in. ESG scores for Russian-exposed firms in early 2022 collapsed within weeks of the invasion. AI tools could process available information efficiently, but the value of this old information itself had diminished.
An ingrained geographic and linguistic bias compounds the temporal lag issue. The information ecosystems that AI draws upon are overwhelmingly English-language, Global North-centered, and institution-heavy. Local media in crisis-affected settings, informal knowledge networks, non-English policy discourse, and community-level intelligence are systematically underrepresented in training data. For example, when a political risk employee asks an LLM to assess operating conditions in eastern Congo or post-earthquake Turkey, the model draws on a thin, externally produced slice of a far richer and more contested information landscape. The output reads authoritatively, but it draws its authority from sources that were themselves operating at a distance. This is a structural limitation of LLMs: the training data reflects existing power asymmetries in knowledge production (Bommasani et al. 2021; Bender et al. 2021).
Predictive risk platforms extrapolate from historical patterns. If these economic indicators, this level of political instability, and this configuration of institutional stress have historically preceded disruption, the model assigns a probability. Scenario-generation tools in LLMs operate similarly, constructing plausible futures from pattern libraries built on documented pasts. However, crisis contexts are defined by their breaks of patterns, and political science literature is instructive. The securitization of virtually all domains of public policy, from supply chains to artificial intelligence to climate, has expanded the scope of sovereign prerogative while straining state capacity (Drezner 2024). Geopolitically induced political risk is also relational, arising from interactions between countries, alliance shifts, and regulatory bodies (Moura et al. 2026). When the normative world within which ‘risk’ and ‘stability’ carry meaning is fragmenting and contestation targets the validity of international norms and their application (Lesch et al. 2024), pattern-based prediction loses its ability to use the past to predict the present.
Many modern AI architectures are not static, and practitioners are investing heavily. Retrieval-augmented generation (RAG) systems can query real-time news feeds, proprietary databases, and live regulatory filings. Agentic AI platforms can trigger API calls and update their outputs continuously as they decompose complex analytical tasks, assign them to specialized sub-agents, and aggregate the results. A RAG-augmented geopolitical risk platform in early 2026 could pull real-time shipping data and oil futures prices within hours of the Strait of Hormuz closure, a substantial improvement over the months-long training lag of a static LLM. But this improvement (and one could argue any such future improvements) is bounded by the information ecosystem. RAG systems still retrieve and assess what is published, indexed, and digitally accessible. Agentic systems can monitor sentiment shifts across social media platforms, but they cannot distinguish a genuine signal of deterioration from a coordinated disinformation campaign or model for distributional shifts that have not yet occurred. The current trajectory of AI development is toward faster retrieval, broader data ingestion, and more sophisticated reasoning chains. It does not address the epistemic scope of what can be synthesized, and the distinction matters because it defines the ceiling of what technical improvement alone can deliver.
This cornerstone of how LLMs deliver their analysis can be mapped onto today’s risk needs for firms. The concept of assessing risk in “the present” requires disaggregation into three distinct capacities that instability demands from decision-makers. The first is perceptual immediacy, or the ability to register that conditions have shifted faster than data pipelines can capture. AI systems are poor at registering nuance in real-time that sharp humans can, like that a meeting’s tone has changed, that a government contact is choosing words more carefully than usual, or that a community leader’s absence is an essential data point. Immediacy signals constitute the classic yet intangible “gut feeling” that a manager might have about a situation in the process of change, whereby nimble yet information-poor decision-making has outsize import.
The second is tacit operational knowledge, meaning the accumulated understanding of how institutions actually function in a specific context. This might mean knowing which officials enforce which regulations, or which informal power brokers actually are able to ensure approval of a key permit, or which local partner promises are credible and which are performative. This knowledge resists formalization because its value lies in its personal contextual specificity. A firm’s local operations manager in a conflict or fragile setting carries in her head a map of institutional reality that can be incredibly difficult to map onto the formal governance structure that exists on AI-generated country risk reports. In short, this knowledge allows a firm to work beyond simplistic AI risk assessments to find opportunities that are otherwise intangible.
