Agentic Commerce Fulfilment: What Happens After an AI Agent Presses Buy?
Your warehouse has been volunteered for a miracle. The confirmation email has already gone out.
Imagine the order arriving at 1:58pm, paid in full, wearing a next-day delivery promise and carrying the absolute confidence of something that has never had to find its own stock. The AI shopper has compared the options and selected yours. Congratulations. Somewhere behind the checkout, your last available unit has apparently developed reproductive capabilities, because two sales channels have promised it to different people and neither intends to surrender custody.
The WMS sends a picker to bin D14. D14 contains a cable tie, a damaged carton and enough empty space to accommodate the entire ecommerce strategy. Replenishment insists the stock is coming. Dispatch insists the truck is leaving. The order confirmation insists the customer has made an excellent choice. Three versions of reality have entered the building, and yours is the only one wearing safety boots.
Then the second item turns out to be in another fulfilment centre. The basket splits, sprouts another freight charge and starts eating its own margin. Someone suggests upgrading the service, as though “express” can reach backwards through time and collect a parcel nobody has packed. Customer Service asks whether delivery is still on track. You look at the empty pick face, the closing collection window and the email announcing how seamless the new shopping experience is.
There it is. That familiar little pulse behind the eye.
Because you already know where this goes. You will find the stock, interrupt another pick, negotiate a collection, approve something expensive and drag the order over the line using favours the operating model does not officially acknowledge. Tomorrow, the dashboard will record another successful delivery. The extra handling will disappear into labour. The upgrade will disappear into freight. Your afternoon will disappear completely.
And the meeting will conclude that the new channel is performing well.
That is the expensive trick: a supply chain can keep the customer’s promise by quietly breaking its own economics.
Agentic commerce fulfilment is where the beautifully frictionless purchase arrives to collect on everything your business said it could do. Available stock. Feasible dispatch. A delivery date with some connection to the working calendar. The agent has accepted the offer. Now your allocation rules, warehouse queues and carrier arrangements have to produce it, preferably without your operations manager personally holding the whole transaction together by the collar.
The trouble starts when each system receives a smaller version of the obligation. The OMS gets the order lines. The warehouse gets the pick tasks. The carrier gets the consignment. Somewhere between those handovers, the reason the customer chose you gets stripped off like packaging nobody thought they needed.
Everybody completes their bit. Nobody notices that the original promise is now wandering through the network without adult supervision.
You can survive that through intervention. You probably already do. But every rescued order teaches the business a dangerous lesson if nobody records what the rescue cost: keep selling this promise. Apparently, we can do it.
Until the afternoon you cannot.
Then the same operation that relied on your improvisation wants to know why you failed to follow the process.
TL;DR
The AI found it, compared it, bought it and paid for it in seconds. Marvellous. Now somebody has to pick the bloody thing. Agentic fulfilment is where pristine digital intent meets inventory allocation, warehouse queues, carrier cut-offs, split shipments and the ancient operational tradition of “who changed that rule?”. The dangerous orders are the ones you rescue so successfully that nobody fixes what went wrong. If the dashboard records the delivery and loses the rescue bill, failure has just supplied its own business case for expansion. The interface simplified the purchase. It did not simplify the physics.
Agentic fulfilment starts where the exciting AI demo ends
Agentic fulfilment is the physical and digital execution required after an AI shopping agent completes or assists with a purchase. It connects the order created through agentic commerce to inventory allocation, order management, warehouse execution, carrier selection, tracking, delivery, exceptions and post-purchase updates. The central challenge is not placing the order. It is preserving the promise as that order moves through multiple systems and providers. That last sentence is the whole article.
The AI presses Buy and somewhere in the warehouse a thermal printer starts laying eggs. Out comes a pick ticket for stock that exists numerically, a next-day label for a collection that exists historically, and a packing instruction requiring three items currently enjoying separate lives across two fulfilment centres. The order has been alive for eleven seconds and already needs transport, counselling and a small injection of shareholder funds.
Picture the warehouse trying to make it real. A picker excavates the supposedly available stock and finds a display model with one foot missing. Replenishment offers a sealed carton containing the wrong colour with tremendous confidence. The OMS spots another unit interstate and quietly splits the shipment, whereupon the original freight allowance becomes a touching little memorial to what everybody thought this would cost.
