ChatGPT adds a selfie-to-try-on step for fashion discovery
ChatGPT has launched virtual try-on for clothing and accessories. When a supported product appears in shopping results, the shopper can take or upload a selfie, select the try-on control and generate an image showing the item on them. Products can then be saved to Favorites or organized in Library folders.
The workflow sits between discovery and the retailer’s product page: browse, visualize, shortlist, then continue toward purchase. It can reduce the need to describe fit or styling in words, but the output is a generated preview, not proof of size, fabric behavior or exact color. Merchants should therefore check whether the selected listing, variant and destination page remain consistent after a shopper saves the preview.
ChatGPT has launched virtual try-on for clothing and accessories. When a supported product appears in shopping results, the shopper can take or upload a selfie, select the try-on control and generate an image showing the item on them. Products can then be saved to Favorites or organized in Library folders.
The workflow sits between discovery and the retailer’s product page: browse, visualize, shortlist, then continue toward purchase. It can reduce the need to describe fit or styling in words, but the output is a generated preview, not proof of size, fabric behavior or exact color. Merchants should therefore check whether the selected listing, variant and destination page remain consistent after a shopper saves the preview.
Poppy gives an AI agent one approval record for a commerce action
The new draft of Personal Agent Protocol includes an operations extension for actions such as exchanging an item. A store first proposes the exact terms; the shopper or their personal agent approves that revision; only then does the store perform it. The same operation ID follows the action across a website, API, MCP tool or support conversation.
The practical safeguard is “at most once”: if a request times out and the agent retries, the exchange should not run twice. This turns approval into a traceable workflow instead of a vague chat confirmation. The limitation matters too: draft 0.1 can still change, the extension is optional, and payments, shipment notifications and attachments such as return labels remain open topics.
The new draft of Personal Agent Protocol includes an operations extension for actions such as exchanging an item. A store first proposes the exact terms; the shopper or their personal agent approves that revision; only then does the store perform it. The same operation ID follows the action across a website, API, MCP tool or support conversation.
The practical safeguard is “at most once”: if a request times out and the agent retries, the exchange should not run twice. This turns approval into a traceable workflow instead of a vague chat confirmation. The limitation matters too: draft 0.1 can still change, the extension is optional, and payments, shipment notifications and attachments such as return labels remain open topics.
❤2
ThredUp tests AI output with a human review gate
ThredUp CEO James Reinhart presented a slide deck he had not seen before: Claude created it and his assistant checked it before the conference. The experiment shows a narrow content workflow inside a secondhand ecommerce company—generation first, human quality assurance second, executive use last.
The cost signal is equally concrete. ThredUp’s token consumption grew faster than Reinhart expected, so the company needs better forecasts even as it continues broad employee use. No productivity saving was measured in the report. A practical rollout therefore needs two separate controls: a reviewer who can verify the output, and a usage budget that records model spend by task instead of treating all AI activity as one line item.
ThredUp CEO James Reinhart presented a slide deck he had not seen before: Claude created it and his assistant checked it before the conference. The experiment shows a narrow content workflow inside a secondhand ecommerce company—generation first, human quality assurance second, executive use last.
The cost signal is equally concrete. ThredUp’s token consumption grew faster than Reinhart expected, so the company needs better forecasts even as it continues broad employee use. No productivity saving was measured in the report. A practical rollout therefore needs two separate controls: a reviewer who can verify the output, and a usage budget that records model spend by task instead of treating all AI activity as one line item.
❤2
Instacart turns ordinary aisle footage into a stock signal
Instacart’s Store View uses computer vision to analyze shelf footage captured by shoppers, store associates or side-facing cameras on Caper carts. The practical change is modest but useful: a grocer can add a view of on-shelf availability without first installing a new camera network or sending a separate inspection team through every aisle.
This is an opt-in enterprise workflow, not a promise of perfect inventory. The footage has to exist, and the system still needs accurate catalog and store data to connect a visible shelf gap with the right item. Instacart does not publish an accuracy rate or measured reduction in out-of-stocks on the product page. A sensible pilot would therefore compare detected gaps with physical checks in a limited set of stores before using the signal for replenishment or customer availability.
