Your customers are splitting your network into more carriers and more nodes. Your cost and tracking systems cannot see that shape. That is why the return on AI will not add up.
We spent Monday 14 September at the Analyst Day at PARCEL Forum '26, at the Gaylord Palms in Orlando. Four sessions. Two threads ran through all of them, and nobody in the room joined them up. This is the join.
The shopper is. Margot Juros, Research Director for Worldwide Retail Platforms and Technologies at IDC, opened her session on evolving consumer fulfilment needs with a finding that should worry anyone who plans capacity: the biggest factor in deciding where to shop, beyond price and convenience, is now the variety of delivery options on offer. Two years ago it was returns. Speaking at PARCEL Forum '26, Juros put the share of shoppers who would walk away from a merchant that fails on same-day options, free returns or the ability to choose a slot at 60 per cent.
North American shoppers, she added, over-index hard on one thing in particular. Live tracking. It is their number one factor by a distance, well clear of the global pattern.
So the specification for your network is now being written by someone who has never seen your depot. It reads alongside what retailers said they are actually judging you on, and it points the same way.
If your customer is a retailer rather than a shopper, none of that reaches you as a preference. It reaches you as a clause. It turns up in the service schedule of a tender: a named delivery window, a locker or pick-up option in postcode bands you cover badly, a scan event wanted on the retailer's own tracking page within minutes rather than at end of day, a service credit when you miss. The shopper's taste becomes a procurement team's contractual language two steps later, and it is tested at renewal.
Because the promise got specific, and one national network cannot hold a specific promise everywhere at a price anyone will pay.
Joanne Strong, Senior Manager and Warehouse Management System Practice Lead at Deloitte, told the Analyst Day about ordering from a Best Buy store at six in the evening and getting the item the next afternoon inside a two-hour window. The speed is not the interesting part. The machinery is. That order came out of a micro-fulfilment site, and Best Buy holds the promise across five to ten small carriers rather than one national network.
Andre Pharand, Founder and CEO of Pharand Advisors and formerly a Global Managing Director at Accenture, described the same movement from the carrier's side. Delivery is fragmenting, he argued, and the growth is going to whoever sits closest to the buyer. Chris Kina, Senior Director Analyst for Logistics, Customer Fulfillment and Network Design at Gartner, put a forecast on it that morning: large-carrier dominance is declining, asset-light operators are taking share, and the US is heading towards the European shape of hundreds of parcel carriers rather than a handful.
There is a marker for how far this has already gone. In calendar year 2025, Amazon Logistics moved 6.9 billion parcels in the US against USPS's 6.2 billion (Pitney Bowes Parcel Shipping Index, 2026 report). The largest parcel network in the country belongs to a retailer.
And the node count is climbing with it, much of it fulfilment capacity out of estate you already own.
Because volume is easier to buy than margin. Pharand walked the room through InPost, and InPost's own numbers make the point better than any characterisation. A locker network runs node to node and never pays for the last hundred metres. In its 2025 annual report, the locker-led Polish business reports an adjusted EBITDA margin of 49.0 per cent, up from 47.0 per cent. The UK and Ireland segment, in the same report, fell from 16.9 per cent to 2.8 per cent, which InPost attributes primarily to its April 2025 acquisition of Yodel, a business that "operated with substantially higher costs per parcel". Nothing changed about what a locker costs to serve. The mix changed.
His wider warning was blunter. Last-mile costs in the US have been climbing faster than last-mile revenues, which he called unsustainable, and the residential mix makes it worse. A home stop carries one or two parcels where a business stop carries three or four, so you spread the cost of the stop across fewer items. Now add carriers and nodes to that picture. The per-parcel economics stop being one average and become a question you have to answer per carrier, per node, per lane.
Integrations at a speed no procurement cycle currently allows, and one tracking vocabulary across all of them. Take ten carriers and forty nodes as an illustration rather than a description of your business. The arithmetic holds at whatever numbers you actually run.
One carrier means one data format, one scan taxonomy, one exception model. Ten carriers mean ten of each, while the customer still expects a single tracking page and a single promise. Every carrier you add is one more translation job, sitting on a smaller share of the volume.
The failure is never dramatic. It looks like this. A carrier changes a status code, or sends the same status under a slightly different name, and nobody notices, because the parcel is still moving. The scan arrives unmapped, so it never becomes an event in the exception language the rest of the operation runs on. The tracking page keeps saying in transit, because in transit is what a system says when it has nothing newer to show. No alert fires, because alerts are triggered by recognised failure states and this one is not recognised. The parcel sits at a node while the hub team reads it as the last mile's problem and the last mile reads it as the hub's. Then the customer calls, and the agent opens the same tracking page the customer is already looking at and has nothing to add. The retailer's account manager calls next, asking why their page says one thing and their shopper says another. Somebody eventually walks the aisle and finds it. By then the operation has spent far more attention on one parcel than it was ever worth, and the retailer has learned something about your visibility that no service review will unlearn.
None of that is exotic. It is the six steps a parcel operation actually runs failing at a single join, and the number of joins rises with every carrier you add.
