The landscape · Retail and e-commerce
Retail and e-commerce: what AI is really doing
Brief me in 2 minutes
AI in Retail and e-commerce.
AI now reads the search query, the camera-roll photo and the grocery list before a person ever sees the order, and the traffic numbers back that up: Shopify says AI-agent traffic to its merchants tripled in a single quarter, and AI referral traffic to US retailers rose 393% in the first quarter of 2026. What is still mostly plumbing and turf war is the last step, the actual payment: OpenAI, Google and Meta each built a way for an agent to check out, Amazon is fighting to keep agents off its own site, and PYMNTS Intelligence found only 24% of Americans would let one pay on their behalf.
The game: we show you 5 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 5
Shopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
Early
The protocols (ACP, UCP, AP2, Link) are real and shipped by real platforms, but the checkout share they move is still small and contested, and the incumbent with the biggest catalog is actively blocking agents rather than integrating them.
How it works. OpenAI turned ChatGPT's shopping features into one-tap purchases with Instant Checkout, then open-sourced the Agentic Commerce Protocol it built with Stripe so any merchant can plug in. Google answered with a rival Universal Commerce Protocol layered on its own Agent Payments Protocol, and Meta's Muse agent started buying for users through Stripe's Link.
Overstated: Merchants have already hung up mid-call on Meta's Muse agent the moment they realized the caller was AI, and Tools for Humanity built a separate product just to prove a human, not a bot, authorized the purchase, so the trust problem is still unsolved underneath the protocol war.
Shift 2 of 5
AI decides what you see when you search for a product, and the traffic already proves it
Real now
This is Shopify's own measured merchant traffic and a nationwide traffic count, not a demo, though Amazon's fake-photo feature shows the same AI is also being pointed at problems nobody asked it to solve.
How it works. Shopify says AI-agent traffic and orders to its merchants tripled year over year, and nationwide AI referral traffic to US retailers rose 393% in the first quarter of 2026 as more shoppers used AI assistants to browse. Amazon started showing AI-generated product images in search results instead of real photos, and fashion app Daydream lets a shopper point a camera at an outfit saved in their own photos to shop the individual pieces in it.
Overstated: Senators have asked the FTC to investigate why Amazon's own Alexa for Shopping reportedly cannot filter for Made in USA products, and the chatbot's own reported answer was that doing so would redirect sales away from Amazon's largest seller base, so AI search still serves the platform's revenue before a shopper's stated preference.
Shift 3 of 5
Retailers hand shelf-level inventory and CPG deduction disputes to AI, and pricing regulators are catching up
Early
The shelf-data and deduction-automation products are already live and billed at real retailers and CPG brands, but the pricing side of this shift is still being defined by which state bans what next, not by a settled market.
How it works. Instacart built software that stitches together more than 10 million shelf-level data points a day from a network of shopper phones, so grocers and CPG brands can see what is actually in stock instead of relying on periodic manual counts. Glimpse runs AI agents over the deduction notices retailers use to subtract money from CPG invoices, flags the invalid ones, and automatically files the dispute paperwork to recover cash.
Overstated: Instacart's own pitch describes the data gap it fills more than proof it lifts sales, and Maryland's ban only covers one food category at stores over 15,000 square feet while explicitly exempting loyalty programs, so most AI personalized pricing elsewhere remains untouched.
Shift 4 of 5
One AI agent now runs an entire customer service thread, and it gets judged on conversion, not novelty
Real now
Rivet is a named, shipped system replacing a real call-center workload with measured results, not a pilot; the limit is what shoppers themselves will delegate, not the technology.
How it works. Crocs replaced a chatbot that routed 80% of its 178,000 monthly service inquiries to a human agent with Rivet, a single Salesforce Agentforce-built agent that handles product search, sizing, order tracking, returns and checkout in one conversation, built in five months with an outside implementation partner. Zendesk, meanwhile, bought agentic customer-service startup Forethought, which by 2025 was already supporting more than a billion monthly customer interactions.
