Expert authority in AI-mediated decisions

Expert authority in AI-mediated decisions is the relational capacity of professionals to have AI-influenced decisions understood, accepted, and acted on by the people affected by them. It is not reducible to having a formal role, possessing domain expertise, or clicking final approval. When AI makes a decision that an expert cannot readily overrule, the expert may still carry the work of contextualizing, explaining, and defending that decision to a client, candidate, customer, manager, or other audience.

2025-mayer-let-me-explain provides the first direct evidence for this problem. Across banking, biotechnology, and recruitment, predictive AI shifted core decisions away from loan consultants, seed sorters, and recruiters while leaving them responsible for presenting those decisions to clients. The authors describe explaining practices as an alternative pathway for experts to reconstruct authority: explanations worked better when clients recognized human expertise as distinctive and when sustained interaction let experts learn and respond to evolving client needs. Where interaction was sparse or explanations did not incorporate those needs, clients could not make sense of or endorse the outcome and expert authority weakened.

This makes explanation part of the actual design of work, not a downstream communication layer. An organization can centralize or automate a decision while retaining a human-facing role that is expected to translate opaque output, preserve a relationship, answer objections, and secure acceptance without the discretion to revise the decision itself. That configuration changes the human role from decision-maker toward relational interpreter, but whether it is responsible or sustainable depends on the truthfulness, contestability, and practical value of the explanation—not merely on whether the client accepts it.

The evidence also creates a necessary tension. In the banking case, experts sometimes masked AI output to maintain the appearance that their own expertise determined the decision. That may protect perceived authority in the short run, but it is not equivalent to transparent, accountable explanation. A workflow that leaves experts responsible for decisions they cannot understand, contest, or tailor risks making the human both a legitimacy buffer for the system and the visible bearer of client dissatisfaction.

Connections

• Relates to human-ai-collaboration because collaboration is incomplete if it specifies task execution and review but not who can explain, contextualize, and stand behind an AI-supported decision to affected people.
• Relates to responsible-ai-deployment because responsible deployment requires intelligible, contestable decision pathways rather than asking workers to legitimize opaque outputs they cannot meaningfully challenge.
• Relates to algorithmic-assessment because recruiters may remain responsible for presenting AI-supported candidate decisions even when the assessment system has already narrowed their discretion.
• Relates to work-redesign because automation can reallocate work into relational translation, client learning, objection handling, and accountability without formally recognizing or resourcing those tasks.
• Relates to human-ai-agent-interaction-design because an expert’s ability to explain, intervene in, and recover from a decision depends on whether the system exposes usable evidence, uncertainty, control points, and escalation routes.

Contradictions & open questions

• 2025-mayer-let-me-explain shows that client acceptance can reinforce expert authority, but acceptance is not sufficient evidence that a decision was transparent, fair, accurate, or appropriately contestable. The field needs studies that keep client understanding, acceptance, procedural justice, error correction, and substantive decision quality analytically separate.
• The first evidence concerns predictive-AI decisions in three client-facing professional settings. It remains open whether GenAI and agentic systems make expert authority easier to enact through adaptive explanation or harder to enact by producing fluent but unverifiable rationales.
• The organizational allocation problem is unresolved: when an expert must present a decision but cannot overrule it, which authority, training, evidence access, appeal route, and accountability should accompany that role so the expert is not merely a human legitimacy layer?