a16z Charts of the Week: AI Can Help You Pick Products—Why Is It Still Hard for It to Place Orders?

AI Can Help You Pick Products—Who Is Ready to Let It Place Orders?

Asking AI to find a suitable pair of running shoes is an entirely different level of authorization from letting it charge your credit card. "Agentic commerce," as explored in the latest a16z weekly chart report, involves AI assistants taking the step from recommending products to autonomously completing orders on behalf of consumers. This evolution confronts resistance from three distinct sides: platforms seeking to protect customer relationships and advertising revenue, merchants wary of returns and dispute liabilities from mispurchases, and consumers who remain hesitant to hand over final buying decisions.

According to traffic data from cybersecurity provider Cloudflare, the proportion of "403 Forbidden" status codes encountered by automated crawlers and AI agents accessing web services has expanded visibly since mid-year. Because this traffic bundle also includes training crawls and search scraping, it is impossible to deduce how many shopping agents were blocked, nor can the entire increase be attributed to retailers. Cloudflare itself allows websites to manage search, training, and user-representing agents separately. a16z notes that platforms such as Amazon are tightening access for shopping proxies, showing that platform openness matters alongside technical capability.

The primary motivation for e-commerce platforms resisting automated proxy buying is preventing "disintermediation." In traditional retail models, consumers search on-site, browse recommendation feeds, and encounter ads—interactions that deliver advertising revenue and direct customer relationships to the platform. If AI agents step in as intermediaries and users merely confirm a final price in an external chat window, opportunities to serve recommendations and display ads diminish, and customer relationships are far more likely to be captured by third-party assistants.

A more direct obstacle stems from post-transaction fulfillment and liability. In real-world e-commerce operations, returns and payment disputes already strain merchants.

According to data compiled by organizations like the National Retail Federation (NRF), merchandise returns have grown increasingly acute in recent years. Total returned merchandise across all US retail has climbed above $800 billion, with online return rates approaching 20%. While returned merchandise is not a 100% net loss—merchants can recover part of the cost through reinspection, refurbishment, or discounted resale—the operational friction of two-way logistics, warehouse restocking, and packaging damage remains a severe financial drag.

When automated proxy shopping enters the picture, this friction can escalate further. If an AI agent orders the wrong specification due to misunderstanding, or interprets "find the lowest unit price" as purchasing an entire shipping container of socks at wholesale, resolving who authorized the quantity and who absorbs the mispurchase becomes a contentious dispute. This is a hypothetical scenario to illustrate the risk, not an actual documented incident.

Chargeback disputes present an even harsher reality. Merchant fraud surveys indicate that over 80% of surveyed retailers have encountered "dispute abuse"—instances where consumers make legitimate purchases but subsequently exploit chargeback dispute mechanisms. Standard refund requests or chargebacks are not inherently fraudulent. However, if an agent-initiated transaction is contested as "purchased by an assistant without authorization," the merchant must prove consent. If the card issuer ultimately sides with the buyer, the merchant risks having funds clawed back.

Beyond merchant and platform hesitancy, consumers themselves remain reluctant to surrender full purchasing control.

Surveys demonstrate that consumers are by no means hostile to AI in retail. 44% of respondents agree that AI helps identify better deals, 41% are willing to consider previously unknown items recommended by AI, and 37% completed a purchase over the past year based on AI recommendations. If AI demonstrates that a given product is the optimal choice, more than 30% of consumers even express a willingness to pay a slight premium.

However, once the task shifts from "help me find" to "buy it for me," consumer attitudes diverge sharply.

When presented with the option of allowing AI to select and directly purchase items, roughly 42% of respondents explicitly state they do not trust AI to make final buying decisions on their behalf. Even for daily essentials, apparel, or electronics accessories, the willingness to delegate purchases hovers around one-quarter. Being receptive to recommendations is fundamentally different from handing over payment credentials; a significant trust gap remains between the two.

Clickstream data also indicates that AI currently serves primarily as a comparison tool during the shopping process.

According to Similarweb's analysis of US desktop clickstreams, the share of shopping journeys visiting AI prior to reaching a transaction website climbed from 4.5% to 11.4% over a two-year span. This highlights that consulting AI before buying has become noticeably more common in this US desktop sample. The data runs through June and cannot be directly generalized as representing current shopping habits across all devices.

Among shopping sessions that used AI, 76.1% engaged it during the middle of the journey. This aligns with researching specifications and comparing models, though touchpoint position alone does not prove the specific questions asked. By contrast, AI served as the initial touchpoint in only 23.2% of sessions and as the final touchpoint before entering a transaction in just 18.4%. In comparison, search appeared at the first, middle, and final touchpoints in 50.0%, 59.5%, and 56.2% of its respective sessions, showing that search remains prevalent at both the start and right before checkout. However, because denominators represent sessions using each respective tool, subtracting the two shares cannot be interpreted as a replacement rate; furthermore, multiple uses per session mean touchpoints do not sum to 100%.

