first mile
vs. last mile
Two names for the same gap, standing at opposite ends of the road. One was solved by building outward. The other cannot be.
last-mile problem noun — the difficulty of completing the final leg between shared infrastructure and an individual endpoint. first-mile problem noun — in artificial intelligence, the inversion of the above: the tendency of initiatives to fail at the start of the data pipeline rather than at the model.
What the last-mile problem actually describes.
The phrase comes from telecommunications. A carrier could run trunk lines across a continent at reasonable cost per kilometre, because that capacity was shared by everyone downstream of it. The stretch that resisted was the last one: the individual connection from the network into each specific building. It could not be shared, it had to be built once per premises, and it involved streets, permits, landlords and physical labour.
Logistics borrowed the term for the same reason and the same shape. A container crossing an ocean moves at a low cost per parcel because it is consolidated. The final leg from a local hub to one door is a vehicle, a driver, a street and a single recipient — and it is widely described as the segment where delivery cost and complexity concentrate.
The structure is identical in both fields, and it is worth naming precisely, because it is what got inverted: the core is efficient because it is shared; the edge is expensive because it is not. Infrastructure at the centre, gap at the perimeter.
Why "first and last mile" is also a logistics term
In supply chain the two ends are often discussed together. The first mile is the movement from a seller or manufacturer into the carrier network; the last mile is the final delivery to the recipient. Both are costly for the same structural reason — neither end consolidates the way the middle haul does.
That is a different sense from the one used on this site. The AI usage borrows the geometry of the metaphor, not the freight: what matters is where the gap sits relative to what has already been built.
Same road. The gap moved.
Draw both eras on the same axis and the argument stops needing words. The amber block — the part nobody has solved — jumps from one end to the other.
The two problems, line by line.
Read down either column and it holds as a description of its own era. Read across and every row is an inversion of the row beside it — which is the whole claim the term makes.
| Dimension | Last mile | First mile |
|---|---|---|
| Era | Telecom & logistics · 1990s–2000s | Artificial intelligence · 2026– |
| Where the gap is | At the edge — the final connection into each home | At the source — the data's first mile |
| What's abundant | Trunk infrastructure & bandwidth | Frontier models, one API call away |
| What stalls it | The last hop into each premises | Everything upstream of the prompt |
| Nature of the work | Physical — streets, permits, cable, labour | Organisational — definitions, ownership, provenance, governance |
| How it fails | Loudly — the connection simply is not there | Quietly — the model answers anyway |
| Who can solve it | A provider, building outward, once per endpoint | Only the organisation that owns the source |
| Can it be bought | Largely yes — it was a capital problem | No — the work is specific to your own data and teams |
Distribution became trivial. Readiness did not.
Every previous wave of technology made you wait at the edge. Electricity, phone lines, broadband, cloud — in each case the capability existed centrally long before it reached you, and the interesting question was when the infrastructure would arrive.
AI broke that pattern. A frontier model is available to every company on earth on its release day, through the same interface, at the same published price. There is no queue and no build-out. The last mile of AI solved itself.
What did not solve itself is the other end. The model still cannot do anything until an organisation hands it something to reason over — and that hand-off runs through data that was created for another purpose, moved by pipelines nobody owns, defined differently by each team, and governed at the table but not at the copy. That is the first mile — seven layers of it — and no vendor can ship it to you, because it is made of your own systems and your own agreements.
Which produces the uncomfortable corollary: the part that is now scarce is the part you cannot buy. When every competitor holds identical intelligence, advantage moves to whoever has the cleanest, best-governed, most model-ready first mile.
Where the first-mile framing comes from
AI doesn't have a last mile problem. It has a first mile problem.
The framing has since been picked up independently. Protiviti applied a parallel first-mile lens to AI infrastructure, and LOMA applied the same framing to agentic AI in insurance.
Three arrivals at the same metaphor, from infrastructure, consulting and insurance, is usually a sign that the metaphor is describing something real rather than being clever about it.
First mile, last mile, answered.
What is the last-mile problem?
The difficulty of completing the final leg between shared infrastructure and an individual endpoint. It comes from telecommunications, where it describes the connection from a provider's network to each customer's premises, and it is used widely in logistics for the final delivery leg from a local hub to a door. In both fields the shape is the same: the trunk is efficient because it is shared, and the final stretch is expensive because it is not.
Does AI have a last-mile problem?
Technically yes — and it is already solved, which is the whole point. AI's last mile is distribution: getting model capability from where it is built to whoever wants to use it. In every previous technology wave that was the expensive part, and here it collapsed. A frontier model reaches every company on earth the moment it ships, through the same API, at the same published price. No build-out, no queue, no waiting for the infrastructure to reach you.
What is not solved is the opposite end: the data an organisation has to create, collect, govern and prepare before the model has anything to reason over. Hence the framing — AI doesn't have a last-mile problem, it has a first-mile problem.
What is the last mile in AI?
The last mile in AI is the stretch between a finished model and the person or system using it: deployment, integration, the interface. It is short and largely commoditised — an API call, an SDK, a managed endpoint — and because it is short, it is a poor place to look for either failure or advantage.
This is the inversion of the telecom and logistics last mile, where that final leg was the expensive one. When an AI initiative stalls, the cause is almost never in the last mile. It is upstream, in the seven layers the data had to survive first.
What is the difference between the first-mile and last-mile problem?
They are mirror images — the same gap at opposite ends of the road. The last-mile problem has infrastructure at the core and a gap at the edge: the highways are built, the individual homes are not connected. The first-mile problem in AI inverts that geometry: the intelligence now sits at the edge, one API call from anyone, and the gap has moved to the source, in the data that has to be created, collected and prepared before the model sees anything.
What does the first and last mile problem mean in logistics?
In logistics both terms describe the two ends of a shipment's journey. The first mile is the movement from the seller or manufacturer into the carrier network; the last mile is the final delivery from a local hub to the recipient. Together they are the segments where cost and complexity concentrate, because neither can be consolidated the way the long middle haul can. That is a different sense from the AI usage here, which borrows the geometry rather than the freight.
Why did AI invert the last-mile problem?
Because distribution stopped being the hard part. A frontier model reaches every company on earth the moment it ships, through the same API, at the same price. The last mile of AI solved itself. What did not solve itself is everything upstream of the prompt — the messy, siloed, undocumented, ungoverned data an organisation has to feed the model before it produces anything worth trusting.
Where does the term first-mile problem come from?
The framing was articulated by Anu Jain, Founder & CEO of Nexus Cognitive, on The AI Forecast, a Cloudera podcast: "AI doesn't have a last mile problem. It has a first mile problem." Protiviti has since applied a parallel first-mile lens to AI infrastructure, and LOMA has applied the same framing to agentic AI in insurance.