After the Order Is Placed: What Operations Teams Are Really Handing to AI Agents This Peak
We have written about agentic commerce a few times this year, and each time it was mostly about the agent that does the buying, most recently in August when we asked who owns the delivery promise when the shopper is a machine. This piece is about the other side, which gets less attention and which we think matters more to anyone running fulfilment and delivery this quarter. It is about agents doing the work that starts after the order is placed: spotting the late orders, working through carrier exceptions, deciding which orders a person needs to look at before they ship, moving a shipment to a different carrier, and telling the customer what is going on. It is unglamorous work and there is a great deal of it in November and December.
It is also what a lot of people will be pitched this week. DELIVER America is in Las Vegas on 7 and 8 October, we will be there, and the public programme has more sessions on AI in operations than on almost anything else apart from carriers. So we went through the programme, and through what has been published and announced over the last few weeks, to try to separate what is being said from what is being deployed, and then to set out what we think has to be in place underneath before anyone lets an agent touch a live order.
What you will hear on the floor
The DELIVER programme gives a fairly good picture of where the conversation is. On Wednesday there is a roundtable with Suresh Balachandran of AbbVie called "AI With Proof: What Actually Changed in Operations?", which promises to "move beyond the AI hype" and compare "what has genuinely changed, what remains stuck in pilot mode, and what it takes to redesign workflows successfully". The same afternoon there is a roundtable with Aniket Ghonge of Amazon on domain experts building their own AI systems, which will cover "the governance, skills and safeguards needed to turn frontline knowledge into scalable solutions", and Pipe17 has a session on modernising order management for agentic commerce that includes "making exception management proactive with AI". Linnworks, also on Wednesday, is speaking about automating the fulfilment decisions behind each order "using intelligent rules".
On Thursday Deborah Dull's keynote on the self-driving supply chain is about bringing together "data, AI, automation and human expertise", and Shippo has the session that is closest to the subject of this article. It is called "Less Work, Better Peak: 5 Jobs High-Volume Shippers Don't Need To Do Anymore", and its description is a good summary of the work we are talking about: "Someone guesses a delivery date. Someone decides which carrier to use, again. Someone hunts for an order at the pack bench. Someone answers 'where is my order' a few hundred times." The session covers "what it takes to turn each one over to machine learning and AI agents".
Reading those descriptions together, two things stand out to us. The vocabulary has moved from the shopper to the operation, which was not really the case a year ago. And several of the sessions carry a qualifier, whether that is proof, pilot mode, governance or safeguards. The organisers seem to expect a room that has already heard the pitch and would now like to see the evidence.
Why this work, and why this quarter
The volume is the first reason. Adobe's holiday forecast, published on 28 September, expects US online sales of $275.1 billion between 1 November and 31 December, up 6.7% on last year, with $47.5 billion of that in Cyber Week. The same forecast expects AI-referred traffic to US retail sites to rise 130%, which is the buying side of the agent story and the part that tends to get the headlines. The operations side is less visible but follows directly from the first number, because more orders in a short window means more orders that go wrong in a short window.
The carriers did well last peak, and the tail is still large. ShipMatrix's figures for December 2025, as reported by FreightWaves on 12 January, put on-time delivery at 97.2% for UPS, 95.3% for FedEx and 94.1% for USPS, all better than the year before. Those are good numbers, and even on ShipMatrix's measure, which leaves out weather delays and allows ground parcels an extra day, they still mean that somewhere between roughly three and six parcels in every hundred arrived late in the busiest month of the year, depending on the carrier. Across a retailer's peak volume that is a lot of individual orders that someone has to notice, make a decision about and explain to a customer.
The customer, meanwhile, expects to be told. The AlixPartners 2026 Home Delivery Survey, as reported by FreightWaves in July, found that 90% of consumers want to be notified when a delivery is running late and given an updated arrival time, and that 88% say a late delivery that is not remedied weakens their future purchases with the retailer.
So the work is real, it is repetitive and time-sensitive, most of it is about reading status data and applying a policy the business already has, and it lands on the same small team every year. We think that is a fair case for handing some of it to software that can read and act, and probably a stronger case than most of the uses agents are being proposed for.
What has actually shipped
It is worth being specific about what vendors have released, as opposed to what is on a slide.
project44 announced a portfolio of six agents on 8 April, one of which is an Exceptions Management Agent for detecting and resolving shipment exceptions such as missed pickups and delivery failures. The company said in the same release that agent-driven interactions on its platform had grown from 500 a week to more than 30,000. Manhattan Associates announced Manhattan Marketplace on 19 May, a place for customers and partners to find and deploy agents inside its warehouse, transportation and order management products; the release says those agents inherit "the same deterministic spine and operational guardrails as our core software", and that customers will get access "in upcoming quarters".