This creates an intricate challenge for commercial monetization: ad attribution. If a consumer discovers a product on social media, turns to AI to compare specifications, and ultimately switches to a mobile app or direct brand website to complete payment, which channel receives credit? This cross-platform, multi-stage drop-off makes it exceptionally difficult for merchants to measure direct returns on AI channel ad spend.

Maturing commercial infrastructure and algorithmic evolution may eventually resolve payment authentication, return protocols, and attribution standards. For now, however, the AI captured in these datasets serves primarily as an informational shopping advisor, still far from universally securing autonomous buying power.

TikTok Shop: A Small Minority of Sellers Capture Most Sales

Shopping on TikTok Shop often begins with a video: demonstrating results, explaining usage, or featuring a creator's recommendation to spark initial purchase intent. Compared to an AI agent comparing technical specifications, this style of social commerce serves as a reminder that shopping encompasses discovery, trust, and entertainment. Its sales distribution is also extraordinarily concentrated.

According to Marketplace Pulse's tracking of over 150,000 merchants and millions of products on TikTok Shop, a store's content volume correlates strongly with sales performance. Among the 71% of tracked sellers with available video and follower data, merchants who posted more than 1,000 self-operated videos accounted for just 8.4% of all sellers, yet this high-output cohort captured 60.3% of sales in that grouped sample. In contrast, sellers posting between 0 and 10 videos represented over 20% of accounts, but their sales contribution was negligible.

Video counts include only assets published directly by "seller-operated accounts," excluding external affiliate creator content, though this does not mean sales exclude affiliate-driven revenue. While high video publishing frequency strongly correlates with sales scale, it does not mean any seller can mechanically churn out 1,000 videos and automatically generate equivalent revenue growth.

Ranking sample sellers by cumulative sales highlights an even steeper tier of head concentration.

Ranked by estimated cumulative sales, the top 0.1% of sellers (around one hundred stores) commanded nearly a quarter (24.7%) of all estimated cumulative sales in the analyzed sample. Expanding the view to the top 1%, the sales share climbs to 58.9%; the top 10% captured 93.3% of estimated cumulative sales in the sample. Conversely, the vast long tail making up the bottom 50% accounted for a meager 0.1% of total sales. Dollar shares are estimated from unit volume multiplied by average list price, with list prices capped at $300 and excluding 16% of sellers lacking price data. Ranked instead by unit volume, the top 1% accounts for 49% of units sold.

Here exist two easily conflated "~60%" figures: one is the aforementioned 8.4% cohort of sellers who posted over 1,000 self-operated videos, capturing 60.3% of sales in that sample; the other is the top 1% of sellers ranked by estimated cumulative sales, who generated 58.9% of sales. These reflect two completely different groupings and cannot be used interchangeably.

Extreme concentration does not imply permanent dominance. Marketplace Pulse found that monthly active leaderboards fluctuate considerably compared to cumulative rankings: going viral once is very different from sustaining top-tier sales. At the same time, if more purchasing decisions are eventually delegated to AI agents in the future, the edge held by top creators and sellers who rely on high-volume video storytelling to persuade human buyers may not operate in the same way. This remains an open question. Compelling content inspires desire, while agents assist in evaluation; both functions can coexist within a single shopping journey, meaning agentic shopping and social commerce will not necessarily displace one another.

As AI Buildouts Grow More Costly, Who Cuts Demand First?

A longstanding proverb in commodity markets observes: "The cure to high prices is high prices." When the cost of a good rises sufficiently, high expenses naturally dampen downstream demand; once demand softens, upward price pressures often ease. This feedback loop takes time and is shaped by supply dynamics, inventory cycles, and other variables.

Two sets of data within AI infrastructure prompted a16z to reflect on this supply-and-demand rule: component pricing versus computer shipments, and financing costs versus new debt issuance.

AI buildouts have surged demand for chips and electronic components, exerting upward price pressures across the supply chain. Data from the US Bureau of Labor Statistics shows that the year-over-year increase in the Producer Price Index (PPI) for semiconductors and other electronic components spiked sharply from historically stable levels, recently exceeding 25%.

Cloud hyperscalers rushing to secure core compute silicon have not been deterred by short-term high prices; on the contrary, they continue to revise capital expenditure forecasts upward. Part of these expanded budgets may simply absorb higher component costs: higher capex does not automatically mean purchasing a proportionally greater volume of equipment. Surging component costs may also spill over into the broader consumer electronics market: while cloud giants continue aggressive procurement, individual consumers and enterprise buyers may delay PC upgrade cycles.

According to quarterly tracking by IDC, global PC shipments fell noticeably into negative growth following an earlier mild recovery, with the two most recent quarters shown in the chart dipping to roughly -4% and -20% year-over-year. If component price increases pass through to finished device retail prices, individuals and businesses may postpone computer upgrades. That said, shipment declines may also be influenced by advance inventory stockpiling or supply constraints; these charts do not isolate the specific role played by pricing alone.

A similar supply-and-demand feedback loop is emerging not only in physical hardware, but also across financial debt markets. As data centers and energy infrastructure projects swell in scale, corporate bonds and debt instruments are increasingly utilized to fund these buildouts.

