Executive proposition
Artificial intelligence is usually described as a new technology for doing work. That description is too small.
AI is changing who can know, who can propose, who can decide and who can act inside an organisation. It is therefore changing not only productivity, but power.
For most of corporate history, information moved upward and permission moved downward. Employees produced fragments of reality. Managers collected them. Senior managers compressed them. Executives interpreted the summaries and authorised action.
Hierarchy was not merely a system of accountability. It was an information-processing architecture built for a world of limited human attention.
That constraint is collapsing.
An employee with access to capable AI can now analyse data, interrogate documents, model scenarios, write software, challenge a strategy and produce an executive-quality recommendation without waiting for the traditional chain of interpretation.
This creates the defining contradiction of the AI enterprise:
Leaders demand machine speed while preserving human permission chains.
They ask employees to use AI, but not to alter how authority works. They ask the organisation to transform, but protect every layer built for the previous system. They want distributed intelligence without a corresponding redistribution of agency.
Governance is essential. But governance and hierarchical control are not the same thing.
Nor can governance begin with the model. Before an organisation distributes machine intelligence, it must confront the neglected condition beneath it: the quality of its own memory. What does the company actually know? Who defined it? When did it stop being true? Which facts may a machine retrieve, which claims may it make and which consequences is the institution willing to own?
An AI strategy without a data-quality strategy is not an intelligence strategy. It is an acceleration strategy for institutional error.
The company of the future must democratise cognition, govern consequence and make responsibility explicit. It should be permissionless where failure is reversible and deliberately slow where consequences are not.
The objective is not to eliminate hierarchy. It is to replace a hierarchy of information with a hierarchy of consequences.
Information moved up. Permission moved down.
The traditional company solved a real computational problem.
No individual could perceive the whole organisation. Information was expensive to collect, difficult to compare and slow to distribute. Managers became compression layers. Each level reduced complexity for the level above it.
The frontline saw reality but lacked authority. The executive possessed authority but saw reality through summaries.
Power accumulated around four scarcities:
- Access to information.
- Capacity to interpret it.
- Permission to decide.
- Ability to execute.
Seniority often provided all four. The more elevated the position, the broader the information, the greater the interpretive legitimacy and the larger the field of authorised action.
This arrangement produced coordination, but it also produced distance. Bad news weakened as it travelled upward. Context disappeared inside presentations. Political skill influenced which facts survived. Decisions returned downward after the conditions that motivated them had already changed.
The hierarchy was slow because human cognition was scarce.
It was also powerful because information was scarce.
AI attacks both assumptions.
The most important corporate resource is becoming available on demand.
The shift is no longer hypothetical. Stanford’s 2026 AI Index reports that organisational AI adoption reached 88%. It also describes a rapidly falling barrier to advanced capability: generative AI reached 53% population adoption within three years, faster than either the personal computer or the internet.1
Microsoft’s 2025 Work Trend Index, based on 31,000 workers across 31 countries plus Microsoft 365 telemetry, found that 82% of leaders considered that year pivotal for rethinking strategy and operations. 81% expected agents to be moderately or extensively integrated into AI strategy within 12 to 18 months, while 46% said their organisations were already using agents to automate workflows or processes fully.2
The speed pressure is equally visible. Microsoft found that 53% of leaders believed productivity needed to increase, while 80% of the workforce reported lacking enough time or energy to complete its work. Employees were interrupted, on average, every two minutes during the working day.2
AI arrives inside this capacity crisis as intelligence on demand.
But it does not distribute only efficiency. It distributes capabilities that organisations previously concentrated in specialised functions and senior roles: research, synthesis, analysis, modelling, writing, coding and strategic argument.
Evidence suggests that these capabilities can disproportionately benefit people with less prior experience. In a study of 5,172 customer-support agents, access to a generative AI assistant increased productivity by 15% on average. The largest improvements occurred among less experienced and lower-skilled workers, while the most experienced workers saw smaller gains and, in some cases, small quality declines.3
This matters politically.
If AI transfers elements of the organisation’s best practices to less experienced workers, expertise does not disappear. But part of its operational advantage becomes reproducible. The distance between the person who traditionally advised and the person who traditionally waited for advice begins to shrink.
AI does not make everyone equally wise.
It makes sophisticated cognition less dependent on rank.
Companies want the output of decentralisation without the loss of central control.