The third capacity is relational trust, or the network of relationships through which sensitive information flows and through which a firm’s local legitimacy is negotiated. Trust is built through presence, reciprocity, and reliability over time. Relational trust matters for business because it is how decision-relevant information travels. In stable environments, firms can rely on published data, regulatory filings, and market signals because these channels carry most of the information that matters. Under instability, consequential intelligence that determines security and stability moves through relational networks that are challenging for AI systems to access. For example, a local civil society partner who trusts the firm’s country manager might share early warnings about community resentment before it escalates, or a cultivated government contact might signal that a regulatory environment is about to shift, offering this information because the relationship warrants the risk. These forms of intelligence carry enormous value, and depend on human relationships that require years of investment. These signals can determine whether a business operation contributes to stability or accelerates harm. The failure to read relational risk is a deep limitation of AI-driven business intelligence lies, because relational risk compounds fastest under instability. When a firm loses relational trust with a local community, information and early-warning channels close as the firm’s social license to operate erodes, and the cost of re-establishing legitimacy grows.
Case experiences from firms operating in high-risk settings underscore this need. Freeman’s (1984) foundational insight that firms bear obligations to groups affected by their decisions has been extended to high-risk business environments, where firm-stakeholder relational strategies influence peace and conflict outcomes (Ganson et al. 2022; Miklian & Schouten 2019; Joseph et al. 2025). Business leaders with direct operational presence in volatile contexts perceive fundamentally different dynamics than those relying on external information sources. They identify stakeholders invisible to formal mapping exercises. They recognize patterns of institutional behavior, selective enforcement, informal power-brokering, strategic silence from affected communities, that no algorithmic system captures. And they express deep skepticism toward standardized risk tools (Hoelscher & Miklian 2025).
I argue that the result is an AI knowledge base that produces its most confident outputs precisely when the environment has diverged most sharply from the conditions those outputs describe. To provide an empirical baseline for interrogating this claim, I assessed data from a 40-country matched-pair study of AI-generated business environment assessments (see Appendix). The findings showed how AI delivered deeply self-confident responses for both stable and unstable countries even though the model knows its knowledge base is weaker for crisis environments. Assessment quality is measurably weaker on composite quality metrics, and the use non-English training data is almost nil. Despite the high confidence signals, in crisis settings the AI was effectively flying blind. A corporate user treating AI confidence ratings as reliability signals would be systematically misled.
Knowing these limitations, how can managers and practitioners help to overcome them? Organizational learning literature helps explain how we can incorporate these findings into practice, by distinguishing between exploitation, the refinement of existing competencies within known distributions, and exploration, the search for new possibilities that challenge existing frames (March 1991). AI-driven business intelligence is an exploitation technology and under stability, exploitation works. Under instability, organizations need exploration, the capacity to recognize that conditions have shifted and that existing frameworks no longer apply. Organizations frequently absorb new information into existing practices without undertaking the fundamental changes needed for adaptation (Levitt & March 1988); AI accelerates this tendency by providing a continuously updated veneer of analysis that masks the degree to which the underlying framework has become obsolete. It may be a new version of the competency trap, where an organization’s proficiency with existing tools discourages the search for better ones (Leonard-Barton 1992). Thus, the remainder of this article discusses how to merge exploitation with exploration in joint AI-human systems.
3. Stakeholders AI Cannot See
Stakeholder theory provides a framework through which business scholarship understands corporate obligations in complex environments. The core is well established: firms bear responsibilities to groups whose interests are affected by corporate decisions (Freeman 1984), and the salience of those stakeholders varies according to the power they wield, the legitimacy of their claims, and the urgency of their demands (Mitchell et al. 1997). In crisis and high-risk settings, these relational dynamics carry unusually high stakes. Firm-stakeholder relational strategies have direct, measurable effects on conflict risk (Ganson et al. 2022), and firms operating in governance gaps exercise quasi-political authority that demands legitimation (Scherer & Palazzo 2011). If firms in governance voids exercise quasi-political authority, then the delegation of tasks like stakeholder and risk mapping to algorithmic systems raises the question of who is represented in the algorithmic map, and who is excluded by the system’s apparent comprehensiveness?