There is a pattern in these and in the other announcements we have read, as far as we can tell. What has shipped is narrow, usually one agent for one job. It sits inside a platform that already holds the data and the rules for that job, and the vendors take care to say that the agent works within the platform's existing controls. A fair amount of it is also still described in the future tense.
How far people say they will let an agent go
The surveys from this year are fairly consistent with each other, and they describe an industry that is interested and cautious at the same time.
RELEX's State of Supply Chain 2026 report, based on a survey of 514 retail, manufacturing and supply chain leaders in January and reported by DC Velocity on 25 March, found that 67% were more confident in using AI for supply chain decisions than a year earlier and that 71% plan to invest in generative and agentic AI over the next three to five years. In the same survey only 10% said they would trust AI to make fully independent supply chain decisions, and 54% said they prefer AI to make recommendations while people finalise the decision.
An Ortec survey of 400 North American transportation, logistics and supply chain executives, reported by DC Velocity on 15 January, found that 42% were not yet exploring agentic AI at all and 23% planned to pilot it within twelve months. The concerns they named were integration cost (32%), a lack of explainability in the models (26%) and poor data quality (22%). BCG's survey of 30 large global logistics companies, published on 24 August, found that 97% of executives rank AI as a strategic priority and only 13% say it is delivering measurable financial impact, and it names fragmented data in legacy systems as one of the main reasons. And Gartner's prediction from June 2025, which is still the one most often quoted, is that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of "escalating costs, unclear business value or inadequate risk controls".
Our reading of all this is that almost everyone intends to use agents and very few intend to let them act alone. The middle ground, where the agent does the reading and proposes or takes a limited action and a person approves anything that matters, seems to be where most operations teams are heading. We think that is the right instinct.
What has to be true underneath
If that middle ground is the destination, then in our experience four things have to be in place before an agent goes anywhere near a live order.
The data has to be normalised before the agent arrives. Salesforce's study of 2,025 agentic AI decision-makers, fielded in May and published on 27 August, found that only 31% of those that had deployed agents had fully unified their data first, and that those who had unified the relevant data first reached a return in 7.3 months against 8.8 months for those who deployed first and fixed the data afterwards. PYMNTS, reporting on 21 September on an IDC paper that predicts 45% of the world's 2,000 largest companies will have adopted agent-driven supply chain orchestration by 2029, put the precondition plainly, which is that agents cannot scale without clean, governed and interoperable data. In post-purchase terms this is quite concrete. An agent deciding whether an order is late needs "late" to mean the same thing whichever carrier has the parcel, and every carrier reports status in its own vocabulary. If nobody has translated those into one model, the agent will get some of these wrong and will have no way of knowing that it has.
The rules have to exist already, and the agent has to be able to read them. An agent that moves a shipment to another carrier should be choosing among options the business has already said are acceptable for that parcel. An agent that decides an order needs a second look should be applying the business's own definition of what needs a second look. If those rules live only in the heads of the operations team, the agent has nothing to stay inside. Manhattan's phrase about a "deterministic spine" is a reasonable description of what is needed, and it is worth noticing that Linnworks is at DELIVER talking about rules rather than agents.
The agent's reach has to be limited by something other than its own judgement. That means it acts as a named user with a role, it can only use the tools that role allows, and what it does is logged in the same place and the same way as what a person does. The Salesforce study found that organisations with below-average governance were nearly twice as likely to discover an agent operating outside its parameters only after a consequential error, 32% against 18%.
There has to be a place where a person approves. The 54% in the RELEX survey who want AI to recommend and a person to decide are describing a workflow, and in practice that workflow needs a queue of some kind, where an order or an action waits with an owner, a priority and a deadline, where the agent can put things and prepare the case, and where a person releases or rejects. Gartner listing "decision governance" among its top supply chain technology trends for 2026, in a release dated 30 June, suggests the analysts see it the same way. We would add one thing about what the agent says to the customer while an order is waiting. We do not think an agent should promise a new delivery date on the strength of a doubtful carrier scan, and it is better for it to say the order is being looked into and to come back when there is something reliable to say.
Where we would start this peak
For a team that wants to use agents this season rather than next, we would start with the jobs where the agent reads far more than it writes. That means things like a morning review of late and at-risk shipments, a diagnosis of why a particular shipment is stuck, a first pass over the orders that are waiting for review, or a summary for the operations lead of what changed overnight. These are useful on the first day and very little can go wrong. We would let an agent act on its own only where a wrong action is cheap to undo, and we would not introduce a new write permission in the last week of November.