Instacart’s Store View uses computer vision to analyze shelf footage captured by shoppers, store associates or side-facing cameras on Caper carts. The practical change is modest but useful: a grocer can add a view of on-shelf availability without first installing a new camera network or sending a separate inspection team through every aisle.
This is an opt-in enterprise workflow, not a promise of perfect inventory. The footage has to exist, and the system still needs accurate catalog and store data to connect a visible shelf gap with the right item. Instacart does not publish an accuracy rate or measured reduction in out-of-stocks on the product page. A sensible pilot would therefore compare detected gaps with physical checks in a limited set of stores before using the signal for replenishment or customer availability.
❤1
One handheld workflow can turn an inventory query into a pick path
Stratix describes a store-associate workflow built around shared mobile devices. An employee can ask for inventory in natural language, generate a pick path, open training guidance or launch a task without moving through several menus.
For click-and-collect, that narrows the sequence to one device: confirm stock, receive the route, pick and scan the item, then hand the order to the pickup desk. The useful control point is the scan; a generated route should not override the store’s inventory record or substitute for item confirmation.
The article is sponsored by Stratix and presents vendor capabilities, not an independent deployment study. It gives no measured result for pick time, stock accuracy or completed orders, so those remain the metrics a retailer would need to test.
Stratix describes a store-associate workflow built around shared mobile devices. An employee can ask for inventory in natural language, generate a pick path, open training guidance or launch a task without moving through several menus.
For click-and-collect, that narrows the sequence to one device: confirm stock, receive the route, pick and scan the item, then hand the order to the pickup desk. The useful control point is the scan; a generated route should not override the store’s inventory record or substitute for item confirmation.
The article is sponsored by Stratix and presents vendor capabilities, not an independent deployment study. It gives no measured result for pick time, stock accuracy or completed orders, so those remain the metrics a retailer would need to test.
❤2
Amex puts the shopper’s constraints before an AI checkout
American Express has introduced an Intent Intelligence Framework for purchases made by AI agents. Its Create Intent API records the shopper’s instructions—such as merchant, product preferences and maximum budget—before payment. The shopper authenticates and approves that record, which produces an Intent ID and a proof token.
The agent then receives a secure payment credential instead of the customer’s card number. At checkout, the recorded constraints can expose a mismatch: an agent instructed to buy one brand or size should not silently substitute another. The same record can support fraud reviews and disputes after the purchase.
This is a framework inside Amex’s ACE developer toolkit, not evidence that autonomous buying is already widespread or dispute-free. Amex says it is discussing implementation with payment partners.
American Express has introduced an Intent Intelligence Framework for purchases made by AI agents. Its Create Intent API records the shopper’s instructions—such as merchant, product preferences and maximum budget—before payment. The shopper authenticates and approves that record, which produces an Intent ID and a proof token.
The agent then receives a secure payment credential instead of the customer’s card number. At checkout, the recorded constraints can expose a mismatch: an agent instructed to buy one brand or size should not silently substitute another. The same record can support fraud reviews and disputes after the purchase.
This is a framework inside Amex’s ACE developer toolkit, not evidence that autonomous buying is already widespread or dispute-free. Amex says it is discussing implementation with payment partners.
❤2🆒1
A cleaner product feed starts on the product page
Google can combine two inputs for a shoppable item: `Product` structured data on the page and a feed in Merchant Center. The page markup can expose price, availability, ratings, shipping and return details; the feed adds a managed catalog source. Using both helps Google compare what the storefront shows with what the merchant submitted.
The practical workflow is to generate markup from the same catalog record that supplies the feed, then test it whenever price or stock changes. Google separates merchant listings, where a shopper can buy the item, from editorial product snippets. Markup makes a page eligible for richer search treatment; it does not guarantee placement, ranking or an AI recommendation.