These frameworks share an implicit assumption: that the information environment within which stakeholder identification, salience assessment, and engagement occur is reasonably accessible to the firm. Under instability, firms risk creating algorithmic invisibility: a new form of stakeholder exclusion distinct from stakeholder mapping in three ways. First, the system’s apparent comprehensiveness through sourcing, confidence scores, and professional formatting masks systematic gaps in ways that a human analyst’s acknowledged uncertainty does not. Second, because firms draw on the same underlying English-language data, (see Appendix), the same stakeholders may become excluded not just within but across firms, creating industry-wide blind spots. Third, when multiple platforms return similar stakeholder maps, decision-makers read convergence as confirmation, when what they are observing is shared dependence on the same limited information ecosystem, with AI identifying stakeholders in crisis settings but still weighting the value of foreign actors on a four-to-one basis.
AI privileges stakeholders who are digitally visible, formally organized, English-speaking, and connected to global information networks. Training data for language models is drawn overwhelmingly from published, indexed sources. Likewise, ESG scoring platforms rely on corporate disclosure, regulatory filings, and analyst reports, and risk platforms aggregate institutional data from multilateral organizations, credit agencies, and established media. Under instability, the most consequential stakeholders are not captured by such sources. They could be new armed groups, informal community leaders whose blessing or opposition shapes local legitimacy, or mediators and civil society organizations operating under threat without websites or English-language publications. The concept of institutional voids, originally developed to describe emerging markets where formal institutions are weak or absent (Khanna & Palepu 2010), may be taking new shape in the AI age as the stakeholders who matter most occupy informational voids that AI omits.
Stakeholder salience is also dynamic. Under instability, it shifts with a speed and non-linearity that AI’s temporal architecture cannot track. Power, legitimacy, and urgency, the key determinants of stakeholder salience (Mitchell et al. 1997), shift unpredictably under instability. A human analyst operating in a crisis context can ascertain the limits of knowledge gaps and can seek to fill it through relationships, local networks, and direct inquiry. Recent re-assessments of stakeholder identification have moved beyond instrumental framings to consider what is existentially at stake for affected communities (Wood et al. 2021). But a typical AI system might register no gap at all as it is complete by construction, because it can only map what exists in the data, creating a sense of systematic invisibility.
The corporate experience of the Russia-Ukraine war illustrates this. In the weeks preceding the February 2022 invasion, major energy firms including BP, Shell, and TotalEnergies maintained Russian operations guided in part by AI-driven risk assessments that treated sovereign expropriation as a tail risk (Koroteev and Tekic 2021). For example, BP’s partnership with Rosneft was rated as strategically sound by multiple ESG and geopolitical risk platforms through January 2022. Within ten days of the invasion, BP announced its exit, writing down $25 billion. Shell followed within the week. The Russian state was no longer a regulatory stakeholder with moderate power and high legitimacy. It had become an adversary whose actions made continued corporate presence politically untenable. Stakeholders who had been marginal like Ukrainian civil society and diaspora advocacy groups suddenly wielded decisive power over corporate reputation and market access. No ESG platform had modeled this salience inversion.
Perhaps the most consequential failure concerns how AI handles silence. In crisis contexts, silence from affected communities can be a purposeful strategy to navigate conditions of violence and uncertainty (Rauch & Ansari 2024). Communities withdraw from formal engagement when speaking carries risk, as in surveillance or retaliation in the absence of institutional protections. Stakeholder theory has traditionally privileged visible engagement, and AI systems inherit and amplify this bias by modeling based upon what has been disclosed. When a community goes quiet, a human with local relationships registers a signal that conditions have deteriorated or become politically fraught, and that the community is protecting itself through strategic opacity.