Then someone marks it urgent. Urgent reproduces. Soon the replacement pick, the stock transfer and the second parcel are all urgent, three descendants of one purchase fighting over the same collection while the original order sits in the dashboard looking remarkably uncomplicated.
It cost the business money to teach itself that this was profitable. Now somebody wants to scale it.
Agentic commerce has attracted enormous attention because the purchasing interface is changing. Rightly so.
McKinsey estimates AI agents could mediate between $3 trillion and $5 trillion of global consumer commerce by 2030 under moderate scenarios. It also argues that as agents take on more purchasing activity, retailers with clean inventory data, predictable fulfilment and reliable service become more attractive default suppliers.
Big number. Interesting future. Now someone has to pick the order.
ACP makes that ownership brutally clear. The agent manages the checkout interaction. The merchant remains the merchant of record. And the merchant still owns fulfilment after successful payment.
The ACP seller documentation leaves fulfilment with the merchant. Every offspring of that checkout is yours to feed.
Microsoft's Copilot Checkout uses the same basic architecture: the AI experience may originate the transaction, but the merchant's existing systems still handle shipping and order fulfilment. The interface changed. Gravity did not.
Someone still has to make reality agree with the transaction.
The customer outsourced the shopping. Your warehouse inherited the homework.
Checkout removed the friction. Dispatch would like to know why it is holding all of it.
“Buy these three things and have them here by Friday.” Behind checkout, imagine the basket hitting the OMS and hatching into three consignments. One immediately demands express. Another cannot travel until replenishment locates its legs. The third has been allocated to a warehouse that technically holds stock, although the stock is currently trapped behind a promotional display nobody has authority to move because it belongs to Marketing.
Now the pick tickets start recruiting. A supervisor leaves receiving. A picker abandons a replenishment run. Someone borrows the only suitable carton from an order that was foolish enough to arrive without a crisis attached. That order becomes a crisis. It borrows a picker. The original basket has developed a small workforce and is now producing secondary emergencies faster than Dispatch can assign them fluorescent stickers.
On the customer’s screen, three ticks. Effortless.
The interface simplified the intent. It did not simplify the physics.
There is commercial pressure behind those ticks. Baymard Institute’s published 2026 checkout data lists excessive additional costs at 40% and slow delivery at 20% among reported abandonment reasons after setting aside “just browsing”. Those are human shopper responses, not AI-agent results. Shipping cost and timing were already capable of killing the sale. Delegating checkout does not make them somebody else’s concern.
So the merchant sells the attractive combination. Then Operations discovers it can preserve the price, the deadline or the afternoon’s original workload, but keeping all three requires a favour from someone currently ignoring its calls.
This is friction displacement: the unresolved work has moved out of the buying journey and into fulfilment. Count it in extra touches, split charges and orders shoved backwards to make this one look easy.
Somewhere near packing, the order that donated its carton is now late. Its customer asks why.
Apparently, there has been an unexpected delay. The delay is neither unexpected nor particularly mysterious. It is helping fulfil somebody else’s frictionless purchase.
The machine can accept the order instantly. Your warehouse cannot download another shift.
The website accepts another order without breaking eye contact. Somewhere downstairs, the pick queue starts developing vertebrae. It curls around replenishment, swallows the afternoon wave and sheds a fresh layer of URGENT labels. The dispatch supervisor opens another pack bench by relocating a mountain of returns, which immediately becomes someone else’s mountain of returns. Capacity has apparently increased. Floor space would like to dispute the announcement.
Warehouses already know this creature. Black Friday feeds it. Promotions breed it. An influencer mentions one product and suddenly a perfectly ordinary pick face becomes a place people speak about in the past tense. AI-assisted shopping adds another route into that same physical operation, where buying decisions can become orders without creating a single extra minute before collection.
The channel is growing. Adobe reported that traffic from generative AI tools to US retail sites increased 693.4% year on year during the November–December 2025 holiday season. Those were AI-referred visits, not a count of autonomous purchases or proof of equivalent warehouse-volume growth. I cannot confirm that AI shopping will produce sudden fulfilment surges. The figure establishes that the route into the shop is changing; it does not tell you how many pickers to roster.