The modern executive message is unambiguous: move faster, experiment, adopt AI, challenge assumptions, transform the business.
The organisational message is often the opposite: request access, wait for approval, remain inside the function, do not expose unfinished work, do not alter the process, do not bypass the hierarchy.
The result is an impossible instruction:
Move faster. But ask permission.
Use AI. But do not change how decisions are made.
Transform the company. But preserve every existing role.
Be autonomous. But remain inside the hierarchy.
This is not evidence that every senior leader consciously fears losing power. The data cannot establish that motive, and many restrictions are justified by security, legal obligations, privacy, reliability or customer protection.
But a structural incentive exists whether or not anyone names it.
When status depends partly on privileged access to information, technology that broadens that access threatens status. When managerial value depends on carrying, summarising and validating information, technology that performs those functions threatens the role. When authority is associated with being the person who knows, intelligence available to everyone feels like institutional disorder.
Resistance rarely declares itself as resistance. It appears as process.
“We need another committee.”
“Only the centre of excellence can use the advanced tools.”
“We cannot begin until the policy is complete.”
“All outputs must move through the existing approval chain.”
Any one of these statements may be responsible. Together, they can reconstruct the old monopoly around the new technology.
Governance then becomes the vocabulary through which hierarchy protects itself.
The productivity is real. The institutional value is not automatic.
McKinsey’s 2026 global survey captures the gap. Nearly nine in ten respondents reported regular AI use in at least one business function, and 80% said AI had improved their individual productivity. Yet only 37% attributed any positive EBIT impact to AI, essentially unchanged from the previous year. Just 6% qualified as AI high performers, meaning they attributed at least 5% of EBIT to AI and described its impact as significant.4
This is not primarily a model-performance problem. It is an organisational-design problem.
Individual employees can become faster while the enterprise remains slow. They produce more drafts, analyses, code and recommendations, but those outputs enter the same meetings, approval chains, budgets and functional boundaries. The local work accelerates; the institutional decision does not.
AI poured into the old operating model can increase activity without increasing movement.
McKinsey’s high performers offer a revealing contrast. Nearly three-quarters reported fundamentally redesigning workflows around AI, compared with only one-quarter of other organisations. They were also more likely to show senior leadership commitment, define processes for measuring impact and actively mitigate a broader set of AI risks.4
The organisations moving fastest are not abandoning governance.
They are combining transformation with governance.
That distinction is fundamental:
Weak governance asks whether AI is allowed. Mature governance defines how far it may act.
The org chart is ceasing to describe where work happens.
The organisational chart assumes that capability follows position and work follows reporting lines.
AI breaks both assumptions.
A junior employee may direct a collection of agents capable of research, analysis, design, coding and communication. A five-person group may perform work that previously required several departments. An individual may possess more synthetic analytical capacity than a traditional management layer.
Microsoft argues that human-agent teams may replace parts of the static org chart with a dynamic “work chart” organised around outcomes. Its research also found a significant leadership–employee capability gap: 69% of leaders reported using AI regularly compared with 45% of employees, and 67% of leaders were familiar with agents compared with 40% of employees.2
This gap contains a warning.
AI may theoretically democratise intelligence while organisations distribute its strongest forms first to the people who already possess authority. If access, data and autonomy remain concentrated at the top, AI may reinforce hierarchy rather than flatten it.
The relevant question is therefore not whether AI is available somewhere in the company.
It is who may use which intelligence, against which information, with what freedom to challenge existing decisions and with what authority to act.
Without those answers, the formal company and the operational company separate. The formal company consists of roles, committees and reporting lines. The operational company consists of prompts, agents, hidden workflows, private datasets and machine-mediated decisions.
That second company may become faster than the first—and increasingly illegible to it.
AI does not repair the organisation's memory. It industrialises whatever it finds there.
Executives speak about models because models are new, visible and purchasable. They speak less about product taxonomies, duplicated customers, contradictory definitions, forgotten permissions, stale documents and spreadsheets whose authority depends on who last edited them.
Yet this is where enterprise intelligence begins.
A model can be brilliant and the company around it epistemically broken. It can reason perfectly over the wrong price, retrieve an obsolete policy, combine two incompatible definitions of revenue and answer with such fluency that the error acquires the appearance of institutional truth.
The machine is not hallucinating alone. Sometimes it is faithfully expressing the contradictions the organisation has spent years refusing to resolve.