Borrowing costs have also shifted. The risk premium (credit spread) of 15+ year tech sector bonds relative to US Treasuries widened progressively against standard high-grade corporate debt as the wave of bond issuance got underway, climbing to 1.2 to 1.3 percentage points. Because the two indices carry different credit ratings, underlying credit risk and market expectations also influence the gap, meaning it cannot be attributed solely to rising AI debt issuance.

More expensive financing can prompt borrowers to reconsider their timing, and the subsequent pullback in debt issuance aligns with this interpretation.

AI-related debt issuance across global credit markets briefly surpassed a monthly peak of over $100 billion around mid-year before declining, with September issuance roughly half that of August. Financing costs may represent one contributing factor, though issuance schedules and project timelines can also shift monthly numbers.

The hypothesis raised by a16z is that high prices may not immediately deter the most aggressive AI infrastructure builders, but could first prompt everyday consumers to delay PC upgrades and lead corporate borrowers to moderate financing. Whether weakening demand will eventually cool prices further remains unanswered by the data.

Young Americans Earn More—Why Is Their Confidence Lower?

In the popular imagination, younger cohorts like Generation Z are frequently depicted as a struggling generation uniquely battered by inflation and employment headwinds. Yet microeconomic financial portraits compiled by banking institutions diverge substantially from this prevailing narrative.

Economic analysis from PNC Bank reveals that when comparing Gen Z in 2025 (at ages 19 to 27) with Millennials at the exact same age in 2015, inflation-adjusted real earnings did not decline—they rose by 20% to 40%. Across lower income brackets, median earners, and top earners alike, the younger generation enjoys stronger earnings than Millennials did when emerging from the shadow of the Great Financial Crisis. Furthermore, across PNC's same-age benchmarks, Gen Z's employment and early balance sheets look healthier: higher employment rates, lower unemployment, higher net worth, and a substantially higher rate of retirement account ownership.

However, higher earnings have not translated into a comfortable day-to-day cash cushion.

Internal banking transaction data reveals that Gen Z spends 66% of monthly income on living expenses and credit card payments, higher than the 60% recorded by same-age Millennials. Meanwhile, liquid reserves in deposit accounts would sustain only 25 days of typical spending if income were interrupted, down from 32 days for Millennials at the same age.

Shifts in budget allocations and spending baskets show a similarly divided pattern.

In broad consumer expenditure surveys, households under 25 saw total budget shares increase for retirement savings (+2.3 percentage points), housing (+1.5 percentage points), and transportation (+1.4 percentage points), while education shares fell (-2.3 percentage points) and utility shares narrowed (-1.2 percentage points). In card spending shares on the right, gambling (+3.0 pp) and electronics/appliances (+2.4 pp) gained the most weight, while grocery food/beverage retail (-3.8 pp) and apparel (-2.0 pp) saw the sharpest declines. These figures track changes in expenditure budget shares, not dollar spending growth rates. The left chart is based on BLS total household budgets while the right reflects PNC card transactions; their different denominators mean they cannot be combined.

Cash runway and retirement savings measure two distinct things: how long current balances can absorb an immediate shock, versus what has been set aside for retirement. These metrics can easily move in opposite directions. Stronger earnings and early asset accumulation have not made young adults more optimistic.

In the University of Michigan's multi-decade consumer sentiment tracking, confidence among 18-to-34-year-olds typically ran higher than the broader population, enjoying a longstanding "youth confidence premium" of 5 to 15 points. Since 2017, however, this historic premium has almost entirely evaporated, recently dipping below zero multiple times—meaning sentiment in this age cohort was actually lower than the national average. This age bracket does not align perfectly with Gen Z alone. a16z also notes that consumer sentiment surveys inherently carry significant noise, and shifts in response rates can influence readings. Nevertheless, more telling than any single-month reading is the multi-year secular trend of youth sentiment relative to the national average.

Why does the younger generation feel more pessimistic despite taking home substantially higher real incomes?

The a16z weekly report suggests that because income and asset accumulation have improved, financial factors alone struggle to explain this gloom. Yet PNC's original report highlights a pivotal exception: the threshold for purchasing a home has soared.

According to PNC's core benchmark model, calculating monthly mortgage payments on a median-priced US home at prevailing interest rates reveals that servicing a mortgage today would consume 65% of the median income of a 25-to-26-year-old; a decade ago, the same model showed same-age Millennials bearing a theoretical mortgage burden of just 39%.

This measures the modeled burden of purchasing a median-priced home, not the actual mortgage payments currently made by young households. Even if wages outpace general consumer prices, as long as home prices and borrowing costs rise faster, the long-term milestone of homeownership moves further out of reach.

PNC points to housing affordability hurdles and persistent inflation as key drivers of deteriorating sentiment. This offers a data-grounded explanation for why earning more does not automatically spark optimism, serving as a reminder that assessing quality of life requires looking simultaneously at income, accessible liquid cash, and the cost of long-term milestones.

Shopping involves platform rules and authorization; content relies on attention and trust; infrastructure faces component and financing costs; and income must grapple with housing burdens. Technology can execute certain tasks faster, but these charts pose the deeper questions: who is willing to adopt it, who bears the costs, and how the ultimate gains are distributed to individuals.