Overstated: PYMNTS Intelligence found half of Americans have already made a retail purchase with AI's help, but only 24% would let an agent shop and pay for them, exactly the gap between what these assistants already do well and what the checkout-autonomy pitch claims they are about to do.
Shift 5 of 5
The marketplace, the ad engine and the recommendation feed are merging into one AI system a seller doesn't fully control
Early
Seller-side automation and marketplace recommendation upgrades are live and adopted at real scale, but a fully autonomous marketplace agent operating without seller sign-off does not exist yet, by Amazon's own description of how it works.
How it works. Amazon turned its Seller Assistant from a question-answering chatbot into a system that reasons across a merchant's inventory, pricing, advertising, demand and compliance continuously, with sellers now accepting the large majority of its recommendations. Topsort made OpenAI Ads available to the marketplaces it powers, so retailers can reach shoppers inside ChatGPT without building new ad infrastructure.
Overstated: Amazon's own executive frames this as trust earned over time, not autonomy granted upfront: sellers can always review and accept a recommendation or not, which means "AI running your store" still means AI drafting the move and a human clicking accept.
The scary part
What happens to jobs in Retail and e-commerce
What AI is already taking:
- Resolving day-to-day service requests, order tracking, sizing, returns questions, in one continuous AI conversation instead of routing to a human rep PYMNTS ↗
- Deciding what product images and results show up first when a shopper searches TechCrunch ↗
- Reviewing a retailer's deduction invoices and filing the recovery dispute paperwork for a CPG brand TechCrunch ↗
- Monitoring a marketplace seller's inventory, pricing, ad spend and compliance continuously and running standing automations PYMNTS ↗
What stays human:
Handing over the card is still a human call: PYMNTS found only 24% of Americans would let an agent shop and pay for them even though half have already used AI somewhere in a purchase. Ron Johnson, who built Apple's retail stores, put it plainly about a $1,000 laptop: nobody is letting an agent choose and buy one without ever touching it first.
Get better · 15 minutes
The ability to audit an AI shopping or seller agent's actual output, the price it showed, the recommendation it made, the dispute it filed, against the retailer's own raw records instead of trusting the vendor's self-reported acceptance rate.
Try this this week:
Pick a product you're actually about to buy this week. Ask ChatGPT to find it and walk through Instant Checkout, or ask Perplexity's Comet to shop for it, without leaving the chat. Then check the retailer's own site for the same item and compare price, size and stock; any gap between what the agent showed you and what the store's page says is what you should not yet trust unsupervised.
Your move
The seller-operations layer is getting absorbed by the platforms themselves, Amazon's own Seller Assistant now reasons across pricing, ads and compliance, so a thin wrapper around that job is a weak bet. The open space is the audit and trust layer underneath agentic commerce: Tools for Humanity built AgentKit to prove a human authorized a purchase, but nobody has built the equivalent check for a retailer, confirming an agent's claimed price match or stock count against the store's own system of record.
Watch whether a retailer's AI acceptance or resolution number survives being checked against its own raw logs at renewal, the way Amazon's sellers accepting most of Seller Assistant's recommendations should be verifiable against actual seller activity. Red flag in a pitch: a startup selling full checkout autonomy when even Amazon still requires explicit approval for automated seller actions and PYMNTS found most shoppers still won't hand payment to an agent.
Pull whatever AI product-search or shopping-assistant tool your store already runs, and manually search for five of your own SKUs the way senators say Amazon's Alexa for Shopping was tested for Made in USA products. If the assistant can't find or correctly label products you know are in stock, that gap between the vendor demo and your real catalog is your fix list before the next planning meeting.
Learn to read a retail AI vendor's acceptance or resolution rate as a supervised-recommendation number, not full autonomy, the way Amazon's own seller-assistant numbers describe sellers reviewing and accepting a suggestion rather than a system acting alone. Start by asking the account team behind one AI tool your employer already pays for what its actual human override rate is behind the published number.