The operational exposure is the timing. Take a deliberately simple illustration: 300 orders released into an operation able to complete 100 comparable orders an hour represent three hours of work: 300 ÷ 100 = 3. With two hours until collection, only 200 fit. The remaining 100 do not become dispatchable because the confirmation emails went out quickly.
They become tomorrow’s backlog, tonight’s overtime or another request to hold the truck. Pull people from receiving and tomorrow’s stock waits. Pull them from replenishment and the pick faces run dry. The queue starts feeding on the very work required to keep it moving.
That is the decision-speed mismatch worth testing: how much work can the sales channel commit before fulfilment changes what it is willing to promise?
Otherwise, the website keeps selling today while the warehouse is already working tomorrow, and the missing shift exists entirely inside the supervisor’s blood pressure.
The OMS is about to receive orders with more opinions
Imagine Wednesday delivery entering your OMS and being mugged by a routing rule from 2019. The rule takes its deadline, replaces it with STANDARD and sends it to a warehouse that dispatches on Thursday. Product correct. Quantity correct. Address correct.
Wednesday is lying behind the integration wearing one shoe, while the order sails through validation with an immaculate record.
The customer paid for an outcome. Your systems have started processing a collection of ingredients.
The date matters. Baymard’s June 2023 benchmark found that 41% of sites did not provide an estimated delivery date, leaving shoppers to interpret shipping speeds and work out arrival themselves. That is a historical checkout finding, not a measure of today’s agentic orders. It exposes the ambiguity already hiding inside labels such as “express” and “two business days”.
Now give that ambiguity a warehouse login. EXPRESS believes it means leave immediately. The WMS believes it means join the next wave wearing a more expensive label. The carrier starts counting after collection. Customer Service starts counting backwards from the complaint. One service code has developed four personalities, and every one of them can produce documentation.
This is promise drift. The accepted order survives, but the conditions attached to it get stripped away during execution. Eventually someone discovers that preserving Wednesday requires an upgrade. Finance asks who approved Wednesday. Ecommerce produces the confirmation. Operations produces the routing rule. The parcel sits between them accruing a backstory.
The merchant may never receive the agent’s entire shopping rationale. It can preserve the delivery commitment, selected service and other terms it actually accepted. Those conditions need authority over allocation, release and routing. A delivery date buried in an unused field cannot stop a bad decision.
Otherwise, you have given a rule nobody remembers writing permission to renegotiate a promise the customer has already paid for. And it negotiates exclusively by disappointing them.
Your warehouse cut-off just became an ecommerce API dependency
At 2:01pm, SAME-DAY DISPATCH is still above checkout making promises with your name on them. Inside the warehouse, the manifest has closed, the driver has collected his paperwork, and today’s remaining capacity is disappearing through the gate in a vehicle your website apparently believes can be summoned by positive thinking.
Imagine the next order landing. Paid. Confirmed. Due tomorrow. It presents itself to Dispatch with the serene entitlement of something born after the consequences.
Someone calls the driver back. He agrees, this once. A picker gets pulled off replenishment, a supervisor packs the order, and another delivery promise survives on borrowed time. The website notices none of this. It accepts another order. Your favour has just become a fulfilment option available to the general public.
The clock is less accommodating. In an illustrative operation with a 2:00pm collection and 45 minutes of remaining pick, pack and staging work, the latest feasible start is 1:15pm: 2:00pm minus 45 minutes. That assumes no queue or disruption. Keep advertising the promise until 2:00pm and you are selling execution time that has already expired.
The pressure to leave that option visible is real: 20% cited slow delivery in Baymard’s published 2026 checkout-abandonment reasons, after setting aside “just browsing”. That is human shopper research, not an AI-agent failure rate. It helps explain why a business might hesitate to move the date. It does not give the warehouse another afternoon.
A usable dispatch slot expires. So does the labour window before it. Fulfilment availability needs to change when those opportunities close, even if the carrier still lists the service and the warehouse still contains forklifts.
Otherwise, the operation keeps rescuing an offer that should have been withdrawn. Until the driver says no.
Then the website keeps its promise in writing, the customer keeps the screenshot, and you get asked why the warehouse let them down.
Same-day dispatch has a shelf life. The banner thinks it is immortal.