And whoever controls the definitions still controls the organisation. Democratising access to a centrally curated reality does not democratise truth. It can simply allow everyone to query the assumptions of the powerful more efficiently.
Data is often described as a raw material, as if it existed before interpretation. It does not. Every enterprise dataset is a sediment of human choices: what to count, what to ignore, which categories to create, which exceptions to conceal and whose reality the system was designed to recognise.
“Customer.” “Risk.” “Productivity.” “High performer.” “Fraud.” “Revenue.” None of these is merely a field in a database. Each is an institutional judgment wearing the clothes of a fact.
AI turns those judgments into operating power.
This is why data quality cannot remain a periodic clean-up exercise performed before a migration. Quality decays. Meanings drift. Sources change. Permissions outlive the purpose for which they were granted. A field can be technically complete and institutionally false.
The NIST AI Risk Management Framework makes the connection explicit: AI introduces failure modes tied directly to the quality and representativeness of data, and production systems require ongoing monitoring for drift, anomalies, unreliable outputs and feedback loops.5 The European Union's AI Act similarly treats data governance, relevance, representativeness, completeness and error control as substantive conditions for high-risk systems, and requires post-market monitoring rather than assuming that pre-deployment testing settles the question.6
But the deeper point is not regulatory.
An organisation cannot automate judgment before it knows which version of reality it is automating.
Every consequential data product should therefore have an owner, an origin, a definition, a freshness threshold, a permitted purpose and an expiry condition. Quality should be continuously observed across completeness, accuracy, timeliness, lineage and representativeness. When a critical threshold fails, the system should not merely colour a dashboard red. Its authority to speak or act should contract.
This is not data housekeeping. It is the maintenance of institutional reality.
When intelligence becomes cheap, the quality of reality becomes the scarce asset.
Fluency is not authority.
Most companies govern access: who can open a file, query a system or use a model. Far fewer govern assertion: what the system may present as fact, what it must label as inference, what it may recommend and what it may never say on behalf of the institution.
The distinction is essential.
A machine may be permitted to read a medical document without being authorised to diagnose. It may analyse an employee's record without being authorised to declare the person a poor performer. It may inspect financial data without being authorised to promise a return. It may draft a legal position without being authorised to represent it as the company's position.
Access permission is not claim permission. Claim permission is not action permission.
Yet conversational systems collapse these boundaries psychologically. They speak in complete sentences. They rarely look uncertain in proportion to their uncertainty. They make retrieval, inference, recommendation and commitment sound like variations of the same act.
They are not.
The company must decide which statements require a source, which inferences require a confidence signal, which recommendations require a qualified reviewer and which commitments require a named human principal. It must define forbidden claims not only by topic, but by consequence: statements about safety, employment, credit, health, legality, financial exposure, customer rights and institutional intent.
These boundaries must be versioned, tested and monitored in production. A policy written in prose but absent from the system's behaviour is not governance. It is aspiration.
There is, however, an equally dangerous error: confusing the governance of institutional claims with the governance of human thought. A company should control what an automated system may assert in its name; it should not use “AI safety” to suppress employee dissent, inconvenient evidence or legitimate disagreement. Otherwise the machine becomes something older than intelligence: an instrument of orthodoxy.
The boundary should be hard around deception, unsupported certainty and unauthorised commitment—and radically open around questions, hypotheses and challenge.
And the rule must apply upward as well as downward. If an employee's AI-generated claim requires evidence, an executive's claim should not become true through rank. The same machinery that constrains the model should expose the unsupported certainty of human authority.
This is where AI governance becomes politically difficult. It promises to limit machines, but done honestly it also limits people who have long been permitted to speak without showing their work.
The question is not only whether the machine is correct. It is whether the institution has earned the right to make the claim.
The fastest company may become the company that no longer understands itself.
Every AI system added to a workflow increases capability. It may also add invisible dependencies, ambiguous ownership, inconsistent rules and decisions that cannot be reconstructed later.
This accumulation is governance debt.
Like technical debt, governance debt is initially experienced as speed. Teams route around controls, connect tools, automate decisions and achieve visible gains. The cost appears later: duplicated agents, contradictory outputs, data leakage, unexplained customer treatment, uncertain accountability and systems no one feels authorised to stop.