What's actually changing
- 1
Shopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
How it works. OpenAI turned ChatGPT's shopping features into one-tap purchases with Instant Checkout, then open-sourced the Agentic Commerce Protocol it built with Stripe so any merchant can plug in. Google answered with a rival Universal Commerce Protocol layered on its own Agent Payments Protocol, and Meta's Muse agent started buying for users through Stripe's Link. Amazon, the retailer with the most to lose, sent Perplexity's Comet agent a legal threat instead of an integration.
Why it matters. Whoever owns the tap-to-buy moment keeps the transaction data and the merchant relationship, so every platform is racing to open that door for agents while the incumbent with the largest catalog and ad business tries to keep it shut.
What's overstatedMerchants have already hung up mid-call on Meta's Muse agent the moment they realized the caller was AI, and Tools for Humanity built a separate product just to prove a human, not a bot, authorized the purchase, so the trust problem is still unsolved underneath the protocol war.
OpenAIGoogleMetaAmazonPerplexityStripe
3 receipts
- OpenAI open-sourced the Agentic Commerce Protocol, built with Stripe, that powers ChatGPT's Instant Checkout for merchants.
it will open source its Agentic Commerce Protocol (ACP), the tech that powers Instant Checkout built with Stripe, so that other merchants and developers can integrate agentic checkout
TechCrunch, 29 Sept 2025 ↗ - Google announced its own Universal Commerce Protocol for AI agent-based shopping at the NRF conference.
Google on Sunday announced a new open standard called the Universal Commerce Protocol (UCP) for AI agent-based shopping, at the National Retail Federation (NRF) conference.
TechCrunch, 11 Jan 2026 ↗ - Amazon sent Perplexity a legal threat over its Comet agent shopping on Amazon on users' behalf.
This week, Perplexity received an aggressive legal threat from Amazon, demanding we prohibit Comet users from using their AI assistants on Amazon.
TechCrunch, 4 Nov 2025 ↗
- OpenAI open-sourced the Agentic Commerce Protocol, built with Stripe, that powers ChatGPT's Instant Checkout for merchants.
- 2
AI decides what you see when you search for a product, and the traffic already proves it
How it works. Shopify says AI-agent traffic and orders to its merchants tripled year over year, and nationwide AI referral traffic to US retailers rose 393% in the first quarter of 2026 as more shoppers used AI assistants to browse. Amazon started showing AI-generated product images in search results instead of real photos, and fashion app Daydream lets a shopper point a camera at an outfit saved in their own photos to shop the individual pieces in it.
Why it matters. The entry point to a purchase is moving from a typed keyword to an AI reading an image or an intent, which rewards clean structured product data over a decade of keyword-focused SEO, and shifts real dollars of traffic toward whichever store shows up well in an AI answer.
What's overstatedSenators have asked the FTC to investigate why Amazon's own Alexa for Shopping reportedly cannot filter for Made in USA products, and the chatbot's own reported answer was that doing so would redirect sales away from Amazon's largest seller base, so AI search still serves the platform's revenue before a shopper's stated preference.
ShopifyAmazonDaydream
3 receipts
- AI-driven traffic and orders to Shopify's merchant stores tripled year over year in the second quarter of 2026.
The company noted that AI-driven traffic and orders to Shopify stores had tripled year-over-year in the second quarter.
TechCrunch, 5 Aug 2026 ↗ - AI traffic to US retailers rose 393% in the first quarter of 2026 compared to a year earlier.
And in the first three months of 2026, AI traffic had risen 393% compared to a year earlier, as more consumers used AI assistants for online shopping.
TechCrunch, 16 Apr 2026 ↗ - Amazon began showing AI-generated product images in shopping search results instead of real product photos.
Amazon announced on Wednesday that it will display AI-generated images of products within its shopping app based on users' search queries.
TechCrunch, 3 June 2026 ↗
- AI-driven traffic and orders to Shopify's merchant stores tripled year over year in the second quarter of 2026.
