origin of the term
Nobody coined it. For fifteen years, people working on data kept reaching for the same shape — the hard part is at the start of the road.
Fifteen years, one shape.
Every use below links to its source. Entries marked a different sense use the same words for a different road.
-
2011
Data capture
“Data in the First Mile”
Kuang Chen, Joseph M. Hellerstein and Tapan S. Parikh (UC Berkeley), CIDR 2011
Database researchers use “first mile” for the hard, human work of capturing data where it is created — in low-resource communities, where the usual assumptions of data management do not hold. The paper proposes separating capture from structuring.
-
2017
Data access
“Solving Data Science’s First Mile Problem”
Jon Loyens (data.world), R Views, April 2017
The first mile becomes finding, getting and understanding the data before any analysis — the stretch where, the post argues, many would-be analysts give up.
-
2019
AI pipeline
“The Last Mile: Where Artificial Intelligence Meets Reality”
Enrico Coiera, Journal of Medical Internet Research, November 2019
The first explicit AI pipeline: in the “first mile”, data are acquired, labelled and cleaned; the middle mile builds the algorithm; the last mile puts it to work. Coiera’s argument is about the last mile — but the geometry is all here.
-
2022
Data ingestion
“Fixing the First Mile of Data Ingestion”
Osmos, August 2022
Data-ingestion vendors adopt the phrase for pulling messy external data in, cleaning it and loading it. From here on, “first mile” starts to appear in product marketing.
-
2023
A different sense
“First-Mile vs. Last-Mile AI Systems in the Era of Foundation Models”
Hazy Research (Stanford), March 2023
A different use: foundation models as “first-mile technology” — quick, exploratory, out of the box — against traditional machine learning in the production last mile. Same words, different road.
-
2025
Data ingestion
“First-Mile Data™”
Adeptia, 2025
An integration vendor trademarks “First-Mile Data” for external and unstructured data entering the enterprise — a sign the phrase had become worth owning.
-
2025
AI data readiness
“AI doesn’t have a last mile problem. It has a first mile problem.”
Anu Jain (Nexus Cognitive) on The AI Forecast, a Cloudera podcast — episode “Why Data Needs a Digital Birth Certificate”, host Paul Muller, 2025
The sharpest wording yet, and the inversion made explicit: models scale instantly; what fails is the quality, provenance and readiness of the data that feeds them. The line this site is built around.
-
2026
AI data readiness
“First-Mile Data: Setting the Foundation for Agentic AI”
LOMA MarketFacts, 2026
The insurance trade association applies the framing to agentic AI: before agents can act, the external data feeding them has to be ingested and made usable.
-
2026
A different sense
“AI Data Centers and the First Mile Infrastructure Fallacy”
Protiviti, June 2026
A parallel use about power, not data: the first mile of AI infrastructure is electricity generation and transmission for data centres. Same inversion, different road.
-
2026
AI data readiness
“Your Data Is Not Ready: Solving the First Mile Gap for Enterprise AI”
Kumar Goswami (Komprise), BigDATAwire, August 2026
“This gap is the first mile problem of AI.” Finding, understanding and classifying unstructured data before it reaches a model — the readiness sense, stated outright.
From capture to readiness.
The road stayed the same; what people pointed at along it changed. Early uses were about getting data at all — capturing it, then accessing it. Once machine learning made the pipeline explicit, the first mile became the acquire–label–clean stage before the model. Vendors then sold tools for ingesting it. And once frontier models became one API call away, the question shifted from “can we get the data?” to “is it ready for a model to use?” — the sense this site defines as the first-mile problem.
Two neighbouring uses share the words but not the road: foundation models as quick “first-mile technology” (Hazy Research, 2023), and the electricity that feeds AI data centres (Protiviti, 2026). They are recorded here so they are not confused with the data sense.
Know an earlier use?
This record is incomplete by design — it grows as uses are found. If you know an earlier or better source, send it through the contact in humans.txt. Additions are dated in the corrections log.
The origin, answered.
Who coined the term “first-mile problem”?
Nobody, as far as the record shows. Database researchers wrote about “Data in the First Mile” in 2011, data scientists called data access their “first mile problem” in 2017, and a 2019 paper in the Journal of Medical Internet Research described the first mile of AI as acquiring, labelling and cleaning data. Anu Jain’s 2025 line — “AI doesn’t have a last mile problem. It has a first mile problem.” — is the sharpest recent wording, not the coinage.
What does “first mile” mean in data and AI?
The stretch before the model: where data is created, captured, accessed, cleaned and made ready. The emphasis has shifted over time — from data capture (2011) to data access (2017), the AI pipeline (2019), data ingestion (2022) and, since 2025, AI data readiness: whether data is fit for a model to use.
Is the first-mile problem in AI the same as first-mile logistics?
No. In logistics and transport, the first mile is the first leg of a shipment or journey, from origin to the main network. The AI sense borrows the geometry of the telecom and logistics “last mile” and inverts it: the hard part is no longer the edge, but the source — the data.