A carrier offering a next-day service does not mean an order can still reach that service today. First it needs release, replenishment if required, picking, packing and handover. Those dependencies do not shorten because the buyer used an AI agent. A shopping decision can be completed while the product is still several physical tasks away from being dispatchable.
The last safe time to sell a delivery promise is earlier than the last possible time to rescue it.
Confuse those two and the recovery window becomes normal operating capacity. The supervisor who occasionally persuaded a driver to wait is now an undocumented extension of the carrier contract. Their personal phone number has become infrastructure.
A fixed sales cut-off can work where the operation has sufficient protection built around it.
Where workload varies, the offer needs to reflect what the fulfilment location can still execute. That might mean withdrawing a service, moving the date or choosing another feasible location before payment locks in the expectation.
The hard conversation is whether the business will accept losing an order it cannot fulfil sensibly. Avoid that conversation and Operations gets a different one later: why did freight cost so much this week?
“Processing” is a very small word for this much trouble
Imagine an urgent order waiting behind a replenishment task that is waiting for a forklift that has been borrowed to clear a staging area occupied by freight awaiting collection. The dependency chain has curled around and started chewing its own tail. On the customer’s screen: processing.
Inside the queue, urgency is running a counterfeit passport operation. Every department has stamped something PRIORITY. The oldest order feels entitled to go first. The premium order believes it purchased nobility. A replacement order arrives carrying the personal endorsement of someone senior, and the pick sequence quietly becomes a contest in who can generate the most uncomfortable phone call.
This is how a warehouse can be working flat out while the most endangered promise receives no useful attention. Activity fills the building. The wrong work consumes the remaining time.
Order age alone cannot settle the sequence. The operation needs to connect remaining work with the latest feasible handover and the consequences of missing it. An order received later may need action sooner because its viable route closes first.
Nor does every delayed task justify a red alert. If everything is an exception, escalation becomes warehouse wallpaper. The useful intervention identifies which commitment is becoming unrecoverable, what would preserve it and who can authorise that action.
“Processing” can remain perfectly acceptable customer language while the work is under control. It becomes an expensive euphemism when nobody underneath it can say whether the order still has a route to the promised date.
Your basket has split. So has everyone’s interest in the total cost.
One purchase enters the network. Three parcels emerge, each with its own label, freight charge and entirely reasonable explanation. Individually, they are model employees. Together, they have eaten the contribution margin and would like another carton.
Consider an illustrative order for a coffee machine, the compatible grinder and a starter kit, intended for a café opening. Different stock locations trigger separate shipments. Each location dispatches its allocated line successfully. The starter kit arrives first. The machine follows. The grinder misses the opening, leaving the buyer with an impressively documented inability to serve coffee.
The fulfilment network has completed several tasks. The customer is still missing the useful outcome.
Splitting may be the correct choice. Consolidation may delay everything. The point is to evaluate the basket before allowing line-level rules to decide its economics by accident. Extra handling, packaging, freight and the timing of dependent items belong in that decision.
The same applies when the carrier is chosen. A cheap service that cannot support the accepted terms was never a valid saving. A familiar carrier does not deserve the parcel simply because its code has occupied the default field since someone who no longer works here set it up.
This is where a parcel carrier scorecard earns its keep. Evidence about service, destination performance and recovery behaviour needs to influence routing. Otherwise the scorecard attends meetings while habit continues allocating the freight.
The parcel has left the system. Unfortunately, it is still in the building.
A label prints and the order becomes “shipped”. The carton remains under a bench, participating in international digital society without having travelled past the tape dispenser. Customer Service quotes the update. The customer rearranges their plans. A printer has just acquired the authority to describe a vehicle movement.
The distinction matters when order events travel back to a shopping platform. The ACP webhook documentation provides for merchant order updates intended to keep ChatGPT aligned with “fulfillment-grade truth”. Google’s UCP order lifecycle guidance likewise documents fulfilment events and post-purchase adjustments. www.agenticcommerce.dev
Neither mechanism makes an inaccurate source event more accurate. It gives that event somewhere else to go.
Our operational reading is that event definitions now deserve the same scrutiny as routing rules. Label created, packed, manifested and accepted by the carrier are different conditions. They support different conclusions about where the order is and what can still be done to it.