- 3
Retailers hand shelf-level inventory and CPG deduction disputes to AI, and pricing regulators are catching up
How it works. Instacart built software that stitches together more than 10 million shelf-level data points a day from a network of shopper phones, so grocers and CPG brands can see what is actually in stock instead of relying on periodic manual counts. Glimpse runs AI agents over the deduction notices retailers use to subtract money from CPG invoices, flags the invalid ones, and automatically files the dispute paperwork to recover cash. Maryland became the first state to ban AI-driven personalized grocery pricing outright rather than just requiring disclosure.
Why it matters. The forecasting, shelf and pricing layer used to run on a planner's spreadsheet and a buyer's instinct; now a vendor model decides what gets restocked, whether a CPG dispute is worth fighting, and increasingly whether two shoppers pay the same price, which is where the real savings and the real legal exposure both land.
What's overstatedInstacart's own pitch describes the data gap it fills more than proof it lifts sales, and Maryland's ban only covers one food category at stores over 15,000 square feet while explicitly exempting loyalty programs, so most AI personalized pricing elsewhere remains untouched.
InstacartGlimpse
3 receipts
- Instacart's shelf-data tool draws on roughly 600,000 shoppers generating more than 10 million data points a day.
Instacart is using data generated by a network of around 600,000 shoppers who visit stores across North America more than 15 times per store per day on average, leading to more than 10 million unique data points each day.
PYMNTS, 21 Sept 2026 ↗ - Glimpse raised a $35 million Series A led by Andreessen Horowitz to expand its AI deduction-dispute automation for CPG brands.
Dispute-tracking fintech Glimpse announced Wednesday that it raised a $35 million Series A led by Andreessen Horowitz, with participation from 8VC and Y Combinator.
TechCrunch, 25 Mar 2026 ↗ - Maryland's new law bans grocers from charging a shopper a different price for tax-exempt food based on personal data.
It bars them from using "dynamic pricing," defined as setting a consumer-specific price based on personal data, to charge a higher price for covered food.
PYMNTS, 21 Sept 2026 ↗
- Instacart's shelf-data tool draws on roughly 600,000 shoppers generating more than 10 million data points a day.
- 4
One AI agent now runs an entire customer service thread, and it gets judged on conversion, not novelty
How it works. Crocs replaced a chatbot that routed 80% of its 178,000 monthly service inquiries to a human agent with Rivet, a single Salesforce Agentforce-built agent that handles product search, sizing, order tracking, returns and checkout in one conversation, built in five months with an outside implementation partner. Zendesk, meanwhile, bought agentic customer-service startup Forethought, which by 2025 was already supporting more than a billion monthly customer interactions.
Why it matters. Customer service AI is now judged against metrics a retailer already tracks, handling time and duplicate-profile cleanup, not a separate innovation budget, which is why it is spreading through retail faster than the flashier shopping-agent story, and why incumbents are buying the startups that built it rather than building it themselves.
What's overstatedPYMNTS Intelligence found half of Americans have already made a retail purchase with AI's help, but only 24% would let an agent shop and pay for them, exactly the gap between what these assistants already do well and what the checkout-autonomy pitch claims they are about to do.
CrocsSalesforceZendeskForethought
3 receipts
- Crocs' old chatbot routed 80% of its 178,000 monthly service inquiries to a human before Rivet replaced it.
Before Rivet, the footwear company's service team was handling 178,000 inquiries a month across voice, chat and email, Salesforce said in a customer case study.
PYMNTS, 22 Sept 2026 ↗ - Only 24% of Americans would let an AI agent shop and pay for them, even though half have used AI in a purchase.
Half of Americans say they've made a retail purchase with AI's help, but only 24% would let an agent shop and pay for them, PYMNTS Intelligence reported
PYMNTS, 22 Sept 2026 ↗ - Forethought, acquired by Zendesk, was supporting more than a billion monthly customer interactions by 2025.
and by 2025 was supporting more than a billion monthly customer interactions
TechCrunch, 11 Mar 2026 ↗
- Crocs' old chatbot routed 80% of its 178,000 monthly service inquiries to a human before Rivet replaced it.
