--- title: Who decides what counts as a better deal? subtitle: AI shopping agents reduce the work of choosing. Someone still has to set the terms. kind: Essay date: 2026-10-17 author: Ron Bronson --- In [The Atlantic this week](https://www.theatlantic.com/technology/2026/10/ai-agent-chaos-e-commerce/689010/), Matteo Wong and Charlie Warzel imagine a million shopping agents buying the same paper towels because one brand becomes three cents cheaper. The example is funny until you think about what the agents were asked to optimize. A price comparison is a fairly narrow calculation. Deciding whether the savings justify a switch involves the person who has to live with it. The article draws on Rory Van Loo's [2019 paper, *Digital Market Perfection*](https://repository.law.umich.edu/mlr/vol117/iss5/2/). He called the prospect of millions of automated customers moving together *hyperswitching*. More efficient shopping could create real gains for consumers and, in extreme cases, volatility that individual shoppers never intended. Van Loo described a possibility and stopped short of predicting that every market would behave like a stock exchange. I'm interested in the decision that precedes the transaction. Who decided that the cheapest paper towels were the best ones to buy? ## An agent has to know what better means to its customer In February 2025, I wrote [*Service Design for AI: The Disintermediation Problem*](https://blog.ronbronson.com/service-design-for-ai-the-disintermediation-problem). One example was an agent purchasing something based on what it thinks it knows about its customer. I was concerned with what happens as people lose direct contact with services and the ability to express needs in their own terms. An AI intermediary interprets intent, and that interpretation starts to govern the interaction. A household might care about a familiar brand, reliable delivery, or avoiding another set of returns. Someone else might welcome every small saving. Neither preference is inherently irrational. The agent has to know which criteria apply before "better" can mean anything useful. People also change their minds. A standing preference to buy the cheapest item is different from permission to sign a new service contract, change a bank account, or cancel an existing relationship. ## Choosing takes work, even if decision fatigue is overstated Decision fatigue is an appealing explanation for the rise of shopping agents. People have limited time. Comparing insurance plans, checking bills, and contesting fees can consume an entire afternoon. It's easy to see why people would hand that work to software. The psychology is less settled than the familiar story that every choice drains a fixed supply of willpower. A [2010 meta-analysis of choice overload](https://doi.org/10.1086/651235), covering 50 experiments, found an average effect close to zero, with large differences between studies. A [2015 review](https://doi.org/10.1016/j.jcps.2014.08.002) identified conditions that matter, including how complicated the options are and how uncertain people are about their preferences. In a [23-lab replication](https://pubmed.ncbi.nlm.nih.gov/27474142/) of a widely used ego-depletion test, the estimated effect was also very small, with a confidence interval crossing zero. There are good reasons to reduce the work of choosing without assuming that people become incapable of judgment after making enough decisions. Researchers Evan Risko and Sam Gilbert call the use of external aids to reduce mental effort [*cognitive offloading*](https://pubmed.ncbi.nlm.nih.gov/27542527/). Reminders and calculators already do this. An agent extends the idea by taking actions in outside systems. That changes the design problem. If an assistant brings every minor choice back for approval, the work reappears as a stream of notifications. If it acts on every inferred preference, the consumer may discover the consequences only afterward. Someone has to specify the routine work the agent can do, the circumstances that require approval, and the evidence that will count as a result. ## Authority belongs to each action, and outcomes need their own evidence I raised a related question in [*Who Holds the Pen?*](https://blog.ronbronson.com/who-holds-the-pen-why-ai-agents-need-a-manager-not-just-a-manual), about AI systems operating inside public agencies. A rule tells a worker what is generally allowed. Discretion enters when the circumstances do not fit neatly within that rule. When software performs the next step, authority has to attach to the specific action and its conditions. My earlier writing on service delivery comes from the same concern. In [*Stuff Should Work*](https://blog.ronbronson.com/stuff-should-work-on-doing-civic-tech%E2%80%A6), I described the gap between policy as written and service as experienced. An organization can meet a processing target while a person still waits for the thing they needed. The status in a system and the outcome in the world are different facts. The [Tallmere 311 demonstration](/writing/tallmere-311/) Vloora built