Last-mile AI: Pointed at the wrong problems

Every front-door outcome traces back to decisions made at three different time horizons: structural decisions made months in advance, recurrent decisions made weeks ahead of time, and real-time decisions made in the moment. Last-mile AI earns its place when it improves the right decisions at the right time. Going beyond decision-making and into day-to-day operations, the supply chain industry seems to have reached a consensus on where last-mile AI belongs. That's a problem. Routing AI investments are saturated AI adoption in routing and visibility already sits above 70% according to Bringg research in The 2026 Last-Mile Performance Outlook . The report surveyed 150 retail and logistics executives at companies with more than $1 billion in annual revenue and found that routing and visibility are the most invested-in applications in last-mile delivery. And it's valuable. Route optimization improves on-time rates, reduces windshield time, and tightens delivery windows. But when an entire market concentrates AI investment in one function, the question changes from, “ Does routing AI work?” To, “ What are we not fixing if we’re all fixing the same thing?” To answer that, start with the decisions themselves. Routing AI is a real-time decision that inherits every structural and recurrent choice made before it, anywhere from one week, one year, or even a decade ago. The operational decisions routing AI doesn't touch Take recurrent decisions , for example. Every week, a planner builds the next week's carrier allocation against projected volume. In most operations, that means pulling last week's delivery data, cross-referencing carrier rate cards and SLAs, and manually adjusting zone assignments. If volume spikes unexpectedly or a carrier underperforms, the weekly plan breaks, and dispatchers spend the rest of the week compensating as a result. The report found that some mission-critical operations are still largely manual: 48% of billing and invoice reconciliation 42% of carrier management 39% of exception handling These aren't back-office curiosities. They're the workflows where operational cost compounds and where the people closest to end customers spend time on tasks that don't improve customer experience. The AI blind spot Almost three out of four executives (68%) plan to make additional investments in routing and planning AI despite already being the most-adopted workflows. The investments follow visibility: routing produces dashboards and surfaces in operational reviews, so it’s easy to point to when leadership asks what AI is doing. The workflows at the bottom of the investment list don't produce dashboards. Billing reconciliation, carrier management, and exception handling happen in the background, but that's where operational cost compounds. An unreconciled billing discrepancy leaks margin on every carrier invoice. A carrier allocation built on last week's data and gut feel sets the cost floor for the entire week before a single route runs. Yet, only about 14% of enterprise executives plan to increase investment in billing reconciliation and carrier management. These are the workflows most directly connected to the metrics where performance is weakest. Cost per delivery sits at only 36% overachievement, the lowest figure in the dataset and the metric executives rank first in severity. Operational efficiency trails by a similar margin. More routing AI doesn't touch either one. It improves on-time delivery, which is already the strongest metric in the dataset at 63% overachievement, and produces diminishing returns on a problem already largely solved. The investment mismatch creates a competitive gap most organizations don't recognize. The weakest metrics sit in the decisions routing doesn't reach. That's exactly where AI investment runs thinnest. Every point of underperformance on cost per delivery and operational efficiency traces to decisions that remain largely manual. Three questions last-mile AI should answer The underserved decisions don't need more dashboards, they need a different kind of help. AI that can advise, act, and explain. Three questions separate AI that earns its place from AI that merely occupies it. "What should we do?" Some decisions involve genuine tradeoffs that persist beyond the moment; for example: Add capacity next week and costs rise, but on-time rates improve Hold headcount and the budget stays flat, but late deliveries increase and the customer service team absorbs the impact Shift volume between carriers and cost per delivery changes, but service levels shift unpredictably across zones. These are the recurrent and structural decisions where a planner or operations lead needs to see projected outcomes before committing. AI in this role acts as an advisor: it simulates scenarios, projects the cost-to-service tradeoff of each option, and explains its reasoning so the human can weigh factors the data doesn't capture. The planner still decides, on;y with better reasoning because th