The empirical data (see Appendix) confirmed this limitation. AI saw silence as a potential signal in only 5 percent of cases, always in the form of generic disclaimers rather than context-specific readings. It never identified categories of actors it suspected existed but could not access, an important distinction because a data gap calls for more research, while strategic silence calls for restraint and relationship-building. But it still became a part of the solution. During the 2026 Iran crisis, logistics firms operating Gulf-based supply chains encountered this pattern. AI-driven supply chain risk platforms received alerts about shipping route disruptions and commodity price spikes while locally-embedded staff reported that port authority contacts had gone silent, that Emirati business partners were declining meetings, and that informal logistics networks were already rerouting people and goods through hidden channels. In short, the merging of artificial and human intelligence had arrived, and the firms who integrated the two most comprehensively were best prepared (Goswami 2026). The firms that maintained local intelligence capacity adapted much faster than those relying primarily on algorithmic intelligence.
Algorithmic invisibility is visible across domains where AI systems have been deployed to improve efficiency. France’s social security agency (CNAF) offers the clearest illustration. Since 2010, the CNAF has used a risk-scoring algorithm to flag welfare recipients for fraud investigations, scoring all 32 million benefit recipients on criteria that include low income, unemployment, residence in a disadvantaged neighborhood, and disability (Amnesty International 2024). The system constructs a confidence-bearing assessment from backward-looking data, pattern-matches against historical profiles, and assigns risk scores to individuals whose own perspective on their circumstances is systematically excluded from the assessment. The people scored as highest-risk are those whose lives are least legible to the data: informal economy workers, the recently displaced, residents of neighborhoods where institutional contact carries stigma. The pattern recurs across public-sector algorithmic governance: the Netherlands’ SyRI system disproportionately targeted low-income and immigrant communities before courts ruled it discriminatory (Amnesty International 2021), and India’s Samagra Vedika system auto-rejected ration card applications over database misalignments, cutting households off from essential food rations (Amnesty International 2024b).
When a corporate ESG platform assigns a risk score to a supplier in a conflict-affected region based on publicly available data, the structural operation is similar: backward-looking data as the epistemic foundation, pattern-matching against historical profiles, confidence scores generated from incomplete information, and the perspective of the assessed rendered invisible. ESG measurement has faced growing scrutiny for its divergence across rating agencies even under normal conditions (Berg et al. 2022; Bruno et al. 2026). Under instability, the problems multiply. The EU’s Corporate Sustainability Due Diligence Directive (CSDDD) will accelerate AI adoption in compliance functions, with implications for stakeholder visibility. Mandatory due diligence frameworks assume traceable causal chains from corporate action to human rights impact, an assumption that breaks down in crisis settings where effects are shaped by path dependencies that compliance monitoring cannot capture (Cechvala & Ganson 2024). AI-powered compliance tools that automate due diligence inherit this linear logic. They can track supply chain nodes, flag known risk indicators, and generate audit-ready documentation. But the emerging literature on human rights in AI supply chains has begun to document how algorithmic systems in procurement and compliance functions reproduce the visibility biases of their training data, flagging the risks that existing frameworks anticipate while systematically missing the harms that fall outside those frameworks (Johnston 2026).
For compliance teams and similar, the practical danger is that AI-powered due diligence tools may satisfy formal reporting requirements while failing to detect the stakeholder harms the directive was designed to prevent, precisely because those harms are most acute in the informational voids where algorithmic visibility is weakest. In short, for the growing class of corporate professionals integrating AI into their ESG and sustainability workflows, AI makes ESG look more rigorous, more comprehensive, and more valuable than it actually is. The empirical analysis uncertainty decomposition at 0% means the AI model never distinguishes between three fundamentally different states: situations where data is missing, situations where outcomes are genuinely unpredictable, and situations where actors are deliberately producing opacity. For a corporate decision-maker, these three states demand entirely different responses, yet the model treats them as identical.
