the first-mile glossary
The vocabulary of the road before the model. 20 terms, each defined once, plainly, and from the first mile's point of view.
Defined once, plainly.
Most of these words have a generic definition you can find anywhere. What follows is the generic definition plus the thing that makes it matter upstream of a model — because that is the part usually left out.
first-mile problemnoun · the term
In artificial intelligence, the tendency of initiatives to fail at the first mile of the data pipeline — where data is created, collected and prepared — rather than at the model. Named by inversion of the last-mile problem: models now scale instantly through the cloud, while the data that feeds them does not. Distribution became trivial; readiness did not.
See — The field definition
first-mile datanoun
Data as it exists at the start of the pipeline: produced by an operational system or a human process, for a purpose that was not analysis and certainly not inference. Raw not in the sense of unprocessed but in the sense of not yet fit to be read by a model. Data can be cleaned, modelled and served from a governed warehouse and still be first-mile data.
See — First-mile data
first-mile data ecosystemnoun
The set of systems, contracts and practices standing between where data is created and where a model can read it, across seven layers: creation, collection, contracts, quality, provenance, governance and readiness. Called an ecosystem rather than a stack because it includes things that are not tools — ownership, agreements, definitions — and because no single product covers it end to end.
See — The seven layers
last-mile problemnoun
The difficulty of completing the final leg between shared infrastructure and an individual endpoint. From telecommunications, where it names the connection from a carrier network into each premises, and used widely in logistics for final delivery. The trunk is cheap because it is shared; the last stretch is expensive because it is not. Infrastructure at the core, gap at the edge — the geometry AI inverted.
See — First mile vs. last mile
model-ready datanoun
Data that a model can consume and produce a defensible answer from. Model-readiness is not accuracy; it is six properties — resolved, defined, traceable, current, governed and evaluable. Correct data routinely fails every one of them, which is why "our data is clean" and "our data is ready" are different claims about different things.
See — The six properties
data readinessnoun
The state of data being fit for a specific downstream use, judged against the question about to be asked of it rather than against the job it was created for. Readiness is relative: the same table can be ready for a report and unready for a model. In the first mile it is the layer with the fewest owners.
data provenancenoun
The recorded origin and custody of a piece of data: which system produced it, under what logic, and who touched it on the way. Distinct from lineage in emphasis — lineage traces the path, provenance establishes the claim. The governing constraint is simple: you cannot govern what you cannot trace, and you cannot defend an AI decision whose inputs you cannot reconstruct.
data lineagenoun
The mapped path of data through the systems that move and transform it, from source through each job to the point of consumption. Useful for impact analysis and debugging. Its characteristic first-mile failure is stopping at a boundary — complete up to the warehouse edge and absent exactly where a model's input begins.
data contractnoun
An explicit agreement between the team producing data and the teams consuming it, covering shape, types, semantics, the meaning of a null, what may change and with how much notice. The most social and least technical layer of the first mile, and for that reason the one most often skipped — which is why a silent upstream change breaks a downstream model weeks later.
data qualitynoun
Whether data is true: completeness, validity, distribution, referential integrity, and the anomalies that indicate something upstream moved. Quality is necessary and insufficient for AI. A demo forgives messy input; production bills you for it; and a model bills you quietly, because it produces a fluent answer either way.
See — Quality, the third layer
entity resolutionnoun
Determining that separate records refer to the same real-world thing, and reconciling them into one identity. Also called record linkage or deduplication. The first-mile stake is high: ambiguity a human silently reconciles, a model confidently averages — the same customer under four identifiers becomes a total that counts them four times.
data governancenoun
The rules deciding who may use which data, for what purpose, and under which obligation: access, purpose limitation, retention, residency, consent and the classification that drives them. In the first mile the requirement is that policy travels with the data. Rules enforced on the source table but not on the extract taken for a prototype describe intentions rather than systems.
data freshnessnoun
How recently data reflects the reality it describes, measured as the lag between an event and its availability downstream. Freshness is not a global setting but a per-question requirement, so most staleness incidents are a mismatch rather than a failure: data accurate as of last night, feeding a decision about this morning.
data observabilitynoun
Continuous monitoring of the health of data and the pipelines carrying it — volume, freshness, schema, distribution and lineage — so that breakages are detected rather than reported by whoever they broke. Observability tells you a pipeline is behaving. It does not tell you the data means what the consumer assumed it means.
data silonoun
Data held in a system or team where it is functional for its owner and effectively unavailable to everyone else — not through prohibition, usually, but through absent documentation, unclear ownership and no path to request it. Silos are a first-mile condition rather than a storage problem: the data exists and no one can responsibly use it.
See — Readiness
schema driftnoun
Unannounced change to the structure of incoming data: a column added, removed, retyped or renamed by a producer with no obligation to tell anyone. The loud failure mode of a missing data contract — loud because pipelines break visibly, which makes it the more forgiving of the two drifts.
See — Data contract
semantic driftnoun
Change in what a field means while its structure stays identical — the definition of "active" quietly narrowing, a status code acquiring a new use. The quiet counterpart to schema drift, and far more dangerous in the first mile: nothing breaks, every test passes, and every answer computed since the change is subtly wrong.
evaluation setnoun
A held-out collection of inputs with known-good outputs, used to measure whether a system is actually working. In the first mile it is the property that closes the loop: without something to compare output against, you have not built a system, you have built an opinion generator with excellent latency.
grounding datanoun
The organisation-specific material supplied to a model at inference time so its output is anchored in fact rather than in parametric memory — documents, records, retrieved context. Grounding does not repair the first mile; it inherits it. Ungoverned, unresolved or stale grounding data produces confident, well-cited, wrong answers.
See — Not wrong, not ready
data productnoun
A dataset treated as something with an owner, consumers, a documented interface and a service expectation, rather than as the exhaust of a system that happens to be queryable. The framing matters in the first mile because it assigns the seat that is otherwise empty — someone accountable for whether the data is usable, not merely present.