# Online Retail Killed the Salesperson. AI Is Bringing Back the Ghost > Source: > Published: 2026-08-04 19:20:36+00:00 17 min read ## TL;DR — tap to expand the short version - Ecommerce spent two decades on discovery and never built the layer that tells a shopper whether an item will work for her, what this piece calls the **commerce confidence gap**. - It shows up in the numbers. NRF projects **$849.9 billion** in U.S. returns for 2025, 15.8 percent of sales, while Appriss counts**$706 billion**. Fraud accounts for a minority share of both. - Bracketing, buying an item in multiple sizes and returning most, is a customer doing the fitting room’s job herself. - The tools built to close it, AI assistants, live video, personalization, human escalation, pay for themselves only where uncertainty and price are both high; on commodity catalogs they’re overhead on an easy decision. - The tools carry their own trust problem. Only **35.4 percent** of shoppers trust AI recommendations, and 75 percent trust sponsored results even less. Table of Contents On a Reddit thread about clothing returns, a shopper who posts as *joidea* confessed her habit. She buys often from ASOS, returns around 80% of it, mostly items that didn’t fit or that she’d ordered in multiple sizes to compare in person. The retailer once emailed to say it was watching her account. She kept shopping there anyway. Further down the same thread, another commenter described a spouse’s version. He bought the same item in three sizes, kept one, sent two back, roughly $200 a trip and maybe a third of it staying in the closet. That’s evidence less of how common the behavior is than of how shoppers explain it to themselves. Neither called it indecision; it read more like a workaround, a return box doing a fitting room’s job. Retailers call this bracketing and treat it as fraud prevention or a logistics cost, missing what it’s telling you. Bracketing is a customer solving, at her own expense and the retailer’s, a problem upstream couldn’t solve for her. The website couldn’t tell her if the jeans would fit, so she bought her way around the uncertainty and let the reverse-logistics system absorb the cost. She isn’t abusing the return system. She’s using the only tool the retailer gave her. Bracketing is a customer solving, at their own expense and the retailer’s, a problem upstream couldn’t solve for them Much of e-commerce innovation over the past two decades optimized around one goal, helping people find and buy things fast. Search bars, filters, recommendation engines, and reviews all serve discovery, and they’re good at it. None solved confidence, the narrower, older problem of knowing whether *this specific item* will work for *this specific person*. This piece calls that unresolved space the commerce confidence gap. Confidence, here, means something specific. It’s the customer’s belief that a product will match her expectations before she spends money. Fit, trust, personalization, video, recommendations, everything discussed below, are different attempts to build that belief. Discovery closed one side of the gap decades ago; the belief itself is still missing. Discovery asks what products exist. A search bar takes “black dress” and returns thousands of results in under a second. Confidence asks a different question entirely. *A black dress for someone 5’4″, broad-shouldered, attending a summer wedding, who hates a tight sleeve*. Nothing on a product page reads that query. For a retailer, that gap shows up as a line item. The company pays to ship and restock a return, or issues a discount to keep an angry customer from walking. A bad guess that converts still chips away at lifetime value, since a customer stung once shops more cautiously next time. A good salesperson closed that gap through specific moves. They asked the question a customer hadn’t thought to ask and offered proof beyond their own opinion. Retail [economists](https://pubsonline.informs.org/doi/10.1287/isre.2014.0520) have long split product uncertainty into two buckets, quality (is it well made) and fit (does it suit me). Ecommerce solved the first; it’s still choking on the second, which drives returns on experience goods like apparel and furniture, versus search goods like a laptop, where the spec sheet tells you everything before you buy. The uncertainty carries an emotional cost too. Order a size up by mistake and appearance self-esteem [takes a real hit](https://doi.org/10.1016/j.jcps.2013.07.003), prompting some to spend more elsewhere to feel better. The reverse holds too. A suspiciously small size label [produces](https://doi.org/10.1016/j.jcps.2011.12.001) a more flattering mental image before the box arrives, part of why vanity sizing persists. None of that shows up on a returns dashboard. Five years ago there wasn’t much to do beyond tightening a return policy. Chatbots, AR, personalization, and live shopping existed in some form, but none were cheap or mature enough to become a practical confidence layer. Language models can now hold something close to a real conversation. Computer vision has gotten cheap enough to map a face or body from a phone. Video infrastructure that once needed a broadcast budget now runs on the same cloud stack as everything