Two reports published within days of each other in late October describe the same economy from opposite ends.
The first is Anthropic’s internal study of how employment is changing at the company that builds Claude, drawing on employment records, agent session telemetry, calendar data, output metrics and 17 employee interviews. Its headline number: by August 2026, hours worked by interactive AI agents outnumbered human work hours six to one.
The second is FirstPageSage’s October agentic commerce study, built from a panel of 4,936 US online shoppers, anonymized order and referral data from 43 ecommerce, travel and insurance brands, and 18,400 shopping prompts run across the major AI platforms between July 13 and October 16. Its headline number: 61.4 percent of US online shoppers now use AI to research purchases, but only 6.8 percent have let an agent complete a checkout.
One report is about writing software. The other is about buying paper towels. Together they frame the agent economy more precisely than any forecast: execution is becoming abundant, and verification is becoming the scarce resource. Anthropic hit that wall inside its own company and responded by building monitoring infrastructure it calls “meta-work.” Commerce is hitting the same wall in retail form, and has not yet built the equivalent. That missing layer, not payments and not consent, is where the 2027 stack gets decided.
Six to one, measured, not predicted
The Anthropic study matters because it is telemetry, not projection. The research group led by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory tracked what actually happened inside a company that reorganized itself around agents, and the ratios are startling.
In March 2026, the combined hours of autonomous and interactive agents passed human work hours for the first time. By August, employee-directed interactive agents alone worked six times as many hours as their human handlers. The people are, in the study’s own words, “a thin layer atop a growing contingent of AI agents.”
The composition of the company shifted underneath that crossover:
- The share of junior-level employees in technical departments slipped from 21 to 14 percent since early 2025, and the proportion of junior technical hires fell faster, from 27 percent to 9 percent.
- Coding-focused agent sessions among non-technical employees (those with no code commits at baseline) leapt from 2 per month in August 2025 to 90 per month by May 2026. A sales representative with no engineering background built a custom pipeline dashboard in a few hours, “connecting all the tools, the MCPs.”
- At five years old, Anthropic books between two and four times as much revenue per employee as Meta and Google did at their own rapid-growth crests.
- A financial manager built an agent that reviews purchase requests. An executive assistant built one for scheduling and email triage, and concluded the company had “already moved past the point of Claude augmenting away the really entry-level admin type role.”
None of that is the interesting part. The interesting part is what the surplus did to attention.
The bottleneck flipped from doing to checking
The study’s most quoted finding is not the six-to-one ratio. It is this: even among non-technical employees, reading all the output of their Claude agents would take half their work week. One engineer described the job as facing “an army of Claudes, with every one asking for my attention.”
Anthropic’s response was to build what the researchers call meta-work: visibility tools, agent monitoring pipelines, early warning systems for jobs at risk of disruption. Whole teams rotated from producing output to supervising it. A product writing team that once crafted greetings for Claude.ai built a context graph that let agents write the copy themselves, and the humans moved into writing model evals and prompts. The study compresses that migration into one sentence that deserves a wider audience than a labor report usually gets:
“Writers who used language to drive human behavior are now using it to drive Claude behavior.”
Hold that sentence. We will come back to it, because it is also the most accurate one-line description of what is starting to happen to product reviews.
The structural lesson from Anthropic is simple and transferable: when execution capacity scales sixfold, the constraint does not disappear. It relocates. It moves from production to verification, from writing the output to deciding whether the output is any good. The companies that win the next phase will be the ones that industrialize the checking, not the doing.
The shopping mirror: 61.4 percent research, 6.8 percent checkout
Now run the same analysis on FirstPageSage’s funnel, which the firm defines around any purchase where an AI platform “builds the cart, fills the checkout, or completes the payment on the shopper’s behalf.” Adoption by stage, with change since July:
| Shopping stage | Share of US online shoppers | Change since July 2026 |
|---|---|---|
| Product research and discovery | 61.4% | +5.8 pts |
| Price and feature comparison | 48.7% | +6.3 pts |
| Review summaries | 39.2% | +4.1 pts |
| Cart building by an agent | 14.6% | +5.2 pts |
| Agent checkout with shopper approval | 6.8% | +3.9 pts |
| Scheduled or price-triggered reorders | 3.4% | +1.1 pts |
| Fully autonomous purchase, no approval | 1.3% | +0.7 pts |
The pattern is identical to Anthropic’s. Capacity is exploding where execution is mechanical. Cart building grew 55 percent in relative terms in three months and approved checkout grew 134 percent, the fastest movement in the dataset. Agent-completed orders rose from 0.27 to 0.68 percent of all US ecommerce orders between Q2 and Q3, the largest quarterly jump FirstPageSage has recorded, with 71 percent of the increase tracing to Meta Muse’s September 8 launch.
But look at where shoppers stop. The stages that require a human to trust machine-assembled evidence collapse sharply: 61.4 percent research, 39.2 percent read AI-written review summaries, 14.6 percent let an agent build a cart, 6.8 percent approve a checkout, 1.3 percent go fully autonomous. Approved checkouts outnumber fully autonomous purchases by more than five to one. Shoppers overwhelmingly keep a human checkpoint.
That checkpoint is the retail version of Anthropic’s half-week of reading agent output. The shopper’s approval click is the last moment a human eyeballs what the machine assembled. And just like at Anthropic, it is becoming the bottleneck the industry is misreading.
