Last-mile delivery AI: ETA optimization and anomaly detection

A delivery van sits stuck at the entrance to an underground parking garage in Neihu Science Park. The driver posts in the group chat: 'Waiting at the B building receiving dock, three vehicles ahead.' But on the system, this delivery's ETA has not updated in forty minutes. The recipient has called customer service twice. The rep stares at a dashboard full of green status lights and can only say, 'It shows you are on the way.'

The system is not broken. It simply never learned that underground parking garages get congested.

What last-mile delivery AI addresses

Last-mile delivery AI applies machine learning and real-time data to the portion of delivery from a vehicle's final hub to the recipient's door. This enables specific arrival windows, such as 3:40 to 4:10 PM instead of 'this afternoon', with genuine reliability.

Delivery times have relied on estimates for decades, but the traditional approach used averages and accumulated field knowledge. A route that runs three hours on average gets logged as three hours. The difference with last-mile work is its inherent variability. Line-haul transport follows predictable patterns, while city streets, building elevators, security checkpoints, and receiving docks introduce continuous unpredictability.

The limiting factor in ETA reliability is rarely computational. When predictions miss by an hour or more, drivers stop consulting them and customer service teams bypass the system to call drivers directly. Once this happens, model sophistication becomes irrelevant, the system has lost credibility.

This article focuses on two areas where ETA optimization generates immediate value: time window forecasting and anomaly detection. Each requires specific data inputs. Without them, models cannot produce actionable results.

Time window forecasting: reaching 30-minute accuracy

Time window forecasting aims to provide specific, reliable delivery windows to customers. Instead of a broad window like 9 AM to 6 PM, it enables predictions like 3:40 to 4:10 PM, precise enough for customers to plan accordingly.

Narrower windows reduce failed deliveries, the largest cost component in last-mile operations. When customers know the precise timeframe, they are more likely to be available. Companies that narrowed their windows from two hours to thirty minutes reported first-attempt success rate improvements of ten percentage points or more. Each avoided repeat delivery eliminates fuel costs, labor time, and operational friction.

Achieving this requires specific data collection. Most implementations fail at this stage:

The fourth requirement is often overlooked. In one case, a model performed well across most neighborhoods but consistently underestimated times in specific areas by five to eight minutes. Investigation revealed these were older buildings without elevators. The model had no data to signal this factor, so it could not account for the extra stair-climbing time.

Anomaly detection: catching problems before complaints

While time window forecasting optimizes typical conditions, anomaly detection identifies disruptions before they reach customers.

Anomaly detection addresses a basic question: is this delivery operating normally? Examples include a vehicle stationary for twenty-five minutes, a route running forty minutes behind forecast, or a package in transit with no scan events for two hours. When the system detects these patterns in real-time, it alerts dispatch or customer service, enabling intervention before customers initiate contact.

The value is operational and perceptual. Rather than waiting for customer complaints, the system communicates proactively: 'Your package is delayed 40 minutes due to traffic.' The same delay produces different outcomes depending on whether the customer learns of it first or discovers it through missed expectations.

Anomaly detection uses most of the same data as time window forecasting, with one additional requirement: real-time streaming. Historical data establishes baselines for normal behavior, but detecting current anomalies requires live location and scan events with minimal delay, ideally measured in minutes. Data that is hours old loses value; if a trajectory is delayed significantly, the customer has already contacted support.

An equally important and often overlooked requirement is labeled historical anomalies. The model must distinguish between normal stops, like driver lunch breaks, and genuine issues like mechanical failure. Many companies have extensive delay records but no categorization of causes. Without labeled data, the model generates false alerts, and field teams stop trusting the system within days.

Prioritize data readiness before model selection

Success in last-mile delivery AI depends eighty percent on data readiness, not model selection. Without timestamped stop events, structured addresses, real-time data streams, and labeled anomalies, model sophistication becomes irrelevant.

Implementation prioritizes field validation. The first step is determining where data collection gaps exist and where manual workarounds are currently necessary. Demonstrations matter less than actual performance once deployed. Systems generate measurable value only when drivers use the predictions and customer service teams act on the alerts.