This is where fragmentation plans quietly die. Operators routinely tell us their carrier integrations are quoted at three to six months each. That is a sensible number for a network adding one carrier a year and an absurd one for a network adding eight. What is needed is not a better integration project but an integration capability: a new carrier live in hours, scans normalised on arrival, one exception language, one promise to the customer whoever is carrying the box. The same holds across borders. Which channel each of those lanes should run on changes both your cost and the data you get back.
It is worth saying plainly what consolidation buys, because plenty of directors have spent two years earning it. Fewer carriers means real buying power at volume, one commercial relationship, one escalation path when a week goes wrong, and far less surface area to break. Those are not small things. A network that fragments without a reason trades a discount and a working escalation route for complexity it then has to fund. If one or two carriers already hold your promise profitably in the catchments that matter, adding a third to look modern will cost you on both sides of the ledger. Fragment when a promise cannot be kept any other way, and not before.
Because the money was committed before anyone agreed what would count as a return. Presenting at the Analyst Day on 14 September, Joanne Strong of Deloitte put the share of organisations calling AI a top strategic priority at 75 per cent, against roughly 16.5 per cent that are quantifying a return on those programmes. Her own diagnosis was that firms invest first and work out the KPIs afterwards.
Kina reported the same gap from the other end. Across the supply chain AI use cases Gartner evaluated, unclear return on investment was the barrier chief supply chain officers named most often. As we read his argument, the cause sits underneath the models rather than inside them: where the operational data is poor and siloed, the return goes with it. Run a good model on data like that and you do not get one bad answer. You get the same wrong assumption applied at every node it touches.
It is the thing you are pointing it at, and the two conference threads are one thread.
BCG calls the gap a deployment problem rather than a technological one, and points at firms buying standalone tools instead of building one connected system. Software vendors diagnose it as a tool-selection problem, meaning you bought the wrong platform. From an operator's chair it looks like neither. Your customers have just redrawn the network into more carriers and more nodes, and your cost and tracking systems were built to describe the old shape. Feed a model a monthly average built for one carrier and it will describe a network you no longer run.
Strong said something in passing that nobody picked up. Centralise the tech stack, she argued, precisely so the physical network can decentralise. That is the join, stated as architecture. Our reading is that the physical side is fragmenting because customers are pulling it apart, and that the information side has to converge at the same rate, or fragmentation just buys you more places to lose money quietly.
Cost and margin per parcel, per carrier, per node, inside the week the parcel moved. That is the input an optimisation engine needs before it can run at all.
Look at the one AI case everyone in parcel still quotes. Kina used UPS's ORION route optimisation as his worked example. In its 2016 annual report, UPS said the first phase of ORION was "now generating more than $400 million in annual cost savings and avoidance". That is a decade old and describes the original deployment, not what the system does today. What matters is why the figure could be stated at all. UPS could already count a mile, a stop and a minute, per driver, per day. The measurement existed before the model did.
Most operators cannot do the equivalent for a parcel that crosses several carriers and a node count nobody has totalled lately. What they have is cost information that arrives after the decision it was meant to inform. An AI business case fails its own approval when nobody can say what the current process costs, per node, per carrier, this week. That is not the model's fault.
GN TEQ builds the layer underneath this for parcel and logistics operators: multi-carrier integrations connected in hours rather than quarters, one harmonised tracking vocabulary across every carrier in the network, and cost and margin visible at parcel level while the week is still running. It sits on top of the ERP, TMS and WMS you already run.
Saudi Post Logistics used it to stop being the last leg of somebody else's shipment and become a cross-border carrier that owns the merchant relationship, the customs declaration, the tracking and the invoice. International commercial shipments went from zero to more than three million. Merchant onboarding went from over two months to two or three days, and delivery success from 85 per cent to 98 per cent, as published by the Universal Postal Union.
Why is the parcel market fragmenting? Because the promise got specific, and a promise that is specific to a postcode is held locally or not at all. The test on your own network is narrow. Take the three delivery promises your largest customers actually advertise, and ask which carrier and which node holds each one today. If it is the same carrier every time, fragmentation is somebody else's story for now. If two of the three depend on drop density you do not have, it is already yours.
How many carriers does a next-day network need? There is no best-practice number, and anyone offering you one is selling something. Two things you already know set the count: the promise you have made your customers, and the drop density in each catchment. Where density is high, one carrier may hold the promise profitably. Where it is thin, you will need several, or a locker and pick-up network, or a node closer in. Do that arithmetic and the carrier count is already decided.
Why do so few operators see a return on AI? Most cases fail at approval rather than in production, because nobody can state what today's process costs. The ones that close share a shape: a single decision that is measurably slow or expensive, a number for that decision which existed before the project started, and one person who stays accountable for the number afterwards. Miss any of the three and the return stays unprovable, whatever the model does.
What should we measure before automating anything? The cost and margin of a parcel, broken down by carrier, by node and by lane, available inside the week rather than several weeks later. Then the cycle time of whatever you want to automate: exception to resolution, tender to booking, contract signature to first normalised carrier scan. Those two sets of numbers are what an AI business case is made of, and most operators find they have neither when they sit down to write one.