makes that distinction concrete. A resident's assistant checks for existing reports and asks for approval before filing a new one. In one test, the fictional city marks a fallen-branch report closed while the branch is still on the sidewalk. The resident disputes the closure. Both statements stay in the record, attributed to the people or services that made them. A shopping agent creates a similar problem when it reports that a subscription was canceled or a cheaper plan was secured. Was the request submitted? Did the company accept it? Were fees added later? Did the service change? An assistant's description of success cannot answer every one of those questions. The friction a company imposes differs from the discretion a customer keeps. Researchers Donald Moynihan, Pamela Herd, and Hope Harvey [describe administrative burden](https://doi.org/10.1093/jopart/muu009) as learning, compliance, and psychological costs that people experience when dealing with public programs. Those costs can deter access. I have also written about how [making an administrative process more efficient](https://blog.ronbronson.com/better-technology-for-cruelty-or-administrative-burden-as-a-service) can reproduce the harm of the rules it implements. Removing needless paperwork serves people, and requiring consent before a consequential action protects their discretion. ## Who does the intermediary work for? Van Loo's [2026 paper, *Consumer Agents*](https://wustllawreview.org/wp-content/uploads/2026/02/12.-Van-Loo_Published-2.24.26-1.pdf), examines how consumer law has concentrated on the buyer and seller while doing comparatively little to protect independent third parties that could help buyers. Businesses can limit data access or obstruct agents trying to compare and switch services. The article in *The Atlantic* raises a related concern about commercial partnerships that influence which choices assistants present. A shopping agent may have permission to act and still work from a distorted set of options. That raises questions about disclosure and access that a consumer-facing approval screen alone cannot settle. One person's permission settings also cannot prevent market-wide hyperswitching. Those effects belong to competition, consumer protection, and financial regulation as well as product design. At the individual level, the customer should be able to decide what an assistant may inspect, when it may act, and what needs approval. The record of those actions should survive a change of AI provider. In my short note [*On Anti-Delegation*](https://blog.ronbronson.com/on-anti-delegation), I worried about placing private knowledge into systems we do not own. Control over an agent's permissions and history belongs in the same discussion. Vloora's work on [consumer agent experience](/writing/towards-a-consumer-agent-experience/) addresses this from the person's side. Stenodisc is a place to manage assistants and keep records of their work. [Reverse underwriting](/writing/reverse-underwriting/) is the longer-term idea that AI products build a performance history by task, so people can judge how much to trust them. Credit in that model means earned trust. A person's permission to delegate remains a separate decision. If I ask an agent to compare internet plans, I may be glad for it to search all afternoon. I can decide in advance whether it may contact a provider, and require a review before it signs a new contract. Afterward, I need to know what the provider confirmed and whether the change took effect. Those are the terms of a service that works for its customer, even when the customer did not do all the work. ## Sources - Matteo Wong and Charlie Warzel, [Imagine Every AI Agent on Earth Buying the Same Cheap Paper Towels](https://www.theatlantic.com/technology/2026/10/ai-agent-chaos-e-commerce/689010/), The Atlantic, October 16, 2026 - Rory Van Loo, [Digital Market Perfection](https://repository.law.umich.edu/mlr/vol117/iss5/2/), Michigan Law Review, 2019 - Rory Van Loo, [Consumer Agents](https://wustllawreview.org/wp-content/uploads/2026/02/12.-Van-Loo_Published-2.24.26-1.pdf), Washington University Law Review, 2026 - Benjamin Scheibehenne, Rainer Greifeneder, and Peter M. Todd, [Can There Ever Be Too Many Options?](https://doi.org/10.1086/651235), Journal of Consumer Research, 2010 - Alexander Chernev, Ulf Böckenholt, and Joseph Goodman, [Choice Overload: A Conceptual Review and Meta-Analysis](https://doi.org/10.1016/j.jcps.2014.08.002), Journal of Consumer Psychology, 2015 - Martin Hagger and colleagues, [A Multilab Preregistered Replication of the Ego-Depletion Effect](https://pubmed.ncbi.nlm.nih.gov/27474142/), Perspectives on Psychological Science, 2016 - Evan Risko and Sam Gilbert, [Cognitive Offloading](https://pubmed.ncbi.nlm.nih.gov/27542527/), Trends in Cognitive Sciences, 2016 - Donald Moynihan, Pamela Herd, and Hope Harvey, [Administrative Burden: Learning, Psychological, and Compliance Costs in Citizen-State Interactions](https://doi.org/10.1093/jopart/muu009), Journal of Public Administration Research and Theory, 2015