else. Skepticism is warranted. The convenient story says AI removes the uncertainty a salesperson once closed. The plainer story is that it moves the uncertainty from a physical interaction into a digital one, and a digital guess isn’t automatically better than a printed chart. Still, the shift is real. In many mature ecommerce categories, discovery is largely handled; what’s left unsolved is the moment right after, and the industry hasn’t found a dominant answer yet. ## Confidence has a cost The scale of the gap shows up in what’s spent cleaning up after it. NRF’s 2025 Retail Returns Landscape [report](https://nrf.com/research/2025-retail-returns-landscape) projected $849.9 billion in U.S. retail merchandise returns for 2025, roughly 19.3 percent of online sales; Appriss Retail’s [2026 report](https://apprissretail.com/2026-total-retail-loss-benchmark-report/) put 2025 returns at $706 billion. The figures don’t reconcile to the dollar, but they agree on magnitude. Appriss puts fraud at 14.2 percent, NRF at 9. Most of it is a customer who didn’t know, before buying, whether the product would work. Money that large goes somewhere. [Happy Returns](https://en.wikipedia.org/wiki/Happy_Returns_(company)) built a network of drop-off “Return Bars” and was profitable enough as an idea that UPS bought it outright in 2022. Optoro runs the same game at scale, in a [returns automation market](https://marketintelo.com/report/e-commerce-reverse-logistics-returns-automation-market) valued around $5.8 billion in 2025. None of that spend closes the confidence gap, it’s the cost of cleaning up after it stays open. Take a single order. A $60 apparel item carries a typical D2C gross margin, around $33 in gross profit, before anything comes back. Industry estimates for the [full cost of processing a return](https://eightx.co/blog/average-ecommerce-return-rate), reverse shipping, inspection, restocking, markdown, cluster in the $20 to $45 range depending on category. Run that at a 20% return rate across 100 orders and gross profit drops from a theoretical $3,300 to near $2,200, a swing of about a third, before marketing spend even enters the picture. That’s the CFO’s version of the confidence gap. Eighty-two percent of consumers told NRF that free returns are an important purchase factor, up from 76% the year before. That’s not just Reddit venting. The same report quantifies the behavior we saw in the thread. Close to two-thirds admit to at least one costly returns habit, and just under half say bending the truth on a return is acceptable when unsatisfied. Joidea and the commenter with the $200-a-trip habit are two data points, not the whole picture, but the survey says they’re not outliers. The workarounds people invent are the clearest evidence of what’s missing. In the joidea thread, commenters traded tactics like using the retailer’s own size chart over a generic one, a YouTube try-on haul, photos viewed on a laptop instead of a phone, a shopper reconstructing the questions a salesperson used to answer. Cart abandonment tells a related but distinct story. [Baymard Institute](https://baymard.com/lists/cart-abandonment-rate) puts average online abandonment near 70 percent, driven mostly by cost and checkout friction rather than confidence. The two get conflated, though confidence tech targets only the second. Ask several sources how often apparel gets returned online and the answers won’t match, though none are wrong. NRF and Appriss put U.S. apparel returns around 12.2% against 4.3% for beauty. European data compiled by [Eightx](https://eightx.co/blog/eu-ecommerce-return-rate-benchmark) tells the same relative story at different absolute numbers, apparel running six to thirty percent depending on whether the count is order-level or item-level. Apparel carries far more uncertainty than standardized goods; the numbers vary mainly because studies count differently, not because the reality is unclear. ## Does your catalog need this? Not every confidence gap deserves closing; the question is whether a given business’s version of it justifies the effort. Three questions get most of the way there, and a finance team already tracks the numbers that answer them, return rate, conversion rate, acquisition cost, repeat purchase. Can a customer evaluate the product from specifications alone? Does guessing wrong create a costly return rather than a cheap, shrug-it-off mistake? Does expertise, a salesperson’s opinion, a stylist’s read, meaningfully change what a customer decides to buy? A catalog that answers yes to most of these, fashion, furniture, luxury goods, beauty, higher-end electronics, has a real gap to close; one that answers no, commodities, repeat purchases, standardized products, mostly doesn’t. | Low price point | High price point | | |---|---|---| Low purchase uncertainty (standardized specs, known brand, replenishment items) | Limited value. Commodity purchases rarely justify additional decision-support layers. | Marginal case. A known laptop model from a trusted brand rarely needs live demonstration. | High purchase uncertainty (fit, feel, subjective performance) | Cheap to test. A size-guessing chatbot for a $30 t-shirt can pay for itself in avoided returns. | Strongest case. Furniture, jewelry, and