The industry’s October answer to the 6.8 percent problem has been infrastructure for consent and money movement. Walmart shipped Agent Profiles and an MCP gateway with per-agent spending caps. World shipped proof of human and AgentKit so a person can delegate verified personhood to an agent. Mastercard, Visa, Forter and others shipped agent identity and fraud scoring for payments. Every one of those layers authenticates the buyer, consents the transaction or moves the money. Not one of them verifies what is in the cart.
FirstPageSage’s own partnership data proves the point without intending to. Merchants inside an agent checkout partnership complete purchases at 1.4 to 3.8 times the rate of merchants outside one. The rails decide who can be bought. Nothing in the entire partnership table decides what deserves to be bought.
Review summaries are the tell
One row in the funnel deserves its own section: 39.2 percent of US online shoppers now have AI summarize reviews for them, up 4.1 points since July.
That number means product reviews have quietly crossed a threshold. The primary consumer of review text on major platforms is no longer the human who writes the buying decision. It is a summarizer that reads thousands of reviews and compresses them into a paragraph the shopper trusts without seeing the source material. The review corpus has become machine-ingested infrastructure.
The category data sharpens the point. Consumer electronics is the most AI-researched category at 68.1 percent of shoppers, yet only 5.2 percent let an agent finish the purchase, with a median agent spending limit of $240. Groceries, the low-stakes category, runs 44.8 percent research and 11.9 percent checkout at a $112 limit. Travel shoppers set agent limits at $620 but complete autonomously at just 6.3 percent. The rule the data reveals: the more the decision depends on evidence, the more shoppers delegate the reading and the less they delegate the deciding.
Which is exactly why the evidence layer is where this breaks. Shoppers in high-stakes categories are outsourcing the ingestion of reviews to models while keeping the final click for themselves. The approval step only works if the summary the human glances at is itself trustworthy. Break the reviews and you break the summary, and the shopper cannot see the break because they never saw the reviews.
The quote that should scare marketplace integrity teams
Return to Anthropic: “Writers who used language to drive human behavior are now using it to drive Claude behavior.”
That sentence was written about marketing copy inside a software company. It applies with more force to Amazon reviews, Q&A sections and listing text. Sellers who spent a decade learning to manipulate human readers, and a review ecosystem already saturated with fakes, are now facing a readership transition. When 39.2 percent of shoppers consume reviews through a summarizer, a manipulated review no longer needs to fool a human. It needs to fool a model, and models read everything, weigh aggregates and cannot hover over a profile to check whether “Verified Purchase” means anything.
The Atlantic argued in October that agents are arriving to fulfill “the promise of perfect commerce” while nobody, shoppers, retailers or platforms, is ready. The unready party that matters most is the evidence itself. Agentic commerce assumes the review layer is signal. A corpus with an estimated 30 percent manipulation rate is not signal. It is a poisoned input wearing a star rating.
This is the asymmetric war of the next phase. Manipulation scales with execution: it is cheap, automatable and now targeted at machines. Verification scales with attention: it is exactly the resource Anthropic’s study shows humans ran out of first. Any trust architecture for agentic commerce that leaves review verification to the shopper’s final glance has already lost.
Industrializing the checking
Anthropic’s answer to six-to-one agent hours was not to hire six times more readers. It was to build verification machinery: monitoring dashboards, evals, visibility tooling, an early warning system. Commerce needs the same move, applied to product evidence instead of code output.
That is precisely the layer GoBuy builds:
- Fake review filtering before summarization. GoBuy strips manipulated reviews out of the corpus and weights authentic ones up, so the evidence an agent or a shopper sees is filtered at the source, not after a model has already averaged the poison in.
- Smart Score 0 to 100, built on review quality, not quantity. A million reviews with a 30 percent fake rate produce a confident wrong answer. Smart Score scores the substance, which is the only thing that survives contact with a summarizer.
- Shortlists, not search results. GoBuy returns the top 7 products per category, because an agent acting for a human needs a verified shortlist, not four hundred rows of sponsored inventory.
- GoBuy Verified. Products holding a Smart Score above 80 across 90 days earn the badge, a durable signal agents can cite and shoppers can trust at the approval click.
- MCP at gobuy.ai/api/mcp. AI agents consult GoBuy before purchasing, which means the trust layer sits where the reading happens, upstream of the cart.
- A Chrome extension that injects the trust panel directly onto Amazon product pages, for the human who still browses the old web.
FirstPageSage’s forward forecast says platforms will start charging merchants for checkout modules in the first half of 2027, and that brands recommended organically will hold “a lasting cost advantage” over brands that pay for placement. Read that carefully. When placement becomes paid, organic recommendation becomes the asset, and organic recommendation is exactly what verified product evidence produces. The evidence layer is not a compliance feature. In a monetized agent funnel, it is the margin.
The scarce resource sets the agenda
Anthropic gave the agent economy its first honest internal audit, and the audit says the constraint is no longer what agents can do. It is what humans can check. FirstPageSage gave commerce its most granular funnel yet, and the funnel says the same thing in shopper behavior: research delegation is mainstream, evidence delegation is growing, and the human approval click is the last verification checkpoint standing.
October’s infrastructure wave built identity, consent and payments. November’s has to build evidence verification, or the six-to-one economy will optimize against a review layer nobody is guarding.
Before your next agent-assisted purchase, see what filtered evidence looks like: gobuy.ai. Building a shopping agent? Connect it to the trust layer at gobuy.ai/agent-docs.