Trust in AI Must Be Earned by Leadership

A leader announces an artificial-intelligence strategy and asks employees to embrace the future. The presentation is polished, the efficiency targets are ambitious and the new platform is already licensed. Yet the people expected to use it do not know what data it collects, which roles may change, how performance will be judged or who can overturn a harmful recommendation.

The leader calls the hesitation resistance. Employees call it experience.

Trust is increasingly presented as the missing variable in artificial-intelligence transformation. That insight is important and incomplete. Trust is not a communication problem to be solved after leaders have made every consequential choice. It is a judgement people form about competence, motives, fairness, vulnerability and the possibility of correction.

Leaders should not ask how to make people trust AI. They should ask what evidence would make confidence warranted, what uncertainty should remain and what institutional protections must exist when the system is wrong.

A current leadership event makes trust the strategic question

BetterUp has scheduled an online conversation with Adam Grant for 10 August 2026 under the title The Hidden Variable: The Impact of Trust on Your AI Strategy. The event page states that trust in leadership predicts AI return on investment better than adoption rates.

This is a provider announcement and promotional claim for an event scheduled later on the publication day. It verifies the topic, date and advertised proposition. It does not by itself disclose the underlying model, establish causality or demonstrate that increasing trust will produce financial returns in every organisation.

BetterUp's accompanying Case for Conditions report says workforce conditions can cause similar adoption patterns to produce different performance outcomes. It highlights mattering, judgement, coaching culture and a distinction between managers who calibrate human-AI work and those who merely automate it. The public summary includes provider-reported associations, not enough methodological detail for an independent causal assessment.

The useful leadership question therefore lies beneath the marketing language. If adoption can rise while judgement and performance decline, what kind of trust should leaders cultivate—and what kind should they refuse to manufacture?

Trust should be warranted, not maximised

High trust is not always good. A person can trust an inaccurate system, a charismatic executive or an opaque process more than the evidence justifies. Low trust can also be rational when an institution has concealed errors, punished questions or shifted risk onto people with little power.

The aim of values-based leadership is calibrated trust: confidence proportionate to demonstrated reliability, transparent purpose, fair process and effective safeguards. Leaders should want people to rely on a system where it is competent, question it where uncertainty is material and refuse its use where the consequences exceed its evidence or mandate.

This distinction matters because adoption dashboards can mislead. Logins, prompts and completed automations show activity. They do not establish that work improved, people were treated fairly or the organisation preserved capabilities it will need when the system fails.

A trustworthy institution makes scepticism useful. It converts responsible doubt into testing, correction and learning instead of treating every question as disloyalty to innovation.

Human leadership fairness changes how people read AI fairness

A paper published in the 2026 Academy of Management Proceedings examines employee trust in embedded AI through signalling theory. Its studies report that people distinguish signals of AI fairness from signals of human leadership fairness, and that perceived leadership fairness shapes how strongly AI fairness relates to trust—especially when AI transparency is low.

The proceedings paper provides an abstract of preliminary conference research rather than a full, mature evidence base. Its populations, measures and organisational transfer should be examined before broad generalisation.

Its central insight is nevertheless powerful: employees do not encounter a system separately from the institution deploying it. They interpret an algorithm through prior experience of leaders. A technically careful system introduced by unfair managers will inherit institutional suspicion. A trusted leader cannot permanently rescue a system that repeatedly causes harm.

Leaders therefore earn AI trust long before procurement. They earn it when appointments are fair, reasons are given, commitments are kept, errors are acknowledged and people can challenge a decision without retaliation.

AI leadership is an operating-model responsibility

A 2026 case study of a large insurance company found that generative AI initially resembled other technology-adoption challenges, but over time altered work content, team relationships and the leader's role. The authors argue that leaders must move from functional control towards orchestration of interdependent human-AI work.

One in-depth organisational case cannot establish a universal leadership model. It offers situated empirical insight and propositions that require testing across sectors, cultures and organisational sizes.

Orchestration should not become a fashionable word for unclear accountability. A leader must still define which work the system may perform, which decisions remain human, when a person must intervene, whose expertise governs exceptions and who answers for the result.

A sound operating model distinguishes four forms of authority: the authority to frame the problem, to generate or analyse options, to make the decision and to review or overturn it. If all four quietly migrate into one system or vendor workflow, human oversight may remain in the policy while disappearing from practice.

