Who Spends the Time Your AI Saves?

An AI agent can turn a few minutes of instruction into days of work for other people. Sometimes that is exactly what a fair, accessible institution should allow. Yet a request that must be inspected or answered can make one person’s productivity depend on someone else’s time. Equal access to agents would not settle the problem. Institutions still need a rule for how much of a shared day any one principal may call upon.

Imagine a public agency opening its inbox on Monday morning. Overnight, one company’s AI agents have filed hundreds of objections to proposed rules, licensing decisions, and information requests. Every submission is lawful and worded differently. Several cite separate passages, so staff cannot dismiss the batch without looking through it.

The executive who authorized the campaign spent a few minutes choosing its goal, while the civil servants who receive it inherit days of sorting, checking, and responding. The company’s dashboard records time saved; the agency’s calendar records time consumed. Calling the whole exchange efficient would be too convenient.

The scene is hypothetical. It says nothing about how common such campaigns are today, but it does isolate a narrow problem. When AI summarizes private notes or sorts personal files, its time saving need not burden anyone else. When it negotiates, applies, appeals, complains, or petitions, the task is relational because success depends on another person or institution doing something. In those cases, the saved time may not have disappeared. It may have changed owners.

Main finding: the minute that moved

Productivity is usually measured from the user’s side of the screen. If a task took an hour and now takes ten minutes, the remaining fifty minutes appear as a gain.

That measure works reasonably well for a closed task. If an agent searches a personal archive, the user’s faster result can be close to a net saving. An open task has another endpoint: someone must read a contract proposal, check an insurance claim, decide an appeal, or investigate a customer complaint. Faster production of the request does not guarantee faster or cheaper judgment at the other end.

Research on administrative burden usually follows the costs citizens bear while learning rules, documenting eligibility, and enduring stressful procedures. It also shows that the placement of those costs can be a political choice. The agentic case reverses the direction without making officials the only people affected. When one actor fills a limited review queue, another claimant may wait longer for a benefit, license, hearing, or answer. Response capacity is shared even when requests arrive one by one.

Economists Justin Rao and David Reiley described a related structure in their study of email spam. Bulk messages were cheap for senders and costly in aggregate for recipients and mail providers. Institutional requests are different. A benefits appeal or public comment may exercise an important right, and the recipient may owe it serious attention. The time problem is harder for precisely that reason: the institution cannot filter the message when receiving it is part of its purpose.

A workplace version appears in the discussion of AI "workslop": a sender quickly produces polished-looking material, then a colleague spends time finding what is missing or wrong. Quality, though, is only part of this column’s problem. A perfectly accurate appeal can still consume review time because a valid demand may be one the recipient cannot responsibly ignore.

Method and scope: from making people wait to making them answer

Time inequality is not a blank field. Robert Goodin’s account of temporal justice treats control over discretionary time as a distributive concern, while Sarah Sharma’s work on uneven temporalities shows how one person’s speed and flexibility can depend on labor that maintains their schedule.

A still closer account appeared in 2026. Stephen Christ calls the ability to convert money, knowledge, and connections into faster institutional outcomes "speed capital". Someone with counsel, procedural fluency, or access to an expedited track can move through a queue faster than someone without those resources.

Agentic AI points power in another direction. The advantaged actor may avoid the queue or make someone else wait, but can also create more entries in that queue. An agent can submit another revision, ask another question, find another procedural ground, or reopen another negotiation while its principal does something else.

Call this a practical claim on response time. The phrase does not name a new legal right. It describes the effective ability to make another person or institution allocate attention because a rule, contract, professional norm, or risk makes nonresponse costly. Power can now lie both in making other people wait and in summoning their time automatically.

What the evidence can show: an agent is more than another assistant

Wealthy people have always hired others to act for them. Lawyers file motions, lobbyists draft comments, assistants negotiate schedules, and organizations distribute work across teams. AI did not invent delegation or procedural pressure, so there is no clean historical break here. The mechanism has still changed.

A human delegate needs wages, coordination, rest, and supervision. The delegate can refuse an instruction, raise an ethical concern, misunderstand the goal, or decide that another filing would be pointless. Lawyers and other professionals may also carry duties and liability of their own. Those frictions do not make human delegation fair, but they connect the number of interventions to additional human relationships and costs.

An AI agent can be copied, run continuously, and directed toward the same objective across many channels. It still consumes compute and money, and its work still needs oversight. What shrinks most sharply is the principal’s additional personal time for the next request. Producing the hundredth variant may take seconds, while determining whether it contains a new legal or factual issue may still take a human reviewer much longer.

This decoupling matters most where institutions promise individualized consideration. The agent operates at machine pace, but the promise is redeemed at human pace. If volume helps the principal delay enforcement, wear down a counterparty, or occupy a larger share of a public queue, saved time becomes leverage over other people’s schedules.

Giving everyone an agent does not necessarily correct the problem. Applicants can automate submissions while agencies automate replies and counterparties automate objections. This symmetry may remove routine work, which would be a real gain. It may also produce an arms race in which every participant must generate and screen more material to keep the same position. Equal access to agents can coexist with a larger social processing bill when scarce human judgment remains at the end of every automated exchange.