Google can combine two inputs for a shoppable item: `Product` structured data on the page and a feed in Merchant Center. The page markup can expose price, availability, ratings, shipping and return details; the feed adds a managed catalog source. Using both helps Google compare what the storefront shows with what the merchant submitted.
The practical workflow is to generate markup from the same catalog record that supplies the feed, then test it whenever price or stock changes. Google separates merchant listings, where a shopper can buy the item, from editorial product snippets. Markup makes a page eligible for richer search treatment; it does not guarantee placement, ranking or an AI recommendation.
❤2
Target puts shelf and backroom decisions in one employee app
Target has upgraded myDay, the app used by store employees, after collecting workflow feedback across 50 locations. The app now groups similar tasks and ranks to-do lists by urgency and value. Team leaders can customize a home screen with shelf-stocking or pickup-fulfillment metrics relevant to their role.
One concrete inventory step is Scan to Divert. When an employee scans newly arrived merchandise, the function uses demand and available space to indicate whether the item should move to the sales floor or remain in the backroom. Target attributes $25 million in reinvestable value to the function, though it does not publish the calculation in the report.
AI-backed insights are planned for a later app update, alongside task assignment and handoff notes; they are not described as live yet.
Target has upgraded myDay, the app used by store employees, after collecting workflow feedback across 50 locations. The app now groups similar tasks and ranks to-do lists by urgency and value. Team leaders can customize a home screen with shelf-stocking or pickup-fulfillment metrics relevant to their role.
One concrete inventory step is Scan to Divert. When an employee scans newly arrived merchandise, the function uses demand and available space to indicate whether the item should move to the sales floor or remain in the backroom. Target attributes $25 million in reinvestable value to the function, though it does not publish the calculation in the report.
AI-backed insights are planned for a later app update, alongside task assignment and handoff notes; they are not described as live yet.
❤2👏1
Lightspeed AI moves from an answer to a prepared retail action
Lightspeed has added an action layer to its retail assistant. A merchant can ask about stock or sales, then have the system prepare a product update, clearance tag, purchase order or inventory transfer inside the same conversation. The proposed change is shown before execution and waits for explicit approval; completed actions are recorded in an audit log.
The important limit is availability. Lightspeed is introducing these capabilities to selected customers through an opt-in beta, with wider access promised later. The company describes the workflow, but has not published an independent measure of time saved or sales gained.
Lightspeed has added an action layer to its retail assistant. A merchant can ask about stock or sales, then have the system prepare a product update, clearance tag, purchase order or inventory transfer inside the same conversation. The proposed change is shown before execution and waits for explicit approval; completed actions are recorded in an audit log.
The important limit is availability. Lightspeed is introducing these capabilities to selected customers through an opt-in beta, with wider access promised later. The company describes the workflow, but has not published an independent measure of time saved or sales gained.
An AI audience builder that stops before activation
Rokt’s mParticle Agent turns a campaign goal into a draft customer audience. A marketer can ask for lapsed buyers, inspect which purchase-history signals the system used, change the inactivity window and exclude recent purchasers. The tool explains its proposed logic so the team can check whether the available data actually supports the segment.
The workflow keeps two decisions separate. Saving the audience does not send it to an advertising or messaging platform; a person chooses the destination and whether to add reach-expansion tools afterward. Rokt suggests starting with a measurable case such as repeat purchase or re-engagement. It has not published an independent result for revenue lift or time saved.
Rokt’s mParticle Agent turns a campaign goal into a draft customer audience. A marketer can ask for lapsed buyers, inspect which purchase-history signals the system used, change the inactivity window and exclude recent purchasers. The tool explains its proposed logic so the team can check whether the available data actually supports the segment.
The workflow keeps two decisions separate. Saving the audience does not send it to an advertising or messaging platform; a person chooses the destination and whether to add reach-expansion tools afterward. Rokt suggests starting with a measurable case such as repeat purchase or re-engagement. It has not published an independent result for revenue lift or time saved.