- 5
The marketplace, the ad engine and the recommendation feed are merging into one AI system a seller doesn't fully control
How it works. Amazon turned its Seller Assistant from a question-answering chatbot into a system that reasons across a merchant's inventory, pricing, advertising, demand and compliance continuously, with sellers now accepting the large majority of its recommendations. Topsort made OpenAI Ads available to the marketplaces it powers, so retailers can reach shoppers inside ChatGPT without building new ad infrastructure. Whatnot bought recommendation-engine startup Shaped to speed up how its live-shopping marketplace matches a buyer to a listing in a video feed that changes by the second.
Why it matters. The marketplace, the ad auction and the recommendation feed are becoming one AI system a seller doesn't fully see inside, so the sellers and brands who win are the ones who learn to operate the algorithm, not just list a product and wait.
What's overstatedAmazon's own executive frames this as trust earned over time, not autonomy granted upfront: sellers can always review and accept a recommendation or not, which means "AI running your store" still means AI drafting the move and a human clicking accept.
AmazonTopsortOpenAIWhatnot
3 receipts
- Amazon's Seller Assistant now reasons across a merchant's inventory, pricing, advertising, demand and compliance, not just answering questions.
Sellers can describe automations in plain language and leave Seller Assistant running continuously.
PYMNTS, 23 Sept 2026 ↗ - Topsort made OpenAI Ads available through its retail-media infrastructure so marketplaces can reach shoppers inside ChatGPT.
Learn how marketplaces can enable OpenAI Ads through Topsort to reach shoppers in ChatGPT
Topsort, Sept 2026 ↗ - Whatnot acquired recommendation-engine startup Shaped to make its live-shopping matches faster and more personalized.
By combining Shaped's technology with Whatnot's existing systems, we can make recommendations faster, more responsive, and more personalized.
TechCrunch, 15 July 2026 ↗
- Amazon's Seller Assistant now reasons across a merchant's inventory, pricing, advertising, demand and compliance, not just answering questions.
Who's doing it
9 companies to know, 5 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
DaydreamAI-powered fashion discovery and visual search
Fashion discovery app that lets shoppers search by photo or voice; its newest feature analyzes an outfit saved in a shopper's own camera roll and finds matching pieces to buy.
Latest · 14 Sept 2026Daydream's image-shopping feature matches an outfit photo against its catalog of roughly 3 million products. TechCrunch ↗
Topsortretail media and commerce-ad infrastructure
Ad-auction and personalization infrastructure that lets marketplaces and retailers run a retail-media advertising business without building their own ad server.
Latest · Sept 2026Topsort made OpenAI Ads available to the marketplaces it powers, extending retail media into ChatGPT's AI-powered discovery. Topsort ↗
$20M Series A, 4 Mar 2024 source ↗
Malachytereal-time product personalization
Real-time, intent-aware personalization engine that reorders a store's product grid based on a shopper's clicks and hovers in the current session, no account or purchase history required.
Latest · 6 Aug 2026Malachyte's personalization engine became generally available to Shopify merchants through a native integration in June 2026. TechCrunch ↗
$10M Seed, 6 Aug 2026 source ↗
Shapedreal-time recommendation and search infrastructure
Machine learning company building real-time recommendation and search systems that combine customer data with large language models for personalized product discovery.
Latest · 15 July 2026Shaped was acquired by live-shopping marketplace Whatnot to power faster, more personalized real-time recommendations. TechCrunch ↗
Whatnotlive-shopping marketplace
Live-video shopping marketplace where sellers auction and sell items on camera; now leans on acquired AI recommendation technology to match buyers to listings in real time.
Latest · 15 July 2026Whatnot acquired recommendation-engine startup Shaped to make its live-shopping recommendations faster and more personalized. TechCrunch ↗
$265M Series E, 8 Jan 2025 source ↗
GlimpseCPG deduction and dispute automation
AI agents that review the deductions retailers subtract from CPG brand invoices, flag the invalid ones, and file and follow through on the recovery dispute automatically.