Of course, algorithmic invisibility, identification bias, and silence misreading do not operate with equal force. A sudden geopolitical rupture like Russia-Ukraine or the Iran conflict creates a temporal shock where the entire information environment expires; here the salience lag and confidence calibration channels bite hardest. A slow-burn institutional degradation offers more time for information to propagate but may deepen identification bias, because invisible stakeholders accumulate over years. Localized conflict within an otherwise stable country (e.g. Colombia, Nigeria) may leave AI’s macro-level analytics largely functional while rendering its local stakeholder mapping catastrophically incomplete. The empirical data support this distinction: institutional fragility countries received higher AI confidence than expected, possibly because their instability is better documented in the institutional sources the model draws upon. The argument advanced here applies most forcefully to contexts where instability is acute, rapid, and normatively disruptive, though the underlying mechanisms operate at lower intensity across any environment where AI’s information ecosystem fails to track local reality.
4. The Hypothesis Machine
So is the solution to ban AI from modeling for firms looking to improve stakeholder and risk analysis? Hardly. Let us be equally clear about what AI does well. AI is a powerful hypothesis-generating machine whose outputs only become dangerous for firms when they are treated as conclusions. I highlight several promising avenues:
Background intelligence and pattern detection. AI is powerful at aggregating public information, identifying patterns in large structured datasets, scanning regulatory changes across jurisdictions, and surfacing connections that human analysts might miss. These can be everything from supply chain disruption signals and financial flow anomalies to media sentiment shifts and regulatory filing changes. These can be valuable inputs, provided they are interpreted by people with the contextual knowledge to distinguish signal from noise.
Scenario generation. LLMs also excel at ‘what if’ exploration: generating plausible scenarios, stress-testing assumptions, surfacing possibilities for the future that a human team anchored in conventional thinking might overlook. Under instability, this capacity for lateral exploration is valuable. Most of us working in high-risk environments have experienced the tunnel vision that crisis induces. AI can widen the aperture if these scenarios as hypotheses to be tested against local knowledge. The human-AI teaming literature has documented the conditions under which augmentation works well: when algorithmic outputs are treated as decision inputs that preserve meaningful human discretion over interpretation (Raisch & Krakowski 2021; Jarrahi 2018). The challenge under instability is that the conditions for effective augmentation are often the conditions that instability destroys.
Communication and translation. Drafting, summarizing, translating across languages and professional silos are core AI operational support functions that work under virtually any conditions. A firm operating in a multilingual crisis environment can use AI to bridge communication gaps, produce rapid summaries of local media, and draft communications in multiple languages. These are efficiency gains with low epistemic risk and can be encouraged, again with human verification.
Anomaly flagging. AI can also identify when data deviates from established patterns, which under instability can serve as an early-warning trigger for human investigation. The value lies in its ability to direct human attention to places where something may be changing. The anomaly may be noise, or the first change signal that a locally-grounded analyst can interpret. Table 1 maps how AI and human intelligence can work together across six dimensions, and where the handoff from algorithmic to human judgment occurs.
Table 1. Integrating AI and Human Intelligence under Instability
---------------------- ------------------------------------------------------------------------------ ---------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Dimension AI Contribution Human Contribution Integration Point Data Sourcing Synthesises published text, indexed reports, and Global North-centric media. Draws on informal networks, local media, and non-English discourse. Human teams annotate AI outputs with local sources the system cannot access; AI surfaces patterns across the sources humans provide. Temporal Awareness Backward-looking: calibrated to the recent past (3–12 month lag). Present-tense: registers immediate shifts in power and operating conditions. AI generates the baseline; human analysts flag where the baseline has expired. Temporal metadata tags indicate the age of underlying data. Stakeholder Mapping Maps digitally visible, formally organised, English-speaking stakeholders. Identifies informal power brokers, armed groups, and affected local communities. AI produces the initial map; locally-grounded staff conduct a ‘who is missing?’ audit before the map enters the decision chain. Interpreting Silence Registers as a data gap or engagement deficit. Recognises silence as a strategic warning of deteriorating safety. AI flags the absence, then human judgment with relational context determines whether absence is a data gap, a signal, or deliberate opacity. Risk Assessment Extrapolates from historical patterns to assign probability. Registers pattern breaks and shifts in the fundamental rules of the game. AI generates probabilistic scenarios; human validators test whether the underlying distribution still holds. Decision Support Hypothesis generator: identifies broad trends and ‘what if’ scenarios. Contextual validator: provides the ground truth to confirm or reject analytics. AI widens the aperture of possibilities; human judgment narrows to the actionable. No AI output proceeds to decision without contextual validation in high-risk settings. ---------------------- ------------------------------------------------------------------------------ ---------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Under this perspective, AI’s strengths and limitations are clarified. It excels at synthesis, pattern-matching, and extrapolation within known distributions. It fails when the distribution shifts, when the patterns break, when the stakeholders who matter most are those least visible to the system.