Consequential people decisions require legitimacy, not convenience

A systematic review of 87 high-ranked studies on AI-enabled human-resource management identifies three recurring tensions: objectivity against embedded bias, efficiency against procedural dignity, and automation against accountable human oversight. It concludes that legitimacy depends on whether decisions are explainable, contestable, auditable and institutionally defensible.

The literature remains concentrated in recruitment and selection, and a systematic review synthesises the available evidence rather than proving that one governance arrangement will work in every institution.

Its practical requirements are demanding: bias testing, named accountability, explanation, thresholds for human review, appeals, audit trails, vendor controls and a route for employee or applicant voice. These are not technical accessories. They are leadership commitments concerning how power may be exercised over a person.

No leader should allow AI to become the invisible author of hiring, appraisal, promotion, compensation, monitoring, discipline or termination. Assistance may inform a lawful decision. Responsibility must remain attached to an authorised person and a reviewable process.

African public administration adds the public-value test

A 2026 review of AI-driven chatbots in African public administration synthesised 32 sources through technology-organisation-environment and public-value perspectives. It identified service backlogs and resource constraints as pressures for adoption, while organisational readiness, leadership, workflow design, data governance, regulatory capacity and sovereignty shaped sustained value and legitimacy.

The review also identifies unresolved trade-offs involving responsiveness, equity, transparency, trust and accountability. It finds gaps in empirical evaluation, especially around effects, equity and responsibility in African local-government contexts.

This is a literature synthesis, not evidence that a particular African chatbot improves a public service. Its importance is conceptual and regional: a fast response is not public value if it excludes people with limited connectivity, gives an incorrect answer without remedy or hides the public officer responsible for the service.

Africa's contribution to global AI leadership should not be reduced to rapid adoption under constraint. It should include governance models that treat relationship, public duty, linguistic plurality, community knowledge and remedy as integral parts of system performance.

Human oversight must be capable, authorised and resourced

Many policies promise a human in the loop. That phrase can conceal a person who sees only the system's final score, lacks time to investigate, does not understand the model, fears disagreeing with management or has no authority to change the outcome.

Meaningful oversight requires cognitive access, domain competence, relevant information, decision time, institutional authority and protection when a reviewer raises a concern. The reviewer should know the system's purpose, limitations, likely failure modes and the evidence needed to depart from its recommendation.

The OECD's work on AI risks in employment highlights privacy, work intensification, bias, discrimination and accountability concerns. Workers reporting AI-related data collection have also expressed substantial concern about pressure, privacy, excessive collection and biased decisions. These are survey-reported experiences rather than proof that every monitored workplace produces the same harm.

Boards should ask more than whether human review exists. They should observe a sample of real reviews, examine reversals and escalations, test whether reviewers can obtain necessary evidence and verify that the people affected can reach a human with power to act.

Psychological safety is part of AI assurance

AI systems create new reasons for employees to remain silent. A junior analyst may notice a false pattern but assume the model knows more. A manager may conceal an error after sponsoring an expensive platform. A contractor may fear that reporting bias will end the relationship. A public officer may follow an automated recommendation because departure requires a written explanation.

Amy Edmondson's research on psychological safety helps explain why a learning institution needs credible permission to ask, disagree, report and admit uncertainty. Patrick Lencioni's emphasis on trust and productive conflict, Brené Brown's work on courageous vulnerability and Marshall Goldsmith's attention to feedback reinforce different parts of the same leadership discipline.

Psychological safety does not remove standards or accountability. It makes problems discussable early enough for responsibility to operate. People must still support claims with evidence, respect confidentiality and accept correction.

Every consequential AI workflow should contain a protected stop mechanism: who may pause the process, what evidence triggers escalation, how urgent cases are handled and how leaders prevent retaliation against a person who raises a good-faith concern.

Popular leadership ideas must become decision architecture

John Maxwell's language of influence reminds leaders that formal authority does not create credibility. Stephen Covey's principle-centred trust asks whether conduct remains consistent when convenience changes. Simon Sinek and Myles Munroe place purpose before technique, while Jim Collins insists that disciplined institutions must become stronger than one personality.