premium apparel carry both the uncertainty and the margin to justify a human or AI-assisted consultation. | One commenter on a Reddit thread about virtual fitting rooms summed up the grid better than any framework could. He’d risk a $75 pair of jeans on home delivery, but wanted a $750 suit judged in person and fitted. Nobody had to explain price times uncertainty to him, he’d already priced it himself. A €5,000 sofa and a €5 tube of toothpaste sit at opposite ends of this grid. High uncertainty and high price is where video consultation and AI-guided recommendation earn their cost fastest; low uncertainty and low price is where the same tools become overhead on a decision that was never hard. ## Ecommerce returns: The case against the confidence economy None of this is as settled as the argument so far suggests. Returns aren’t always a failure to prevent; for some retailers they’re a deliberate feature, converting hesitant browsers who wouldn’t otherwise buy, and the extra revenue outweighs the processing cost, an argument Zappos has made publicly for years. Convenience, not certainty, may be what most shoppers optimize for. Plenty buy quickly and return casually because agonizing over fit isn’t worth their time, and a frictionless return is cheaper than a video consultation nobody asked for. Some categories resist digitization no matter how good the tooling gets; no camera substitutes for the tactile experience shoppers want in a store. Human salespeople brought their own friction and bias too, commission incentives that pushed the wrong product, gatekeeping that made some shoppers feel unwelcome. Removing that person wasn’t a pure loss. These objections don’t argue the confidence gap is fake, just that closing it isn’t the right investment for every retailer. The returns-as-feature case holds for a retailer whose margins can absorb it, but collapses for a smaller operator where a 20 percent return rate is existential. The tools built to close this gap carry a [trust problem](https://industrycontents.com/ai-mention-vs-recommendation/) engineering doesn’t dissolve. Only 35.4 percent of shoppers told Alchemer’s 2026 [Retail Report](https://www.alchemer.com/resources/benchmark-report/2026-retail-report-ai/) they trust AI recommendations completely or mostly, 22.2 percent don’t trust them at all, and head-to-head, AI loses badly to online reviews. [A Harris Poll for Quad](https://www.quad.com/newsroom/americans-say-they-would-lose-trust-in-ai-shopping-if-results-were-sponsored) found 75 percent of Americans would trust AI shopping results less if they knew the recommendations were sponsored, a live concern given how many assistants are built by the retailers selling what they recommend. ## Rebuilding the salesperson’s missing functions Video, chat, personalization, and human escalation are four implementations of the same repair job. Each rebuilds a part of what a salesperson used to do before ecommerce deleted the role. | Human function lost | Digital replacement attempt | |---|---| | Understanding context and asking the right question | AI shopping assistant | | Showing the product in motion, not just in a photo | Live and one-to-one video | | Remembering preferences and purchase history | Personalization engine | | Escalating to real expertise when the stakes are high | Human-AI hybrid, routed by risk | There’s a paradox underneath all four. AI isn’t creating the confidence layer by itself; it’s one component in a broader shift toward assisted commerce that includes plain video and better personalization data too. AI can reproduce information, a fit chart, a shade match, an answer the customer didn’t know to ask for. What it struggles to reproduce is trust. Live video reproduces the in-store demonstration, and it has the most concrete evidence behind it here, two named retailers, real quotes, checkable figures. Tula, a small boutique, had stylists going live on Instagram to unbox arrivals with no way to connect what a viewer watched to what she could buy. The brand moved that format onto Bambuser, a video commerce platform, so products could be tagged and added to cart mid-video. “The ease of adding to cart while an item is being styled has been a game-changer for reducing website friction,” said Alysa Holden, Tula’s e-commerce and marketing director. Tula reports a 45% engagement rate and 43% click-through to a product, figures published by Bambuser rather than independently audited. [Elkjøp, the largest Nordic electronics retailer](https://bambuser.com/article/ecommerce-video-consultation-complete-guide), runs the same pattern at scale. It logs more than 3,000 one-to-one video consultations a week, a 30 percent conversion rate, and a $470 average order value, by its own figures. A boutique and a national chain landing on the same tool is a stronger signal than either alone, though both numbers come from a vendor’s showcase. AI shopping assistants are built to reproduce the “let me ask you a few questions” recommendation. [Rep AI’s 2025 report](https://go.hellorep.ai/hubfs/2025_Rep_AI_eCommerce_Shopper_Behavior_Report.pdf) found conversion jumping from 3.1 to 12.3 percent among shoppers who engaged its chat, and checkout 47 percent faster, a vendor measuring its own product