The external evidence suggests that capability and control are already diverging. Stanford’s 2026 AI Index recorded 362 documented AI incidents, up from 233 in 2024, while describing responsible-AI measurement as lagging behind capability development.1
The problem is not solved by placing a human nominally “in the loop.” A person asked to approve hundreds of machine-generated actions may provide ceremony rather than judgment. Oversight without time, context or genuine authority is not governance.
Governance must be designed into the operating system of work.
NIST’s AI Risk Management Framework offers a useful foundation through four continuous functions: govern, map, measure and manage. But companies need to translate principles like trustworthiness, transparency and accountability into the daily mechanics of authority.5
Policies that live in documents cannot govern systems acting continuously.
The rules must become executable.
Govern the space of possible action, not every individual action.
Traditional governance reviews decisions after people have assembled them. Agentic governance must shape the environment in which decisions become possible.
Every human and machine agent should operate within a constitution defining:
- The objective it may pursue.
- The data it may access.
- The claims it may make and the evidence those claims require.
- The actions it may take.
- The money it may commit.
- The uncertainty it must disclose.
- The evidence it must preserve.
- The conditions requiring escalation.
- The person accountable for the system.
- The mechanism for interruption and reversal.
This is not centralised micromanagement. It is bounded autonomy.
The better the boundaries, the less frequently central authority needs to intervene.
An organisation with vague principles and universal approval requirements will move slowly. An organisation with explicit constraints, observable behaviour and defined escalation thresholds can move rapidly because teams know where autonomy begins and ends.
The purpose of governance is not to create more permission. It is to make safe action possible without repeatedly requesting it.
Autonomy should be allocated like capital.
Companies maintain financial budgets because capital creates both opportunity and exposure. Machine autonomy should be governed with similar seriousness.
An autonomy budget defines how much independent action a person–AI system may exercise before escalation. It should expand or contract according to five variables:
- Impact — How many people, customers or assets can the action affect?
- Reversibility — Can the action be undone quickly and completely?
- Uncertainty — How reliable are the model, information and causal assumptions?
- Externality — Can the action impose costs on people outside the immediate objective?
- Legitimacy — Is machine action socially, morally and legally acceptable in this domain?
This produces a hierarchy of consequences:
| Consequence | Default operating mode |
|---|---|
| Low impact, reversible | Autonomous execution with logging |
| Material but recoverable | Execution within limits, sampled review |
| High impact or difficult to reverse | Human approval with evidence |
| Rights, safety, employment or institutional purpose | Deliberate human decision and named accountability |
Authority should scale with consequence, not seniority.
A junior employee should be able to produce and circulate a better hypothesis than the CEO without navigating five political layers. But no employee, executive or agent should make an irreversible, high-impact decision merely because the technology makes it possible.
This is how cognition can be democratised without making execution chaotic.
In the AI enterprise, work without provenance is like money without accounting.
AI-generated work should carry its own chain of evidence.
Every consequential output should reveal:
- Which system produced it.
- Which information it used.
- Which instructions and objectives shaped it.
- Which assumptions remain uncertain.
- Which policy authorised the action.
- Which human owns the outcome.
- How the decision can be challenged or reversed.
This changes governance from retrospective investigation to continuous legibility.
It also changes power.
In the old company, seniority often allowed assertions to travel without evidence. In a proof-carrying company, claims become inspectable regardless of who makes them. A machine-generated strategy and a CEO-generated strategy should both expose their assumptions.
The democratisation of intelligence should therefore be accompanied by the democratisation of scrutiny.
Nobody should retain epistemic privilege merely because of position.
Some decisions should remain slow precisely because machines can make them fast.
Speed is not a universal good.
Hiring, dismissal, credit, medical access, pricing during scarcity, safety decisions and changes to institutional purpose involve more than prediction. They allocate dignity, opportunity, risk and power.
In these domains, friction may be a moral requirement.
The objective is not always to keep a human inside every operational step. It is to identify the moments at which responsibility cannot legitimately be delegated.
The AI-native company should therefore have two speeds:
- Machine speed for exploration, simulation, drafting and reversible execution.
- Human time for irreversible commitments, contested values and consequences borne by others.
The distinction is not between important and unimportant work. It is between recoverable and irrecoverable action.
Move fast where reality permits repair. Move deliberately where people must live with the result.
What is a manager for when carrying information is no longer a job?
Management will not disappear, but its justification must change.
The manager of the information-scarcity era gathered updates, distributed tasks, translated between levels, reviewed output and acted as the local gateway to senior authority.
