Delivery ETA Prediction for Small Courier Fleets: Windows That Hold

"Your parcel is out for delivery." Everyone who receives that message thinks the same thing next: fine, but what time? An industry that moves billions of parcels a year still answers that question, for the most part, with "sometime today." The customer waits at home or doesn't; the driver arrives, rings, and the parcel goes back on the van for a second attempt. Then two calls to the support line: "What happened?" and "When tomorrow?"

Delivery ETA prediction promises to break that loop. Amazon, UPS and DoorDash have been computing arrival windows with machine learning for years. The question worth asking is whether a courier company with 12 vans, or an e-commerce warehouse running its own last mile, can realistically get there, or whether this remains a game for giants.

This guide takes the problem first, then how the prediction works, and finally what it costs at small scale and what it actually changes. Our AI-by-industry map described logistics as the function where most of your competitors have not moved yet; delivery windows sit right in that gap.

Is delivery slow, or just unpredictable?

In most developed parcel markets, speed has largely been solved and predictability has not. Take Turkey, where we work with delivery businesses, as a data point: official figures show 7% of domestic parcels delivered the same day and 82% within two days of acceptance. The day of arrival is known; the hour is not. That uncertainty, rather than speed, is where complaints and failed first attempts come from.

The scale is worth pausing on. That same market handled more than 1.45 billion postal and parcel shipments in 2025, a record and 5% up on the year before. The top three by volume are large national carriers, but the sector is far from three companies: the regulator lists 59 licensed providers, 40 of them on regional licences. The same shape repeats in every country: a handful of giants and dozens of small and mid-sized couriers, plus hundreds of e-commerce warehouses drawing their own routes.

Two honest gaps. We could not find a verified list of which carriers give customers an arrival window (reports suggest the large players are rolling out live driver maps), and no published first-attempt failure rate for that market. The calculations below use our own assumption for that figure and say so.

How does delivery ETA prediction work?

Delivery ETA prediction breaks a parcel's journey into segments and learns a duration for each from historical data: departure from the depot, driving between stops, service time at each stop (parking, doorbell, signature), and an allowance for finding the address. The simple version uses past averages; the advanced version uses machine learning models that also take in traffic, weather, address quality and individual driver habits.

DoorDash describes this segmentation openly: courier assignment, travel to the restaurant, pickup, travel to the customer, each segment with its own uncertainty, each learned separately. Their DeepETA model produces a probability distribution rather than a single minute. That distinction matters for a small operator too: telling a customer "2:00 to 4:00 pm" instead of "2:37 pm" is both more honest and easier to keep.

The model ladder, from simple to complex:

• Averages and rules: "Average stop time in this district is 4 minutes; multiply by position in the route." Most off-the-shelf routing software does this; it is a fine starting point.
• Gradient boosting: Classic models that learn from tabular data (hour, weekday, zone, parcel count, driver). The sensible second step for a company with a few months of data.
• Graph and transformer models: Deep learning that learns the road network and long-range time dependencies; Uber and DoorDash scale. A small company does not need them.

On accuracy, the widely quoted figures (machine learning improves ETA accuracy by 30-40%, long-tail errors drop by up to 20%) come from a routing vendor's blog, and a public error figure in minutes did not turn up anywhere. For a peer-reviewed starting point, the PeerJ Computer Science paper "Ten quick tips for improving ETA predictions using machine learning in logistics" is solid material to hand to a technical team.

The "I'll just use Google Maps" fallacy

In the last mile, driving time is only one part of delivery time. Service time per stop (finding parking, the building entrance, the doorbell, the signature), the order of stops, and the accuracy of the address affect the total as much as driving does, sometimes more. The map's "18 minutes" says nothing about when the van reaches the third of twelve stops.

This is what makes the last mile the hardest segment to predict: the most stops, the most traffic interaction, the most variable service time. Address quality deserves its own paragraph. A parcel aggregation platform claims that AI-assisted address validation catches more than 35% of address errors before the parcel leaves the depot; single-source vendor claim, but the direction is right. Promising a time window for an address written as "across from the old post office" is pointless. Prediction starts with clean addresses.

Do time windows actually help?

The available evidence says two things: the main cause of a failed first attempt is that the customer did not know when the driver was coming, and every failed attempt generates two separate customer contacts. From those two facts, an arrival-window notification reduces both second-attempt cost and support load. Every number on how much comes from software vendors, so here they are with their labels on.

