When the Algorithm Delegates to You

I was reading Stelmaszak et al. (2025) on algorithm-to-human delegation at Uber when something clicked that I had been circling around for weeks. The delegation literature, including the framework I find most compelling, Baird and Maruping (2021), mostly assumes the human initiates the transfer. The human decides what to hand off, to whom, and under what conditions. The artifact receives. But what happens when the algorithm hands the task back? Not as a failure mode. Not as an exception. As a designed feature of the system. The answer, I think, is that almost every concept we built for human-to-AI delegation starts to strain.

Baird and Maruping (2021) gave the IS field its most explicit delegation framework for agentic systems. They replaced use with delegation as the central construct and identified three mechanisms: appraisal, distribution, and coordination. I wrote about why this matters in my post on why use is the wrong construct. The framework also identifies four archetypes, from reflexive to prescriptive, each with increasing decision-making latitude. What matters here is that all four assume the human is the delegator. The human appraises. The human distributes. The human coordinates. Even the prescriptive archetype, where the system acts autonomously, still assumes the human initiated the transfer. The artifact does not delegate to you. You delegate to it.

Stelmaszak et al. (2025) studied Uber drivers. The algorithm assigns rides, sets surge pricing, manages ratings, and determines which drivers see which opportunities. The algorithm is the delegator. It distributes tasks. It coordinates interdependencies across the fleet. It appraises performance through rating and acceptance metrics. The driver receives the delegation, accepts or declines the ride, and lives with the consequences. Not a failure. Working as designed. The algorithm delegates to you, and your job is to perform.

Once you see this reversal, you start noticing it everywhere. Tesla Autopilot hands control back to the driver when the situation exceeds its capability. That handoff is delegation in reverse. The algorithm has decided that this task, this moment, this decision, is better handled by the human. The human did not ask for it. The human may not be ready for it. AI hiring systems escalate certain candidate files to human recruiters for final review. The algorithm flagged the file as ambiguous or high-stakes and delegated the judgment back. Fraud detection systems route flagged transactions to human analysts. The system has determined that it cannot resolve this case on its own. It is delegating upward, to a more capable agent, exactly the way a junior employee might escalate a decision to a manager.

Murray et al. (2021) provide the vocabulary I think we need here. They categorize four forms of conjoined agency between humans and technology: assisting, arresting, augmenting, and automating. Assisting technologies are tools wielded by humans. Arresting technologies constrain humans by autonomously executing actions when predefined conditions are met, like smart contracts. Augmenting technologies develop protocols and recommend actions, leaving the human to select. Automating technologies both develop protocols and select actions independently. Algorithm-to-human delegation lives in the gap between augmenting and automating. The algorithm was running the show, then it hit a boundary condition and shifted the human from a monitoring role to a decision-making role. The locus of action selection moves, not because the human chose to take it back, but because the algorithm chose to give it away.

This is where Vanneste and Puranam (2025) become relevant in a way I had not anticipated. Their theory is about how perceived agency of AI affects human trust. They identify three causal pathways. Greater perceived agency can enhance perceived ability (agentic things seem more capable), shift the locus of trustworthiness (we judge the AI itself rather than its designer), and amplify betrayal aversion (being let down by something agentic hurts more than being let down by a tool). When the algorithm delegates to you, all three pathways shift. The algorithm has demonstrated that it is agentic enough to make a delegation decision. It is not a passive system failing forward. It is an active agent choosing to transfer responsibility. This increases its perceived ability in some ways (it knows its own limits) but also increases betrayal aversion (it just told you it cannot handle this, so now you have to). And the locus of trustworthiness becomes complicated because you are now judging both the algorithm that delegated the task to you and the algorithm that will handle everything else while you deal with this one case.

Feng and Chandra (2026) describe what they call delegation inversion, where AI agents become de facto principals and delegate tasks back to human agents who become subordinate to the machine's logic. They also discuss AIcratism, where public employees defer inappropriately to algorithmic outputs and become procedural rubber-stampers. The human is no longer the principal. Feng and Chandra identify three delegation interfaces where this reversal creates problems: trust-based, hierarchical, and contractual. At each interface, information asymmetry intensifies, goal misalignment risks escalate, and accountability diffuses across actors who can all point at the algorithm.

I think the organizational design implications are what current theory has not caught up with. Baird and Maruping's framework was built for a world where the human decides. Appraisal, distribution, and coordination are human activities. When the algorithm delegates to you, who appraises? The algorithm may have already decided that you are the right person based on its own metrics, metrics you may not understand or agree with. Who distributes? The algorithm. Who coordinates? The algorithm, while you handle the piece it handed off. You are an agent who did not choose your principal. You are delegated to by a system whose criteria you cannot fully see.

The accountability question sharpens this. In traditional delegation, the delegator bears responsibility for choosing the right agent. When I delegate a task to a colleague, I am accountable for whether that person was the right choice. When an algorithm delegates a driving decision to me at 70 miles per hour, who is accountable for whether I was the right choice? When a hiring algorithm escalates a borderline candidate file to a human recruiter, who owns the outcome if the recruiter makes the wrong call? The human is performing under delegation from a non-human principal, and the accountability structures we have were not designed for that relationship.

Vanneste and Puranam (2025) give one more insight. They argue that making an AI appear more agentic can increase or decrease trust depending on the relative trustworthiness of the AI versus its designer. Trust is not the same as delegation, and this distinction matters here. The algorithm that delegates to you has just demonstrated two things simultaneously: it is capable enough to recognize a boundary condition, and it is not capable enough to handle what lies on the other side. If I am the human receiving that delegation, I have to ask: is this system trustworthy because it knows its limits, or untrustworthy because it just admitted it cannot handle a situation it put me in? The trust calculus is different when delegation flows uphill.

The theoretical gap I keep coming back to is this. Baird and Maruping (2021) built delegation theory for human-to-artifact relationships. Stelmaszak et al. (2025) showed that artifact-to-human delegation is real and consequential. Murray et al. (2021) gave us conjoined agency, which captures the spectrum of human-technology interaction but does not explicitly theorize delegation direction. Vanneste and Puranam (2025) showed that trust dynamics shift when agency shifts. Feng and Chandra (2026) showed that delegation inversion creates accountability vacuums. What we do not have yet is a theory of bidirectional delegation that accounts for who initiated the transfer, who holds the principal role at any given moment, and what happens to accountability when that role shifts mid-task. I think that theory is coming. I think it starts from the moment the algorithm hands you something and says, your turn.