The missing event matters too. If the expected handover does not appear, the order should not retain unlimited confidence merely because nobody has sent a contradiction. Silence needs an owner and a threshold for investigation.
Otherwise, connecting the agent to the status feed simply gives the warehouse rumour a larger audience.
“Cancel it” reaches the screen before it reaches the picker
The customer changes their mind. A request is acknowledged. Somewhere on the floor, a picker scans the final item and pushes the tote towards packing. The digital transaction and the physical transaction have begun divorce proceedings without agreeing who gets the goods.
Where an agent-enabled service supports cancellation requests, the operation still needs to distinguish receipt of the request from completion of the cancellation. Those are separate events. Blurring them can produce a refund travelling one way while the product continues enthusiastically in the other.
There is a technical boundary worth respecting here. ACP documents cancellation of a checkout session; that does not establish that a completed order can be stopped after carrier handover. Post-purchase actions depend on the platform, merchant integration, permissions and actual fulfilment state.
Operationally, the question is how far work has progressed and whether it can still be interrupted. Before release, cancellation may be simple. During picking, someone may need to intercept work. After dispatch, the options depend on the carrier and service, and recovery may require a return.
The confirmation should follow what the operation achieved. Otherwise, “cancelled” becomes an optimistic opinion attached to a parcel that has not been consulted.
Returns close the same loop in reverse. Receiving the parcel, inspecting it, issuing a refund and restoring sellable inventory need their own evidence. If a refund automatically convinces the stock feed that the item is available, the next buyer can purchase yesterday’s unresolved problem before the returns bench has opened the box.
RETURNADO™ territory: the return has barely introduced itself and already has another customer.
The dashboard says success because nobody invited the rescue bill
Now comes the meeting. Conversion is up. Dispatch is green. Delivery landed inside the promise. Everyone has brought a chart, which makes the absence of the actual commercial outcome surprisingly difficult to notice.
The emergency upgrade sits inside freight. The additional pick sits inside labour. The replacement packaging sits inside consumables. Customer Service absorbs the chasing. Operations absorbs the interruptions. An order that required extraordinary intervention reappears in the monthly report dressed as evidence of ordinary capability.
Call that the rescue subsidy: the unplanned money and human effort that keep a weak fulfilment promise looking commercially viable.
The distinction matters because a business can mistake subsidised service for scalable service. Add volume on that basis and the forecast assumes the rescue continues at the same price, with the same people, forever. Eventually a person, collection window or approval limit refuses. The dashboard then records a sudden deterioration in performance, although the underlying dependency was there all along.
Delivery performance therefore needs to sit beside the cost of achieving it. Track unplanned upgrades, avoidable splits and repeat interventions against the order types and routes producing them. Include a view of when exceptions became visible relative to when recovery stopped being feasible.
You are looking for the orders that arrived on time but should never have needed that much saving. They reveal the operating model’s weaknesses before the customer complaint makes them official.
Somebody needs authority before the only option is expensive
The warehouse can see the deadline slipping but cannot approve the upgrade. Transport can quote the recovery but cannot authorise the spend. Ecommerce owns the promise but is unavailable. Finance would like supporting information. The last viable collection departs while an email acquires its fourth recipient.
That is a decision failure with a freight invoice attached.
Connecting systems cannot settle who may change a fulfilment route, absorb a recovery cost or withdraw an offer that has become infeasible. Those rights need to exist before the order forces the question. So do the limits and escalation paths.
Human intervention can be entirely sensible. It needs an owner, a response window and authority proportional to the decision. “Ask the 3PL” is useful only if the 3PL has the information and permission to act. Otherwise you have forwarded the uncertainty and started another clock.
This is the work an independent 4PL control role needs to perform: connect inventory, warehouse execution, freight, cost and service decisions around the accepted outcome. That requires agreed operating authority alongside visibility. A control tower that can only watch the collection leave is an observation deck.
Transport Works belongs in the gap where every provider can explain its own action and the business still cannot explain the result. The job is to close that gap before it becomes another afternoon of personally negotiated miracles.
Test the order that normally ruins someone’s lunch
The demonstration order will behave beautifully. One SKU, local stock, an empty queue and hours before collection. It has been raised in captivity and has no idea what your operation is like.