A routing vendor puts the cost of a failed delivery attempt in the US at about $17.78 (citing another firm's data; single source). The same vendor describes two unnamed customers cutting "where is my order" (WISMO) contacts by 63% in 14 days with notification templates and by 71% in 21 days with a tracking page. An e-commerce shipping platform writes that an "out for delivery" text alone reduces "not at home" failures by 40%; again a single source. Do not paste any of these into a business plan. But it is no coincidence that they all point the same way.

The clearest effect we see in practice shows up at the support desk: the "when will it arrive" question goes largely quiet the moment the customer can see a window and a live location. Counting what share of your delivery-day contacts that question represents today is the first step of the investment decision.

We covered the WISMO side in our guide to reducing WISMO tickets; this article moves from "send a notification" to "what the notification should say."

What does it cost a small courier company?

Arrival-window prediction and automatic customer notification come bundled into off-the-shelf route optimization software at roughly $35-90 per vehicle per month; you do not need to build a model. OptimoRoute's plan with real-time re-optimization and live-ETA customer links is $44 per vehicle per month; Routific offers the first 100 orders a month free and $150 a month up to 1,000 orders. At the enterprise end, Onfleet starts at $619 a month.

Local routing vendors exist in every market, usually quoting on request. The question to ask before price: does the software send the customer a time window and a live tracking link, or does it only draw a route for the driver? Those are two different products. The fuel and time side of routing is in our route optimization software guide; the delivery window is the second layer on top of that setup.

The real yardstick is how many failed deliveries each van produces per day, and what each one costs in driver minutes on the second attempt, fuel, and support calls. Once that number is on paper, $40 a month per van stops being a debate at most companies.

A four-week pilot for one courier company

A model we built; the numbers are assumptions, the direction is real. The company: 12 vans, 80 stops per van per day, 26 delivery days a month, about 25,000 deliveries a month. Assume a first-attempt failure rate of 5% (international sources consider that plausible for parcel delivery): 1,250 parcels a month go to a second attempt. Each eats 10 driver minutes, some fuel, and about two calls.

The pilot steps:

• Week 1: Address hygiene. Pull the last three months of "not found" deliveries and separate the address-error cases. Add "landmark / directions" and "alternative recipient" fields to the order form.
• Week 2: Switch on the routing software for 4 vans. The customer gets a morning message ("today between 1 and 4 pm") and a live tracking link as the driver approaches. Keep the window wide; a narrow window you cannot keep is worse than none.
• Weeks 3-4: Compare the same 4 vans against the previous month: first-attempt success rate, delivery-day contacts, completed stops per driver.

Assume a cautious result: failed deliveries down 30%. Rolled out to 12 vans, that removes 375 second attempts and 750 calls a month and frees about 60 driver hours. Software cost for 12 vans is roughly $500 a month. Multiply by your own wage and fuel figures; at the companies where we have run this calculation, payback has never taken longer than two months.

Frequently asked questions

Should we build our own prediction model?

Not below about 5,000 deliveries a day. The rule-and-average ETA in off-the-shelf software is enough to start; after six months of data you can consider your own model, and even then the first step should be something simple like gradient boosting.

What if we promise a window and miss it?

A missed window generates more complaints than no window at all. Start with a three-hour window in the first month and narrow it to two hours once your hit rate passes 90%. If a window is going to slip, the customer needs an updated message before the driver arrives; that is where the software's "recalculate and notify" feature earns its keep.

How many minutes off is the prediction?

We could not find a public benchmark in minutes; vendors prefer percentage-improvement language. Build your own measure: the share of deliveries that land inside the window you promised. That single metric measures model quality the way the customer feels it.

Don't the big carriers already do this?

Live driver maps are spreading among large carriers; we could not verify carrier-by-carrier window practices. A small company's goal should not be to race the giants; offering the same experience to its own customers for a few hundred dollars a month is enough.

So what should you do?

• Measure your first-attempt failure rate. For one month, count "not at home", "address not found" and "unreachable" separately. No investment decision without this number.
• Fix the address form. A directions field and an alternative recipient are the cheapest improvement, ahead of any model.
• Pilot with 4 vans and a wide window. Three hours, a morning notification, a live link on approach.
• Track four metrics weekly: first-attempt success, window hit rate, delivery-day contacts, stops per driver.
• Keep vendor numbers out of the business plan; put the pilot's own numbers in.

Back to the opening question: "fine, but what time?" Answering it does not require being Amazon; a clean address, off-the-shelf routing software and an honest window are enough. The hard part is not the technology, it is the operational discipline to keep the window you promised. When you decide to build that discipline, we are glad to help design the pilot's measurement table.