Why do AI pilots fail when they reach the field?

A model that works in a demonstration and a system that works in the field are two different things. Demonstrations run on good connectivity, clean data, a dominant language, and forgiving users. Field deployment brings intermittent networks, messy and sensitive data, many low-resource languages, strict governance, and users who have no reason to trust an institution's technology. Systems fail at that last stretch because those conditions were treated as edge cases instead of design inputs.

How is Last-Mile AI different from standard AI deployment?

Standard deployment optimises for model quality and assumes the surrounding infrastructure will hold. Last-Mile AI starts from the operating reality: it designs for the channel people actually have, such as a voice call rather than an app, protects the people represented in the data, keeps a human in command of consequential decisions, and measures results in outcomes rather than output.

Who created the Last-Mile AI framework?

Last-Mile AI was set out by Shahzad Asghar, a United Nations AI and digital transformation leader, drawn from more than 20 years implementing AI and digital systems across UN humanitarian and development operations, including UNHCR's first global interactive voice response appointment system serving more than 700,000 refugees.

Last-Mile AI

Delivering artificial intelligence that works in the real conditions where it is used, not the controlled conditions where it is built. Last-Mile AI is a delivery discipline shaped by 20+ years of implementing AI and digital systems across United Nations humanitarian and development operations.

What Is Last-Mile AI?

Last-Mile AI is the practice of delivering artificial intelligence that works in the real conditions where it is finally used, not the controlled conditions where it is built. The term borrows from logistics, where the last mile, the short final stretch to the customer, is the hardest and most expensive part of the journey. AI has the same problem. A model that performs well in a demonstration often fails in the field because the last mile is full of constraints the demonstration never had to face.

It is a delivery discipline rather than a single technology. Last-Mile AI asks one question of every use case: what has to be true for this system to be trusted and useful in the place where it will actually run? The answer usually has less to do with the model and more to do with connectivity, language, data protection, governance, and the people who depend on the result.

Why AI Fails at the Last Mile

Most AI failures in real operations are not model failures. They are last-mile failures. The system assumed bandwidth that was not there, a language it did not support, a literacy level its users did not have, or a governance step that was quietly skipped under deadline pressure. The capability was real, but it never reached the people it was meant to serve.

This is especially true in humanitarian and public-sector settings, where the cost of failure is measured in delayed services, eroded trust, and harm to vulnerable people. A recommendation engine that mislabels a protection case, or a chatbot that only works in one language, does not just underperform. It can cause real damage. Last-Mile AI treats that gap between capability and trusted delivery as the central engineering problem.

The Constraints That Define the Last Mile

Five constraints recur across almost every difficult deployment. Connectivity is intermittent or absent, so systems must work offline or degrade gracefully. Language is plural, so interfaces and models must handle several languages and low-resource ones at that. Data is sensitive, so protection, consent, and access control come before scale, not after. Governance is mandatory, so ownership, review, and accountability must be designed in. And the operational context is unforgiving, so the system must hold up under stress, staff turnover, and edge cases that no test set anticipated.

Last-Mile AI makes these constraints first-class design inputs. They are not risks to be managed at the end of a project. They are the shape of the problem from the first day.

Principles of Last-Mile Delivery

Start from the operating reality, not the model. Define the use case in terms of who uses it, under what conditions, and what happens when it is wrong. Put data protection and human oversight at the points where judgment, rights, or material decisions are involved. Build for graceful failure, with clear fallbacks when the AI is uncertain or unavailable. And keep monitoring after launch, because a system that worked at handover can drift as data, languages, and operations change.

These principles turn a capable model into a delivered system. They are deliberately unglamorous, because the last mile rewards discipline over novelty.

Last-Mile AI in Practice

The approach is drawn from 20+ years of building AI and digital systems inside United Nations operations. An interactive voice response appointment system reached more than 700,000 refugees and was replicated across five country operations, because it was designed for low-bandwidth, voice-first, multilingual use rather than for a smartphone in a connected city. A refugee feedback platform, selected for the United Nations Global Pulse Accelerator, used AI to make community feedback usable while protecting the people who provided it.

In each case the model was the smaller part of the work. The larger part was the last mile: making the system trustworthy, accessible, governed, and durable in the place it was used.

Explore Further

For the structured methodology behind this approach, see The Last-Mile AI Framework, which sets out the pillars and steps in detail, or download the one-page framework (PDF). To see the principles applied, review the agentic AI projects, the work on AI in the humanitarian sector, and the broader areas of practice.

Get the next essay by email

One practical essay a month on AI governance, agentic AI, and digital delivery in the UN system. No marketing, no forwarding of your address.