Use a controlled test that resembles the work you actually dread: several order lines, competing fulfilment locations, a meaningful deadline and a reason the cheapest route might be wrong. Capture the accepted terms, then follow what happens as allocation becomes warehouse work and warehouse work becomes freight.
Introduce a short pick or a missed release window. Watch the first response. Does the endangered promise become visible while there is still something useful to do? Does an authorised person receive a decision they can make, or an alert inviting them to begin their own investigation?
Test a cancellation during picking. Check what gets stopped, what remains reserved and what the customer is told. Follow a partial return far enough to see whether the financial and inventory records reconcile.
Keep a record of every intervention the standard flow needed. The person who knows which supervisor to call is valuable. Their involvement is also evidence of a dependency the test should expose.
If the demonstration passes because your best operator quietly repaired it, the demonstration has discovered that operator. It has not yet proved the system.
The AI has finished shopping. Now the supply chain has to stop improvising.
Agentic commerce is wonderfully good at making the front of ecommerce look cleaner.
Intent becomes structured. Products become comparable. Checkout becomes programmable. Payment becomes delegated. Order confirmed.
Then the warehouse opens the ticket.
And that is where the future has to survive the present.
The stock still has to exist.
The allocation still has to make sense.
The warehouse still has to have capacity.
The pick still has to be right.
The carton still has to close.
The carrier still has to deserve the parcel.
The exception still needs an owner.
The return still has to reconcile.
Nothing about agentic commerce eliminates the physical supply chain.
It removes more of the time and human tolerance that allowed weak supply chains to improvise around their contradictions.
That is why the winners will not be the businesses whose AI can buy fastest.
They will be the businesses whose fulfilment network can accept a machine-made promise without immediately beginning to renegotiate it behind the scenes.
The AI pressed Buy. The order is yours now.
Transport Works. Because Your Supply Chain Won’t Fix Itself.
Want to know what else happens after the AI presses Buy? Read:
Agentic fulfilment FAQs
What is agentic fulfilment?
Agentic fulfilment is the execution of orders created or influenced by AI shopping agents across inventory allocation, order management, warehousing, shipping, delivery and post-purchase updates. The defining challenge is maintaining the commercial conditions that caused the AI agent to select the offer while the order passes through physical fulfilment systems and providers.
Does an AI shopping agent fulfil the order?
Generally, no. Current agentic-commerce models allow an AI agent to support discovery and checkout, while the merchant remains responsible for inventory, pricing, payment processing and physical order fulfilment.
ACP explicitly assigns order fulfilment to the seller after payment and order creation. The AI can initiate the transaction. The merchant still has to deliver it.
How does agentic commerce affect warehouses?
Agentic commerce can increase the importance of accurate inventory, dynamic allocation, fulfilment capacity, cut-off logic and event data because these factors may influence the buying promise before checkout and need to remain accurate after purchase. The warehouse therefore becomes part of a faster commercial decision loop rather than simply receiving orders after the sale.
What systems are needed for agentic fulfilment?
Agentic fulfilment may involve ecommerce platforms, OMS, WMS, 3PL systems, TMS, carrier APIs and post-purchase services. The important requirement is not the number of systems. It is whether they share enough operational truth and decision ownership to preserve inventory, delivery and cost commitments throughout the order lifecycle. A fully connected stack can still fail if each system optimises a different answer.
What is the biggest risk in agentic fulfilment?
The biggest risk is promise drift: downstream fulfilment decisions stop matching the cost, delivery, availability or service conditions that caused the AI shopping agent to choose the order. The transaction can therefore remain technically successful while the commercial promise quietly deteriorates during execution.
The AI Pressed Buy. What’s Actually Running the Rest?
The Order Is Confirmed. The Hard Part Just Clocked In
INSIGHTS FROM DANYUL GLEESON, FOUNDER, CLUSTER-FREIGHT-FIXER & LOGISTICS CHAOS TAMER-IN-CHIEF AT TRANSPORT WORKS
Danyul Gleeson has spent 25+ years collecting the kind of supply chain scar tissue nobody puts on a capability statement.