❤1
Samsara turns route data into an early compliance warning
Samsara has made Agent Studio, its no-code AI agent builder, generally available. One documented workflow combines a driver’s remaining stops, available driving time and shift time. The agent flags a possible hours-of-service breach three hours before it occurs, giving the fleet team time to adjust the route or assignment.
The workflow is useful to commerce operations because delivery risk is evaluated before the final stop fails, not after a late order appears in a report. Samsara says S. Bertram Foods reduced weekly manual review time by 84%, but the figure comes from the vendor and its customer; the reporting does not provide the full measurement method. The product is live for Samsara customers, while evidence of broad production use remains limited.
Samsara has made Agent Studio, its no-code AI agent builder, generally available. One documented workflow combines a driver’s remaining stops, available driving time and shift time. The agent flags a possible hours-of-service breach three hours before it occurs, giving the fleet team time to adjust the route or assignment.
The workflow is useful to commerce operations because delivery risk is evaluated before the final stop fails, not after a late order appears in a report. Samsara says S. Bertram Foods reduced weekly manual review time by 84%, but the figure comes from the vendor and its customer; the reporting does not provide the full measurement method. The product is live for Samsara customers, while evidence of broad production use remains limited.
Constructor carries an agent conversation through payment
Constructor has launched Agentic Checkout, a Stripe-powered flow that keeps product discovery and payment inside one conversational interface. A shopper can describe a need, answer follow-up questions, compare recommendations, refine the choice and then pay with Link or another supported method without opening a separate checkout page.
The workflow removes one handoff, but it also shifts the control point. Product attributes, availability and payment confirmation have to remain consistent while the conversation changes the shortlist. Constructor says the feature is live; the announcement does not provide an independent result for conversion, abandonment or payment errors. Merchants therefore need to test whether a completed recommendation still matches the item and terms shown at authorization.
Constructor has launched Agentic Checkout, a Stripe-powered flow that keeps product discovery and payment inside one conversational interface. A shopper can describe a need, answer follow-up questions, compare recommendations, refine the choice and then pay with Link or another supported method without opening a separate checkout page.
The workflow removes one handoff, but it also shifts the control point. Product attributes, availability and payment confirmation have to remain consistent while the conversation changes the shortlist. Constructor says the feature is live; the announcement does not provide an independent result for conversion, abandonment or payment errors. Merchants therefore need to test whether a completed recommendation still matches the item and terms shown at authorization.
❤1
A campaign brief becomes an inactive discount draft
commercetools has detailed a Promotions Agent workflow inside Merchant Center. A merchant can paste a campaign brief or describe an offer in plain language. The agent identifies the discount type, products, categories, customer groups and conditions, then asks for missing details such as priority or validity dates.
The useful boundary comes at the end. The system can prepare cart discounts, product discounts or codes, including bundles, shipping offers and localized descriptions, but saving does not make them live. The result remains inactive until a person reviews and activates it. Teams can also set instructions for naming, stacking and maximum percentages. The company has not published independent evidence of fewer pricing errors, so the concrete gain is reduced configuration work, not a proven sales lift.
commercetools has detailed a Promotions Agent workflow inside Merchant Center. A merchant can paste a campaign brief or describe an offer in plain language. The agent identifies the discount type, products, categories, customer groups and conditions, then asks for missing details such as priority or validity dates.
The useful boundary comes at the end. The system can prepare cart discounts, product discounts or codes, including bundles, shipping offers and localized descriptions, but saving does not make them live. The result remains inactive until a person reviews and activates it. Teams can also set instructions for naming, stacking and maximum percentages. The company has not published independent evidence of fewer pricing errors, so the concrete gain is reduced configuration work, not a proven sales lift.
Electronic shelf labels are not personalized pricing
US retailers are expanding electronic shelf labels while AI tools estimate demand response. AP separates three workflows often blurred together: changing a price for everyone, updating a shelf display electronically, and offering a price based on personal data.
Walmart says its labels keep prices current and do not set a different price from a shopper’s face or phone. Relex Solutions told AP that retailers usually update electronic labels on a schedule, often weekly; intraday surge pricing remains rare.