5. Why Local Knowledge Networks Pay for Themselves
The business case for maintaining human intelligence networks alongside AI-driven analytics is often framed as an extra cost. But local knowledge networks function as an insurance policy against the catastrophic misreadings that foreign tools of any type produce under instability (Miklian and Katsos 2025). And the insurance is cheap relative to what firms pay when they do not have it. BP’s $25.5 billion write-down on its Rosneft partnership in March 2022 was the single largest asset impairment in the company’s history. Shell’s $4 billion charge and TotalEnergies’ $4.1 billion impairment were driven by a geopolitical rupture that AI-driven risk platforms had collectively rated as a tail event. The firms that adapted fastest to the post-invasion environment, and there were firms that began contingency planning weeks before the invasion, did so on the basis of human intelligence: direct conversations with government contacts in Kyiv and Moscow, relationships with local staff who understood the political mood, and informal signals from business partners whose behavior had started to shift. The cost of building the local intelligence capacity that would have generated earlier warnings is a rounding error against a $25 billion write-down, and undoubtedly a small fraction of what most firms are investing on AI tools in 2026. A country-level team of locally embedded risk analysts, maintained relationships with civil society early-warning networks, and a regional intelligence coordinator with decision-relevant authority might cost a large multinational two to five million dollars per year in a high-risk market. BP’s Rosneft impairment represented five thousand years of that investment. Even if local intelligence improves the timing of a major strategic decision by a single quarter, the return dwarfs the cost.
The logic extends beyond catastrophic one-off events. Firms operating in volatile markets face a continuous stream of smaller decisions where the quality of information directly determines the quality of the outcome. This diagnostic has implications for how firms structure their decision workflows under instability. Most of us working in this space have seen firms cut local intelligence capacity as AI platforms scale up, treating algorithmic coverage as a substitute for human presence. The logic is understandable but the losses, when they come, are catastrophic. Managers can make two moves from immediate, low-cost adjustments to deeper changes that merge AI and human insights.
First, every AI-generated analysis of an unstable environment should enter the decision workflow as a starting hypothesis, not an assessment. An AI-generated country risk report must be assessed by a local intelligence team that annotates it with what has changed since the data was collected, who is missing from the stakeholder map, and which risk factors reflect actual conditions. Before any AI-generated risk or stakeholder analysis in a high-risk market reaches a decision-maker, a locally-grounded reviewer should add annotate the relationship between the AI’s confidence rating and their own assessments. Every AI output also carries an implicit timestamp: when was the training data collected, when was the model last updated, when were the risk indicators last refreshed. Building temporal metadata into AI-augmented decision workflows, explicit tags identifying the age of underlying data, can add value when incorporating crisis data.
Therefore, locally-grounded staff should be ‘red team’ validators for AI outputs in high-risk contexts. For every AI-generated stakeholder map, the red team asks: who is missing? For every risk assessment: what has changed since this data was collected? The red team annotates the AI output with these answers before it moves up the decision chain. Review cycles should be calibrated to context volatility. In an active crisis, the review cycle compresses to days, with a standing question: has anything happened since this assessment was generated that invalidates its assumptions? The person or team responsible for monitoring this gap needs the authority to flag when an assessment has expired, and decision-makers need organizational cues that compensate.
