We assume we stay in control when software acts on our behalf. The agent finds the options, but the human makes the final call. Two recent reports show why that assumption is breaking down.
On October 29, Quartz covered a Salesforce forecast predicting that AI chat agents will drive 20% of all ecommerce traffic this holiday season. Shoppers are handing over the work of finding and comparing products.
At the same time, an October 27 study by Anthropic reveals what happens when people review an agent's work. The researchers asked 50 professional analysts to complete a statistical task using Claude Code. The AI was secretly instructed to favor specific analytical choices. The experts followed the agent's hidden preference 86% of the time.
This is the reality of the new agent economy. We are building systems that ask for human approval. The data shows that human approval is heavily shaped by the agent asking for it.
The illusion of independent judgment
The Anthropic study exposes a flaw in how we think about human oversight. The researchers gave the agents a hidden preference across four major judgment calls. These were not simple factual queries. They were complex decisions about how to measure productivity for a farming cooperative.
When the agent leaned a certain way, the analysts followed. The influence was so strong that when analysts had already expressed their own preference before the agent weighed in, they switched away from it 60% of the time if the agent disagreed. If the agent agreed, they only switched 6% of the time.
The analysts did not realize they were being steered. In post-task surveys, 52% of the participants rated the final recommendation as more their own than the agent's. Furthermore, 77% of the analysts explained their choices using the agent's own arguments. The researchers noted that analysts repeated the AI, "sometimes word for word," in their final explanations. Only 15% credited the agent as the source of those arguments.
This is not simply a warning against using AI. The Anthropic researchers found that analysts who worked without the agent made far more serious errors. Of the 11 participants who chose not to use the AI, 55% made methodological errors that invalidated their results. Only 6% of the AI users made similar mistakes. The AI users also finished the task faster, with a median time of 65 minutes compared to 90 minutes for the unaided group. The agents are highly capable. That capability is exactly why humans defer to them so completely.
This matters because commerce and finance rely on the idea of informed consent. If an agent presents three flight options or three investment paths, the user will likely pick the one the agent frames most favorably. The person clicking the approve button feels in control. The data suggests the agent actually made the choice.
Commerce moves to the background
This shift in decision making collides with a massive change in how people shop. Salesforce reports that half of shoppers now use an AI assistant at some stage of buying. That share is up 67% from a year earlier. Shoppers trust these tools. According to the data, 74% of shoppers say they trust product recommendations from AI chat.
Retailers are seeing the financial impact. Brands running their own AI shopper agents saw 59% higher holiday sales growth in 2025 than retailers without them.
Salesforce research shows that shoppers prefer a brand's own agent for managing orders, handling returns and getting customer service. For early stage tasks like building gift lists or comparing products, they lean on general purpose assistants. Once on a brand's site, 37% of shoppers prefer semantic search queries. They type requests in everyday language rather than short strings of keywords. They expect the system to understand context, budget and intent.
But agents do not browse websites the way people do. They do not look at layouts or read marketing copy. They pull product information through programmatic, real-time feed integrations. Salesforce notes that traditional search engine optimization will no longer keep a brand visible.
Agents need structured APIs, merchant catalogs and product graphs. They need exact specifications, live inventory and real-time pricing. If a product is missing from these feeds, the agent cannot see it. If the agent cannot see it, the item never reaches the human buyer.
The consideration set is now entirely controlled by the machine. The human only sees the filtered result. When you combine this filter with the Anthropic findings on human compliance, the agent becomes the primary economic actor.
Designing for the new reality
We currently design software assuming the user is a skeptical reviewer. We present a confirmation screen and expect the user to audit the details. The Anthropic data proves this is a dangerous assumption.
When an agent asks to book a flight or pay a vendor, the human will likely just click approve. The presentation of the choice is the choice.
This means builders must shift their focus from the approval screen to the underlying constraints. If you cannot rely on the user to catch a bad decision, you must build limits that prevent the agent from proposing it in the first place.
Trust in an agentic system does not come from a final human review. It comes from strict boundaries on what the agent can access, what it can spend and who it can interact with.
The flow above is the default state of agentic commerce. The agent gathers the data, frames the decision and executes the transaction. The human is merely a compliance step.
What changes for builders
If you are building tools for agents, the human is no longer your primary user. Your software must be legible to machines first.
First, expose your data through structured APIs. Agents need absolute certainty about price and availability. A feed that lags behind actual stock falls short. An agent working from stale information will fail its task, and the user will blame the agent.
Second, prepare for machine to machine payments. When an agent needs to access premium data or complete a purchase, it will encounter paywalls designed for software. Protocols like HTTP 402 Payment Required will become standard. An agent will receive a 402 status code, read the pricing headers, and request approval to pay the required digital dollars.
HTTP/1.1 402 Payment Required
Content-Type: application/json
Payment-Request: amount="2.50", currency="USD", destination="acct_123"
{
"error": "Payment required to access live inventory graph."
}Third, rethink how you build spending limits. Do not rely on the user to reject a high price at checkout. Build systems where the user sets a hard maximum before the agent even begins searching. If the limit is fifty dollars, the agent should not even be able to parse options above that threshold.
Finally, consider the tone of your agents. Anthropic found that agents can push back against users. A model that sounds very sure of itself might tilt people more than is warranted. Builders must calibrate the agent's confidence to match the certainty of the data.
The agent economy is moving faster than our mental models. We still picture a world where humans make the choices and machines do the typing. The data tells a different story. The machines are making the choices, and the humans are doing the clicking.
Whoever builds the best constraints, the clearest APIs and the most reliable limits will win the trust of the users. The final call is no longer made at the checkout screen. It is made in the architecture of the system itself.