For a pricing rollout, log the input category for every change — demand, inventory, competitor price or customer data — and keep the displayed shelf price synchronized with checkout. Personal data requires a separate consent and compliance review; the label hardware itself does not prove personalization.
US retailers are expanding electronic shelf labels while AI tools estimate demand response. AP separates three workflows often blurred together: changing a price for everyone, updating a shelf display electronically, and offering a price based on personal data.
Walmart says its labels keep prices current and do not set a different price from a shopper’s face or phone. Relex Solutions told AP that retailers usually update electronic labels on a schedule, often weekly; intraday surge pricing remains rare.
For a pricing rollout, log the input category for every change — demand, inventory, competitor price or customer data — and keep the displayed shelf price synchronized with checkout. Personal data requires a separate consent and compliance review; the label hardware itself does not prove personalization.
❤1
Hello Retail lets merchants stage a change from a chat prompt
Hello Retail is demonstrating a beta connection between its Product Intelligence system and Claude or ChatGPT. An ecommerce manager can ask a plain-language question such as which site searches returned no results, instead of building a report first. The same connection can prepare a merchandising action—for example, moving selected products to the top of a bestsellers page.
The control point is explicit: Hello Retail says the system checks with the user before making a requested change. That separates catalog analysis from a live storefront edit and gives the merchant a review step. Product Intelligence through MCP is currently a beta, not a general rollout, and the company has not published an independent result for conversion or time saved.
Hello Retail is demonstrating a beta connection between its Product Intelligence system and Claude or ChatGPT. An ecommerce manager can ask a plain-language question such as which site searches returned no results, instead of building a report first. The same connection can prepare a merchandising action—for example, moving selected products to the top of a bestsellers page.
The control point is explicit: Hello Retail says the system checks with the user before making a requested change. That separates catalog analysis from a live storefront edit and gives the merchant a review step. Product Intelligence through MCP is currently a beta, not a general rollout, and the company has not published an independent result for conversion or time saved.
❤1
RELEX keeps a human review step in AI-assisted pricing
Retailers are using AI-assisted pricing tools to process sales history, competitors' prices and other signals, then forecast demand and estimate how a price change may affect sales and profit. RELEX Solutions told AP that roughly half of the US price changes it recorded were decreases, a useful check on the assumption that automated pricing only pushes prices up.
The workflow is still controlled by retail rules. Many stores update electronic shelf prices on a set schedule, often weekly, rather than changing them continuously for each shopper. A pricing team reviews the recommendation, chooses the permitted change and synchronizes it with shelf labels. The system predicts a response; it does not prove how customers will react.
Retailers are using AI-assisted pricing tools to process sales history, competitors' prices and other signals, then forecast demand and estimate how a price change may affect sales and profit. RELEX Solutions told AP that roughly half of the US price changes it recorded were decreases, a useful check on the assumption that automated pricing only pushes prices up.
The workflow is still controlled by retail rules. Many stores update electronic shelf prices on a set schedule, often weekly, rather than changing them continuously for each shopper. A pricing team reviews the recommendation, chooses the permitted change and synchronizes it with shelf labels. The system predicts a response; it does not prove how customers will react.
Guide: test the shopping task before choosing the model
AWS’s Agentic Shopping Assistant is a starting architecture for retailers building conversational product discovery. It connects a retailer’s catalog and customer context to an assistant that can answer questions, compare items and recommend products. The implementation is customized rather than switched on as a universal storefront feature.
A useful pilot should begin with one bounded task, such as comparing two products from verified catalog attributes. Log the source fields used in each answer, require a handoff when stock or policy data is missing, and measure completed product-page visits or qualified carts—not chat volume alone.
The architecture does not remove the retailer’s responsibility for catalog quality, permissions or testing. A persuasive answer built on stale availability is still a failed shopping workflow.