Latest · 25 Mar 2026Glimpse now works with more than 200 retail brands, including Suave and Chapstick. TechCrunch ↗
$35M Series A, 25 Mar 2026 source ↗
Tools for Humanityidentity verification for AI agents
Builds World ID biometric identity infrastructure and shipped AgentKit, letting a website confirm a real, unique human authorized an AI shopping agent's purchase.
Latest · 17 Mar 2026Tools for Humanity launched AgentKit in beta to let websites verify a real human is behind an AI shopping agent's purchase. TechCrunch ↗
Forethoughtagentic customer service
Agentic customer-service platform used by e-commerce and software companies to automate support interactions end to end; a 2018 TechCrunch Battlefield winner acquired by Zendesk in 2026.
Latest · 11 Mar 2026Forethought was acquired by Zendesk after growing to support more than a billion monthly customer interactions by 2025. TechCrunch ↗
$115M total raised (Series A-C), 11 Mar 2026 source ↗
PerplexityAI answer engine with a shopping/browsing agent
AI search company whose Comet browser agent can research and buy products on a user's behalf across retail sites, including Amazon.
Latest · 4 Nov 2025Amazon sent Perplexity a legal threat demanding it stop its Comet agent from shopping on Amazon on users' behalf. TechCrunch ↗
Your move
Our read, built on the receipts above.
The seller-operations layer is getting absorbed by the platforms themselves, Amazon's own Seller Assistant now reasons across pricing, ads and compliance, so a thin wrapper around that job is a weak bet. The open space is the audit and trust layer underneath agentic commerce: Tools for Humanity built AgentKit to prove a human authorized a purchase, but nobody has built the equivalent check for a retailer, confirming an agent's claimed price match or stock count against the store's own system of record.
Watch whether a retailer's AI acceptance or resolution number survives being checked against its own raw logs at renewal, the way Amazon's sellers accepting most of Seller Assistant's recommendations should be verifiable against actual seller activity. Red flag in a pitch: a startup selling full checkout autonomy when even Amazon still requires explicit approval for automated seller actions and PYMNTS found most shoppers still won't hand payment to an agent.
Pull whatever AI product-search or shopping-assistant tool your store already runs, and manually search for five of your own SKUs the way senators say Amazon's Alexa for Shopping was tested for Made in USA products. If the assistant can't find or correctly label products you know are in stock, that gap between the vendor demo and your real catalog is your fix list before the next planning meeting.
Learn to read a retail AI vendor's acceptance or resolution rate as a supervised-recommendation number, not full autonomy, the way Amazon's own seller-assistant numbers describe sellers reviewing and accepting a suggestion rather than a system acting alone. Start by asking the account team behind one AI tool your employer already pays for what its actual human override rate is behind the published number.
What's next
- 1 Oct 2026 Maryland's ban on AI-driven personalized grocery pricing takes effect pymnts.com ↗
Other industries
- FinanceBanks move agentic AI from chatbots to running back-office and dealmaking work
- Software engineeringSpaceX buys the leading AI coding startup outright, for $60 billion
- HealthcareAmbient scribes go from pilot project to clinical infrastructure
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing and resident calls move to voice AI agents that never sleep
- Sales and marketingCustomer support agents now get paid per resolved ticket, not per seat
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- InsuranceClaims move from adjuster-first to AI-first, humans supervise the exceptions
- CybersecurityAutonomous AI agents take over penetration testing and vulnerability hunting
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- EducationNYC and LA ban AI for students, not for teachers
- Media and creativeAI video splits into model labs that train and aggregators that resell
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare across a dozen-plus US cities
- Energy and data-centre powerGrid rules get rewritten mid-buildout, state by state
- GovernmentGSA turns AI adoption into a government-wide subscription, then flips 'free' to metered
- Agriculture and foodAutonomous tractor fleets learn to run the tools behind them, not just the wheel