He has seen the dashboards that looked magnificent while the operation underneath them quietly caught fire, the freight contracts that were apparently “great deals” until somebody read the surcharges, and the logistics problems everybody kept managing because nobody stopped long enough to ask why they existed in the first place.
As Founder of Transport Works, Danyul works across freight optimisation, 3PL management, warehousing and fulfilment, logistics technology, visibility, KPI reporting and supply chain performance across Australia, New Zealand and the USA.
He has a fairly low tolerance for supply chain theatre and a habit of finding the one loose bolt everybody else has been stepping over.
The Freight Files is where Danyul calls out what he is seeing now, what matters next and which logistics problems are quietly getting expensive while everyone else is still talking about them in meetings.
No recycled industry theatre. Just what’s changing, what’s getting expensive, and what’s worth fixing before the answer becomes obvious.
Sources & References
Agentic commerce growth and adoption
McKinsey & Company – Agentic Commerce: How AI Shopping Agents Can Change Retail Used to support the discussion around the growth of agentic commerce, including McKinsey’s estimate that AI agents could mediate $3 trillion to $5 trillion of global consumer commerce by 2030, and the increasing importance of reliable inventory and fulfilment performance. McKinsey & Company
Adobe – Holiday Shopping Season Drove a Record $257.8 Billion Online with Consumers Embracing Generative AI Tools Referenced for Adobe Analytics data showing traffic from generative AI tools to U.S. retail sites increased 693.4% year on year during the 2025 holiday shopping season, demonstrating how rapidly AI-assisted shopping behaviour is developing. Adobe Newsroom
Agentic commerce protocols and merchant responsibility
Agentic Commerce Protocol – Getting Started for Sellers Used to establish the merchant’s responsibilities within ACP, including inventory management, pricing, payment processing and physical order fulfilment after an AI-assisted checkout is completed. Agentic Commerce Protocol
Agentic Commerce Protocol – Agentic Checkout Webhooks API Referenced for ACP’s post-purchase order update architecture and the requirement for merchants to send order lifecycle events so AI shopping experiences can remain aligned with fulfilment status. Agentic Commerce Protocol
Microsoft Advertising – Agentic Commerce and Copilot Checkout Used to support the discussion around AI-native checkout architecture, including Microsoft’s position that merchants remain merchant of record while their existing payment, tax, fulfilment and reconciliation systems continue to execute the transaction. Microsoft Advertising
Post-purchase fulfilment and order lifecycle
Google for Developers – Universal Commerce Protocol: Order Lifecycle (Post-Purchase) Referenced for UCP requirements covering order status updates, fulfilment events, cancellations, returns, refunds and the communication of post-purchase order truth back to the commerce platform. Google for Developers
Google for Developers – Universal Commerce Protocol: Order Lifecycle Implementation Used to support the discussion around multi-item orders, split shipments, shipment events, cancellations, returns and refunds as agentic commerce transactions move from checkout into physical fulfilment. Google for Developers
Ecommerce delivery and checkout expectations
Baymard Institute – 50 Cart Abandonment Rate Statistics 2026
Used to support the discussion around the commercial pressure behind delivery promises, including Baymard’s 2026 findings that 40% of relevant abandonment responses cited excessive additional costs and 20% cited delivery being too slow, after excluding shoppers who were simply browsing. Baymard Institute
Inventory accuracy and operational truth
Nicole DeHoratius & Ananth Raman – Inventory Record Inaccuracy: An Empirical Analysis, Management Science
Referenced to demonstrate the long-established operational risk created when recorded inventory and physical inventory diverge. The study examined nearly 370,000 inventory records across 37 stores of one retailer and found 65% were inaccurate
Disclaimer:
The information in this blog is provided for general informational purposes only and is current as of the date of publication. Customs duties, charges, processes, policies, and rates are subject to change at any time without notice. We make no representations or warranties of any kind, express or implied, about the completeness, accuracy, reliability, suitability, or availability of the information contained in this article. You should not rely on this content as a substitute for official sources. For the most up-to-date and authoritative information, please consult the relevant government agencies, customs authorities, and reference websites directly. Ideas, interpretations, and opinions expressed here are subject to change as regulations, markets, and industry practices evolve. Transport Works and its authors accept no liability for any loss or damage whatsoever arising from reliance on the information in this blog.