AWS’s Agentic Shopping Assistant is a starting architecture for retailers building conversational product discovery. It connects a retailer’s catalog and customer context to an assistant that can answer questions, compare items and recommend products. The implementation is customized rather than switched on as a universal storefront feature.
A useful pilot should begin with one bounded task, such as comparing two products from verified catalog attributes. Log the source fields used in each answer, require a handoff when stock or policy data is missing, and measure completed product-page visits or qualified carts—not chat volume alone.
The architecture does not remove the retailer’s responsibility for catalog quality, permissions or testing. A persuasive answer built on stale availability is still a failed shopping workflow.
Square puts spoken ordering into a self-service kiosk
Square has introduced Kiosk, a restaurant ordering station that combines a touch display, payments and a conversational voice layer. The system is designed to understand the restaurant's menu, listen to a customer and respond while the order is being assembled.
The concrete workflow starts with a spoken request rather than a sequence of menu taps: the kiosk maps the request to available items, presents the resulting order and keeps payment in the same station. That makes menu accuracy the critical control. Item names, modifiers, availability and prices must remain consistent between the voice layer, the screen and the point-of-sale record. Square has announced the product, but the source provides no independent result for order speed, accuracy or sales.
Square has introduced Kiosk, a restaurant ordering station that combines a touch display, payments and a conversational voice layer. The system is designed to understand the restaurant's menu, listen to a customer and respond while the order is being assembled.
The concrete workflow starts with a spoken request rather than a sequence of menu taps: the kiosk maps the request to available items, presents the resulting order and keeps payment in the same station. That makes menu accuracy the critical control. Item names, modifiers, availability and prices must remain consistent between the voice layer, the screen and the point-of-sale record. Square has announced the product, but the source provides no independent result for order speed, accuracy or sales.
Motive links fleet alerts to maintenance actions
Motive has released Agents for fleet operations. A company can set a rule so that a vehicle fault code automatically opens a work order in Motive Maintenance, marks the truck out of service and alerts the responsible team. The same rule model can cover idling, fatigue and other safety events.
Available actions include email, text, app and in-cab notifications; automated driver calls are planned for later in 2026. The fleet chooses both the trigger and the response, so the system does not invent its own maintenance policy.
Bennett Family of Companies started with a narrower setup: office teams receive event alerts, while driver-facing actions remain disabled. That makes the rollout testable before more authority is automated.
Motive has released Agents for fleet operations. A company can set a rule so that a vehicle fault code automatically opens a work order in Motive Maintenance, marks the truck out of service and alerts the responsible team. The same rule model can cover idling, fatigue and other safety events.
Available actions include email, text, app and in-cab notifications; automated driver calls are planned for later in 2026. The fleet chooses both the trigger and the response, so the system does not invent its own maintenance policy.
Bennett Family of Companies started with a narrower setup: office teams receive event alerts, while driver-facing actions remain disabled. That makes the rollout testable before more authority is automated.
Guide: let AI draft the automation, not activate it
Shopify Flow can use Sidekick to turn a plain-language instruction into an automation. A merchant can describe a trigger, condition and action—such as tagging a small order—and Sidekick builds the workflow elements for review.
The useful control is built into the process: generated workflows start inactive. The merchant must open the draft, check whether business-specific exceptions are missing, test it, and manually turn it on. Sidekick can revise the draft through follow-up instructions, but Shopify warns that complex logic may still be incomplete.
Start with a reversible task, test expected and edge-case orders, and record false triggers before automating refunds, customer messages or inventory changes.
Shopify Flow can use Sidekick to turn a plain-language instruction into an automation. A merchant can describe a trigger, condition and action—such as tagging a small order—and Sidekick builds the workflow elements for review.
The useful control is built into the process: generated workflows start inactive. The merchant must open the draft, check whether business-specific exceptions are missing, test it, and manually turn it on. Sidekick can revise the draft through follow-up instructions, but Shopify warns that complex logic may still be incomplete.
Start with a reversible task, test expected and edge-case orders, and record false triggers before automating refunds, customer messages or inventory changes.
❤1