AI can perform growing parts of that work.
The manager of the intelligence-abundance era must instead:
- Define outcomes rather than distribute fragments of activity.
- Design boundaries for human and machine autonomy.
- Build the judgment of people rather than monopolise answers.
- Resolve exceptions that systems cannot legitimately resolve.
- Protect dissent and surface inconvenient evidence.
- Maintain coherence across fast-moving, agent-mediated work.
- Accept responsibility for the conditions under which machines act.
This is a more demanding role.
It is also a less comfortable one because authority can no longer be defended through superior access to information.
Leadership becomes constitutional rather than informational.
The leader’s task is not to remain the most knowledgeable intelligence in the room. It is to design a room in which intelligence can move without allowing responsibility to disappear.
Democratising capability does not make specialists less necessary. It makes superficial expertise less defensible.
The arrival of AI has produced another seductive prediction: if everyone can analyse data, generate code and interrogate a model, the data scientist, engineer, researcher and domain specialist become less important.
This confuses access to an instrument with mastery of a discipline.
A calculator made arithmetic abundant; it did not abolish mathematics. Search made facts accessible; it did not abolish judgment. Generative AI can make the forms of expertise widely reproducible without reproducing all the substance beneath them.
The study of customer-support agents is instructive precisely because its effects were unequal. Less experienced workers gained the most from AI, while the most experienced gained less and sometimes experienced small quality declines.3 AI raised the floor by transferring patterns from strong performers. It did not prove that expertise had become worthless. It showed that one part of expertise could be distributed.
What remains becomes more valuable.
Data scientists will be needed not merely to build models, but to decide whether the measurement deserves to exist; whether the target variable represents the stated objective; whether correlation is being sold as causation; whether missing data hides a population; whether a benchmark survives contact with reality; and whether a system that performs well in aggregate fails the people who matter most.
The highest form of expertise will move from producing answers to protecting the conditions under which an answer can be trusted.
But expertise alone is not enough. An organisation can contain brilliant specialists and still become intellectually sterile if every problem must be expressed in the categories the system already understands.
This is why lateral thinking becomes more—not less—important in an AI-rich world.
Models are exceptionally powerful inside a frame. They extend patterns, combine precedents and optimise legible objectives. The human advantage increasingly lies in moving outside the frame: noticing the absent variable, connecting domains that the organisation separated, asking whether the requested optimisation is worth pursuing and refusing the premise of a perfectly answered wrong question.
Efficiency improves the answer.
Lateral thought changes the question.
Yet organisations routinely demand originality while punishing deviation. They hire unusual minds and subject them to ordinary permission. They celebrate innovation in speeches and require conformity in workflow. AI will intensify this contradiction because it makes conventional competence cheap while the organisation continues to suppress the independence that produces unconventional insight.
Autonomy is therefore not an employee benefit added after the strategy. It is part of the intelligence architecture.
People need protected space to explore, combine, test, contradict and fail without converting every uncertain idea into an authorised corporate act. The organisation should widen cognitive autonomy while bounding consequential autonomy: radical freedom to question; explicit limits on what may be executed in the world.
This is also the future of the specialist. Not a priest guarding access to scarce knowledge. Not a technician reduced to checking machine output. A designer of questions, a challenger of measurements, a translator of uncertainty and a custodian of reality.
AI will commoditise the plausible answer. It will increase the value of the person capable of asking the illegible question.
Faster companies do not automatically create better work.
The World Economic Forum’s Future of Jobs Report 2025, based on more than 1,000 employers representing over 14 million workers, projected structural transformation affecting 22% of current jobs by 2030: 170 million roles created and 92 million displaced. It also estimated that 39% of workers’ existing skill sets would be transformed or become outdated.7
Employers simultaneously expressed expansion and contraction: half planned to reorient their businesses in response to AI, two-thirds expected to hire people with specific AI skills and 40% anticipated reducing staff where AI could automate tasks.7
This ambiguity is not a forecasting failure. It is the nature of the transition.
AI can augment workers, compress teams, remove entry-level tasks, enable new businesses and concentrate returns—all at the same time.
McKinsey’s 2026 survey found that only 14% of respondents from AI-using organisations reported an overall workforce decline attributable to AI during the previous year, far below the 32% who had expected reductions. Yet 39% expected reductions in the following year.4
The honest position is therefore neither mass-unemployment certainty nor effortless abundance.