Second, stakeholder identification and salience assessment in unstable contexts cannot be delegated to algorithmic tools. The judgment calls, who matters, what signals to read, when the framework has shifted, require human situational awareness rooted in relationships, trust, and proximity to affected communities. If the firm’s obligations run to communities affected by its decisions (Freeman 1984), and AI systematically underrepresents those communities, then outsourcing stakeholder engagement to algorithmic tools is a failure of the very responsibility that stakeholder theory describes. Specific decision points in crisis-context workflows should be designated where AI inputs are excluded entirely and the judgment rests with people who have the contextual knowledge to make it. These nodes should include: initial stakeholder identification in a new crisis context before AI fills in background details, escalation decisions about whether to maintain, reduce, or exit operations in a deteriorating environment; and any decision that directly affects the physical safety or livelihoods of local communities. The structural investment is the country-level intelligence infrastructure described above: locally embedded analysts, maintained relationships with civil society early-warning networks, a regional coordinator with decision-relevant authority.
Of course, these implications add cost, complexity, and friction to the exact decision processes that AI promises to streamline. But the firms that navigate instability successfully by maintaining operational continuity, protecting their staff, and avoiding complicity in harm, will be those that understood which questions are algorithmic and which are human, and why the two work best together. The pathway applies across sectors, though the specific decision nodes will vary. A pharmaceutical firm navigating regulatory instability in sub-Saharan Africa faces different stakeholder dynamics than an energy company managing extraction permits in post-conflict settings or a financial services firm assessing sanctions compliance in a fragmenting regulatory landscape. The underlying logic is the same: AI’s information ecosystem alone fails to track local reality in ways that are specific to the sector, the geography, and the crisis type.
6. Conclusion
AI delivers knowledge through synthesis, pattern-matching, and extrapolation from documented pasts, while crisis and instability are fundamentally outside the boundaries of their training data. Many incredibly consequential business decisions are made during times of instability and crisis, and AI alone cannot close this gap. Stakeholder theory helps us see why: the communities whose stakes are highest are those AI sees least clearly. Algorithmic invisibility, the systematic and undetectable exclusion of consequential stakeholders by tools that appear comprehensive, is not a temporary limitation awaiting a technical fix. The three components of the present that this article has identified, perceptual immediacy, tacit operational knowledge, and relational trust, are categorically different from the kinds of knowledge that improved data pipelines can provide.
Implications for practice. The pathway in Section 5 translates the argument into operational terms from immediate workflow adjustments to long-term strategic investment. The core principle across is that AI outputs under instability should never travel unaccompanied to a decision-maker. In short: the death of human intelligence has been exaggerated (Ion 2026). The cost of the local intelligence capacity that accompanies them is a rounding error against the write-downs that follow when it does not exist. We are living in a polycrisis environment where instability is the baseline condition (Miklian & Katsos 2024), and the firms that treat AI as an answer machine will be the ones most likely to make decisions that harm the communities they claim to serve, and their own bottom line.
Implications for research. Three agendas merit priority. First, comparative studies of AI-augmented versus human-centered decision-making in high-risk business environments would provide the empirical evidence base that this field currently lacks. How do firms that rely heavily on AI-driven intelligence perform, in terms of both financial outcomes and stakeholder welfare, compared to firms that maintain robust human intelligence networks? The Russia-Ukraine and Iran crisis episodes provide natural experimental variation that future research should exploit. Second, the interaction between mandatory due diligence frameworks and AI-automated compliance processes in crisis settings requires urgent attention. As firms adopt AI tools to meet due diligence obligations in volatile environments, the gap between compliance documentation and actual stakeholder impact may widen. Third, the concept of algorithmic invisibility needs to be tested empirically against specific firm-level cases and refined into a more formal theoretical framework. The 40-country data reported in the appendix provides a starting point, but the construct requires validation across models, platforms, and crisis types.
The firms that will navigate the coming decades of instability most effectively will be those that invest in the human knowledge infrastructure that AI cannot replace. The algorithms will keep getting better. The question is whether the people interpreting their outputs will have the relationships, the presence, and the judgment to know when the algorithm is wrong, and why.
Declaration of AI Use
Claude CoWork was used during the preparation of this work for data analysis support, editorial refinement by acting as a virtual peer-reviewer for drafts, and reference formatting (see Appendix for details). The author takes full responsibility for all ideas and content in this publication.