For anyone listening to a pitch this week, the four points above turn into questions that are worth asking. What data does the agent see, and who normalised it? Which of our rules can it read, and can it change them? Whose permissions does it act with, and where is the log? And where, exactly, does a person approve?
Where Carriyo fits
Most of what we have built sits underneath those four points. Carriyo holds normalised status data across 140+ carriers, the shipping costs and the rules a merchant has set up for routing and for review, and a native MCP layer through which an agent works with the same permissions and the same audit trail as the person who authorised it. In our next article we will go through how that works on Carriyo today, including how an agent puts an order on hold for a person to review and what it can and cannot do after that.
We are at DELIVER America in Las Vegas on 7 and 8 October. If you are there, come and find us, and we would be glad to talk through where agents would and would not be safe in your own operation. Otherwise, get in touch with our team at carriyo.com/contact.
Carriyo is The Intelligent Commerce Platform, from checkout to doorstep.
---
Sources
1. DELIVER America 2026. Programme and agenda, 7–8 October 2026, Las Vegas. https://www.deliver.events/america/program 2. Adobe. "Adobe: U.S. Holiday Shopping Season to Hit Record $275.1 Billion Online, Rising 6.7% YoY", 28 September 2026. https://news.adobe.com/news/2026/09/adobe-us-holiday-shopping-season-to-hit-record 3. FreightWaves. "Large parcel carriers improved on-time delivery during 2025 peak season" (citing ShipMatrix), 12 January 2026. https://www.freightwaves.com/news/large-parcel-carriers-improved-on-time-delivery-during-2025-peak-season 4. FreightWaves. "Carrier diversification unravels the last-mile delivery duopoly" (citing the AlixPartners 2026 Home Delivery Survey, June 2026), 29 July 2026. https://www.freightwaves.com/news/carrier-diversification-unravels-the-last-mile-delivery-duopoly 5. project44. "project44 launches AI agent portfolio at decision44, built on a decade of context, skills, and orchestration", 8 April 2026. https://www.project44.com/press-releases/project44-launches-ai-agent-portfolio-at-decision44-built-on-a-decade-of-context-skills-and-orchestration/ 6. Manhattan Associates. "Manhattan Associates Unveils Manhattan Marketplace, a Shared Engine for Supply Chain and Commerce AI Innovation", 19 May 2026. https://www.manh.com/about-us/newsroom/press-releases/manhattan-associates-unveils-manhattan-marketplace 7. DC Velocity. "Study: Leaders trust AI for decision support, but not to act on its own" (RELEX Solutions, State of Supply Chain 2026, survey of 514 leaders conducted January 2026), 25 March 2026. https://www.dcvelocity.com/technology/artificial-intelligence/study-leaders-trust-ai-for-decision-support-but-not-to-act-on-its-own 8. DC Velocity. "42% of logistics leaders are holding back on Agentic AI, survey shows" (Ortec survey of 400 North American executives), 15 January 2026. https://www.dcvelocity.com/editorial/featured/42-of-logistics-leaders-are-holding-back-on-agentic-ai-survey-shows 9. Boston Consulting Group. "AI Hasn't Fully Paid Off Yet in Logistics. But It Will." (BCG 2026 AI in Logistics Executive Survey), 24 August 2026. https://www.bcg.com/publications/2026/why-ai-isnt-delivering-roi-logistics 10. Gartner. "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 11. Salesforce. "New Study of 2,025 Agentic AI Leaders: First To Launch Isn't Fastest to ROI" (State of Agentic AI in the Enterprise, fielded 14–28 May 2026), 27 August 2026. https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/ 12. PYMNTS. "AI Agents Take the Controls in Supply Chains" (reporting on IDC research by Stephanie Krishnan), 21 September 2026. https://www.pymnts.com/news/artificial-intelligence/2026/ai-agents-take-the-controls-in-supply-chains/ 13. Gartner. "Gartner Identifies Top Supply Chain Technology Trends for 2026", 30 June 2026. https://www.gartner.com/en/newsroom/press-releases/2026-06-30-gartner-identifies-top-supply-chain-technology-trends-for-2026 14. Carriyo. "Agentic Commerce Meets the Last Mile: When the Shopper Is an Agent, Who Owns the Delivery Promise?" (Carriyo blog, August 2026). https://carriyo.com/blog/agentic-commerce-meets-last-mile-shopper-2026-08-18/ 15. Carriyo. Carriers. https://carriyo.com/carriers