against shoppers likely already closer to buying. Personalization engines bring back the recognition a regular gets from someone who remembers her size. Beauty makes the clearest case, since color is harder to judge from a photo than size is from a chart. [Sephora’s Virtual Artist](https://www.techrepublic.com/article/how-sephora-is-leveraging-ar-and-ai-to-transform-retail-and-help-customers-buy-cosmetics/) has logged over 200 million shades tried on since 2016, though the sales lift is too confounded with a pandemic ecommerce surge and heavy parent-company investment to credit to the tool alone. Hybrid associate models route routine questions to AI and escalate high-stakes ones to a person. None of these four pieces is new. What’s new is deploying all four at once without proportionally more humans. ## What retailers offer today A review of leading retailers across three categories shows how differently companies approach the confidence problem, though three is a snapshot, not a market study. [SHEIN](https://m.shein.com/us/How-to-choose-your-size-a-748.html) offers the fullest stack of the three. It runs a static size chart, a “Check My Size” fit finder, a saved sizing profile, and reviews filterable by size, though it stops short of an on-page try-on. [IKEA](https://www.ikea.com/global/en/newsroom/innovation/ikea-launches-ikea-place-a-new-app-that-allows-people-to-virtually-place-furniture-in-their-home-170912/) does something apparel retailers don’t attempt. Its IKEA Place app places a true-to-scale 3D sofa model into a customer’s room through a phone camera, claiming 98 percent scale accuracy since 2017, though the tool lives in a separate app rather than the product page. [Apple’s MacBook Air](https://www.apple.com/macbook-air/specs/) page offers neither, just precision. It lists exact dimensions and weight, skips the AR viewer, skips the fit recommender. A laptop barely needs one. What the three pages show is that the confidence layer looks different by category. It’s informational and algorithmic for apparel, spatial and app-based for furniture, and functionally absent for standardized electronics. None of the three solves it end to end. ## The honest limits This doesn’t fully close the gap it’s aimed at, and the people building it are often the first to say so. The tools improve the online experience, but you can’t touch the fabric. That means it’s not a full replacement, it’s an alternative that runs alongside the real thing. Comfort matters more than appearance. During the pandemic, many people ordered twenty or thirty outfits and returned whatever didn’t feel right. No visual fidelity fixes that, since comfort is a sensory problem, not an informational one. Shade matching and sizing guidance are informational problems cameras and better data can solve. Comfort isn’t. The thing missing was never a fact she lacked, it was a sensation only the object could provide. Synthetic confidence doesn’t touch distrust in payment security or a customer already burned by hard returns, and in cheap, disposable categories, no demonstration moves the needle, since the purchase was never uncertain to begin with. ## The next competitive layer Ecommerce spent twenty years making it easier to find something to buy, cutting clicks, speeding up search, enabling one-tap checkout. Almost none of that effort went toward the slower problem a salesperson solved, making a stranger feel sure enough to say yes. Joidea is still out there right now, ordering something in two sizes because nobody built her a better way to know which one fits. For now, that guesswork happens on the retailer’s own site, so the retailer still owns the moment and the data that comes out of it. The first era of e-commerce was about helping customers find products; the next will be defined by whoever reduces the cost of uncertainty first. That won’t hold by default. If shoppers start asking an AI assistant instead of a product page whether something will fit, the confidence layer moves away from retailers, and whoever owns that conversation next owns the customer. ## FAQ The commerce confidence gap is the disconnect between a shopper’s ability to locate a product and their certainty it will meet their needs. Discovery tools have improved; the remaining uncertainty still drives checkout hesitation and high return rates. Bracketing is when customers buy multiple variants of an item intending to return most, usually for lack of reliable fit information. It turns the returns process into a virtual fitting room. Online returns run into the hundreds of billions of dollars a year in the U.S. alone, with estimates ranging from $706 billion to $849.9 billion for 2025. A minority of that is fraud; most of it stems from customers who couldn’t be sure a product would work before they bought it. AI tools can show higher conversion rates among engaged users, though that’s partly because interested shoppers are more likely to use them. Building genuine trust in AI recommendations remains a challenge. Categories where fit, feel, or subjective performance matter—apparel, furniture, premium electronics—benefit most from these tools. Commodity goods and repeat-purchase items see little impact, since the risk of guessing wrong there is minimal.