Ram Charan's work on leadership transitions and decision responsibilities helps clarify who should decide at each level. Liz Wiseman's multiplier perspective asks whether leaders enlarge or suppress the intelligence around them. John Kotter's change tradition requires a credible coalition rather than executive announcement alone. Ken Blanchard's developmental approach reminds leaders that direction and support should match readiness.

Robin Sharma, Craig Groeschel and Sam Adeyemi repeatedly translate leadership into daily practice, stewardship and service. Fred Swaniker's institution-building work insists that Africa's future requires leadership capacity at population scale. Adam Grant invites leaders to rethink assumptions; Amy Edmondson shows why people need safety to surface what leaders do not know.

These are established contributions, not a single validated AI-governance model. Their value today is realised when purpose, character, learning, influence and human development become explicit decision rights, review practices and institutional safeguards.

Faith leadership must not baptise automation

Faith-based institutions may use AI for administration, communication, translation, research, learning and service coordination. They carry an additional risk: spiritual language can make a technical system appear more authoritative than it is.

A generated sermon outline, pastoral response or leadership recommendation has no calling, conscience, covenant or capacity for repentance. It may assist a responsible minister. It cannot bear spiritual authority, promise revelation or replace accountable relationship with a person and community.

Christian trust is not gullibility. Scripture joins trustworthiness with truth, tested character, wise counsel, justice and responsibility for the vulnerable. A leader should disclose material AI assistance where it affects confidence, protect pastoral information and keep safeguarding, discipline, appointment and spiritual discernment under legitimate human and ecclesial authority.

Pentecostal leadership can affirm the Spirit's gifts without turning fluency into anointing. The Spirit forms a body whose members discern, correct and serve one another. No machine becomes a member of that accountable body.

Boards should govern the trust chain

Trust in AI is a chain connecting governing bodies, executives, professionals, vendors, data, models, users and affected people. A weak link can invalidate the confidence placed in the whole arrangement.

Before approving a consequential use, a board should require a short trust record. It should state the human purpose, lawful authority, affected people, data sources, known limitations, decision rights, reviewer capability, vendor interests, incident route, explanation and appeal, cost, renewal criteria and exit plan.

The World Economic Forum's 2026 organisational-transformation report, drawing on more than 450 executives in its industry community, identifies human accountability, operating-model redesign, talent, transparency and disciplined experimentation as enabling principles. This is an expert-community report and set of organisational cases, not causal proof of universal returns.

The board's task is not to certify that a system is trustworthy forever. It is to establish the evidence threshold, monitor changing performance, commission independent challenge where warranted and stop or redesign the use when the public promise can no longer be defended.

Use a warranted-trust test

For every important human-AI workflow, leaders should answer seven questions. Purpose: what human good should improve? Authority: who may frame, recommend, decide and review? Evidence: what demonstrates performance, fairness and limits? Capability: can the reviewer understand and override the system? Voice: can workers and affected people challenge it safely? Remedy: what happens after error or harm? Exit: can the institution continue its duty without the vendor or tool?

The answers should be specific enough to alter procurement, staffing, training and workflow. A generic ethics policy cannot substitute for the name of the officer who may reverse a decision or the time within which an appeal will be heard.

Leaders should then compare three forms of trust: trust claimed by the provider, trust reported by users and trust warranted by observed evidence. Agreement is useful. Divergence is diagnostic. High user confidence with weak evidence signals overreliance; low confidence with strong evidence signals a communication, experience or historical-trust problem that leaders should investigate.

Trustworthy leadership does not promise certainty. It makes responsibility visible under uncertainty.

Let trust follow truth

Artificial intelligence may help institutions analyse information, reduce repetitive work, expand access and improve decisions. Those possibilities deserve disciplined exploration. They do not create an entitlement to trust.

People are not barriers standing between leaders and technological return. They are colleagues, learners, applicants, citizens, customers and communities whose knowledge and rights belong inside the decision system. Their questions may be the institution's earliest warning that the strategy is drifting away from its values.

Global South institutions can lead by refusing both technological fatalism and performative caution. They can adopt useful systems while insisting on local purpose, linguistic and social inclusion, public value, professional judgement, accountable relationship and affordable remedy.

Do not begin by asking people to trust the system. Begin by making authority visible, fairness demonstrable, oversight meaningful and correction possible. Then let trust follow truth.