What it cannot show: assistance is not domination

A rule that treats AI assistance as inherently suspicious will punish many of the people who need assistance most. Amelia Arsenault and Sarah Kreps tested that assistance in a 2026 study of public commenting. In a 386-person experiment built around a real policy document, participants with ChatGPT found comment writing easier, and human and AI evaluators rated assisted comments more highly across education levels. The tool did not produce a statistically significant improvement in self-reported policy comprehension. The study also did not test automated manipulation or real mass filing.

Those limits matter, but so does the benefit. A person who struggles with legal language can use AI to turn one genuine grievance into a clear appeal, and someone with several distinct injuries may need several submissions. Collapsing everything from one principal into one item would merely replace an old access barrier with a new one.

The receiving side can automate too. A UK government trial used an AI tool to analyze more than 2,000 consultation responses and reported results close to those of human officials. That does not prove that legal significance, disputed facts, or individual hardship can be delegated safely. It does show that an agent-generated request does not always create an equal amount of new human work.

AI use alone is therefore not temporal domination, and repetition alone is not enough either. The problem appears when assistance becomes multiplicative power over a bottleneck that other people still have to staff.

Meaning: a test for temporal domination

Four conditions separate that problem from ordinary assistance.

1. One principal can produce substantially more interventions while spending little additional personal time. A clearer version of one claim does not meet this condition.

2. The recipient has a legal, procedural, contractual, or practical reason to inspect or answer. Material that can be discarded without consequence belongs to a different problem.

3. Human judgment remains necessary because safe batching or automation cannot remove the added burden without overlooking distinct rights, facts, risks, or reasons hidden inside the volume.

4. The extra work displaces others by delaying claimants, changing a bargaining position, exhausting a limited review budget, or making resistance more expensive for the recipient.

Together, the first three conditions create a temporal externality because someone else pays part of the initiator’s gain in processing time. The fourth can turn that cost into a relation of domination. The principal uses another party’s limited time as an instrument while remaining comparatively insulated from the burden.

The test sets a high bar. A single accessible appeal fails the first condition, while a platform that safely batches equivalent requests may fail the third. A mutually automated negotiation that consumes negligible human attention may fail the second and fourth. The label belongs to the structure of the interaction rather than the mere presence of AI.

Limits and evidence: count people, reasons, and harms

Institutions need not choose between unlimited machine-generated volume and a ban on assisted participation. They already distinguish document count from substantive weight.

After earlier mass-comment controversies, the U.S. Administrative Conference adopted guidance on mass, computer-generated, and falsely attributed comments. It recommended transparent handling of identical material and computer-generated submissions while explicitly refusing to treat broad participation itself as a problem. That balance is important. In a separate investigation, the New York attorney general found that nearly 18 million comments in the FCC’s 2017 net-neutrality proceeding were fake. Fraudulent identities are not equivalent to authorized AI assistance, but the case shows why the number of documents cannot stand in for the number of people represented.

A workable rule would track at least four things: the accountable principal, materially new reasons, distinct people or harms represented, and the human judgment the request truly requires. Reworded duplicates from one principal could be grouped without pretending they never arrived, while distinct claims from one person would remain distinct. A thousand people authorizing the same position would still register as public participation even if the institution answered their shared argument once.

Agent identity can support that accounting. NIST’s 2026 concept work treats identification, authorization, auditing, and non-repudiation as basic questions for software agents. An agent may need its own technical identity, but its acts should remain attributable to the human or organization whose authority it uses.

Technical attribution is only half the answer. Freedom to express a view does not automatically include unique human review for every machine-generated variation of that view. At the same time, people without agents should not lose an appeal, a job, or a public voice merely because they cannot answer at machine speed. Rate limits, batching, and priority rules should govern the demand placed on a shared process without quietly reserving that process for technically fluent users.

Return to the Monday inbox. Its first question should concern the principals who authorized the documents, the new reasons they contain, the people or harms they represent, and the parts that deserve human judgment. AI does not manufacture time. In institutions built around requests and replies, it can automate claims on response time. Justice begins with deciding which of those claims should be allowed to become someone else’s day.

Some articles on SikiT were prepared with the help of AI tools.

Sources

• Temporal Justice, Robert E. Goodin, 2009
• In the Meantime: Temporality and Cultural Politics, Sarah Sharma, 2014
• Institutional time inequality: speed capital and the social distribution of waiting, Stephen R. Christ, 2026
• Administrative Burden: Learning, Psychological, and Compliance Costs in Citizen-State Interactions, Donald Moynihan, Pamela Herd, and Hope Harvey, 2015
• The Economics of Spam, Justin M. Rao and David H. Reiley, 2012
• Why People Create AI "Workslop" and How to Stop It, Kate Niederhoffer, Alexi Robichaux, and Jeffrey T. Hancock, 2026
• Robotic rulemaking, Bridget C.E. Dooling and Mark Febrizio, 2023
• Whose voice counts? The role of large language models in public commenting, Amelia C. Arsenault and Sarah Kreps, 2026
• Administrative Conference Recommendation 2021-1, Administrative Conference of the United States, 2021
• New York Attorney General report on fake FCC comments, Office of the New York Attorney General, 2021
• New Concept Paper on Identity and Authority of Software Agents, NIST NCCoE, 2026
• Government-built Humphrey AI tool reviews responses to consultation, UK Government, 2025

Related articles

• Did I Mean It Before the Machine Said It?
• Who Is Free-Riding on AGI?

The rule remains open where grouping requests would hide distinct rights, facts, risks, or harms.