The outcome will depend on institutional choices: who receives the tools, who owns the productivity gains, which capabilities are developed, which roles remain entry points and whether employees gain greater agency or simply face greater output expectations.
Democratised intelligence can decentralise opportunity.
It can also create a more efficient system of central control.
Technology does not decide between those futures. Governance does.
The AI-native company requires a new agreement between leadership, workers and machines.
1. Intelligence should be broadly accessible
Access to capable systems should not be a status privilege. Differences should be justified by data sensitivity and consequence, not hierarchy.
2. The right to propose should be almost universal
Ideas should travel according to evidence, not reporting lines. AI should make it easier for reality at the edge of the company to challenge belief at the centre.
3. The right to execute should be conditional
Execution authority should depend on impact, reversibility, uncertainty and legitimacy.
4. Responsibility must remain named
“The model decided” is not an acceptable organisational sentence. Every deployed system needs a human owner and an institutional principal.
5. No governed data, no governed intelligence
Every critical source needs explicit meaning, provenance, ownership, freshness and permitted use. Data quality is not a project completed before deployment; it is a condition continuously defended after deployment.
6. The right to speak must be governed separately from the right to know
Access to information does not authorise a system to convert that information into fact, recommendation, promise or institutional position. Fluency must never be mistaken for mandate.
7. Governance must operate at runtime
Controls, monitoring, provenance, limits and escalation cannot exist only in policy documents or quarterly committees.
8. Productivity gains must purchase capacity, not only extraction
If every minute saved becomes another minute of expected output, AI will intensify work without improving it. Some gains should become learning, experimentation, resilience and time.
9. Leadership must surrender epistemic monopoly
Seniority can still confer decision responsibility. It cannot confer immunity from evidence.
10. Expertise must be distributed without being trivialised
AI should widen access to specialist capability while preserving investment in data science, engineering, research and domain mastery. Their value shifts from controlling the answer to safeguarding the question, the measurement and the limits of inference.
11. Cognitive autonomy must be wider than operational autonomy
People should be free to question, explore and contradict far beyond the limits within which people or machines may execute. Innovation requires permission to think without automatic permission to impose consequences.
After hierarchy
AI is not merely automating corporate work.
It is destroying the scarcity of intelligence on which much of corporate hierarchy was built.
This does not make hierarchy obsolete. Organisations still need coordination, accountability and legitimate authority. But the basis of hierarchy must change.
The old hierarchy answered:
Who is allowed to know?
The new hierarchy must answer:
Who is accountable for the consequence?
Companies now face an uncomfortable choice. They can distribute the ability to understand and propose, redesign authority around consequence and construct governance capable of moving at machine speed. Or they can preserve the old permission architecture and accept that their transformation will remain trapped inside presentations about transformation.
The winners will not be the organisations that automate the most. Nor will they be those that accumulate the largest number of models while allowing the foundations beneath them to rot.
They will be the organisations that make intelligence abundant, reality inspectable, action bounded and responsibility impossible to escape.
The greatest barrier to the AI-native company will not be the intelligence of its machines. It will be the fear of the humans whose authority depended on everyone else knowing less.
And the ultimate test of leadership will not be whether leaders can control AI.
It will be whether they can stop using governance to control the people AI has empowered.
The empirical claims in this memorandum are drawn from the sources below. The argument about hierarchy, power and institutional resistance is AJL Innovation Lab’s interpretation of those findings; it should not be attributed to the cited organisations.
1 Stanford Institute for Human-Centered Artificial Intelligence, 2026 AI Index Report ↗.
2 Microsoft WorkLab, 2025: The Year the Frontier Firm Is Born ↗, based on 31,000 workers in 31 countries, LinkedIn labour-market data and Microsoft 365 signals.
3 Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work ↗, study of 5,172 customer-support agents.
4 McKinsey & Company, The State of AI: Global Survey 2026 ↗, survey of 1,719 participants across 97 nations, conducted May–June 2026.
5 U.S. National Institute of Standards and Technology, AI Risk Management Framework ↗.
6 European Union, Regulation (EU) 2024/1689 — Artificial Intelligence Act ↗, especially Article 10 on data and data governance and Article 72 on post-market monitoring.
7 World Economic Forum, Future of Jobs Report 2025 ↗, based on more than 1,000 employers representing over 14 million workers.