Skip to content
AI Engineer Guide.

concept · concepts · beginner · concept 2 of 12

AI vs ML vs data vs research

0 of 12 complete

Why it matters

Job titles are a poor map of the work. Two companies will call the same role "AI engineer" and mean a prompt author at one and a distributed-systems engineer running inference at the other. The consequence is practical: teams hire for a title, then discover nobody owns the evaluation set, the data contract, or the serving path.

The distinctions matter most at the seams. A research team can hand over a method with no serving story. An ML team can hand over a model with no data pipeline that keeps its features current. A data team can hand over tables with no definition of what "correct" means for the product question. Each handover looks complete and leaves the product broken six weeks later.

Mental model

One construction site, four trades. The architect decides what should exist and why it has never been built that way. The structural engineer proves the loads hold. The surveyor establishes where the ground actually is. The general contractor delivers a building someone can occupy, and owns the schedule, the inspections and the snag list.

The analogy breaks because construction trades are stable and licensed, while these four roles blur. On a small team one person is all four, and that is legitimate — as long as somebody consciously covers the contractor's duties: integration, testing, operations and the decision about what is good enough to ship.

How it works

Mechanism, step by stepInteractive
  1. 1Research optimises for a claim that survives scrutiny: a new method, a measured effect, an explanation. Success is reproducibility and contribution, not uptime. Artefacts: papers, ablations, released checkpoints. It is allowed to fail, and often should.
  2. 2ML engineering optimises for model quality against offline metrics, and for a training pipeline that can reproduce it. Artefacts: feature pipelines, training jobs, model registries, evaluation reports, drift monitors.
  3. 3Data engineering optimises for trustworthy, timely, well-defined data. Artefacts: ingestion, warehouse models, data contracts, lineage, quality tests. Its output is a guarantee about a table, not an answer to a question.
  4. 4AI engineering optimises for a user-visible outcome under latency, cost and safety constraints, using models as components. Artefacts: prompt and context assembly, retrieval pipelines, output schemas, eval suites, traces, fallback paths, runbooks.

Step 1 of 4

Research optimises for a claim that survives scrutiny: a new method, a measured effect, an explanation. Success is reproducibility and contribution, not uptime. Artefacts: papers, ablations, released checkpoints. It is allowed to fail, and often should.

Step 1/4 · step 1

Figure. The mechanism in this page's own steps. The figure adds sequence, not new claims.

What to notice. A step can be read alone, but the order is the mechanism. The control lets you stop on any one.

Text version

The page's steps: 1. Research optimises for a claim that survives scrutiny: a new method, a measured effect, a… 2. ML engineering optimises for model quality against offline metrics, and for a training pi… 3. Data engineering optimises for trustworthy, timely, well-defined data. Artefacts: ingesti… 4. AI engineering optimises for a user-visible outcome under latency, cost and safety constr…

The roles are best separated by their objective function, then by the artefacts they leave behind.

  1. Research optimises for a claim that survives scrutiny: a new method, a measured effect, an explanation. Success is reproducibility and contribution, not uptime. Artefacts: papers, ablations, released checkpoints. It is allowed to fail, and often should.
  2. ML engineering optimises for model quality against offline metrics, and for a training pipeline that can reproduce it. Artefacts: feature pipelines, training jobs, model registries, evaluation reports, drift monitors.
  3. Data engineering optimises for trustworthy, timely, well-defined data. Artefacts: ingestion, warehouse models, data contracts, lineage, quality tests. Its output is a guarantee about a table, not an answer to a question.
  4. AI engineering optimises for a user-visible outcome under latency, cost and safety constraints, using models as components. Artefacts: prompt and context assembly, retrieval pipelines, output schemas, eval suites, traces, fallback paths, runbooks.

In practice the boundaries are drawn by who is paged. If a retrieval regression wakes someone at night, retrieval is owned by whoever is on call, regardless of title. Role clarity is really ownership clarity: for each component, who decides, who is paged, and who signs off on a change.

Research
Optimises novelty and evidence; artefact is a paper or checkpoint
ML engineering
Optimises offline model metrics; artefact is a reproducible pipeline
Data engineering
Optimises data correctness and freshness; artefact is a contract
AI engineering
Optimises the shipped outcome; artefact is a system with a quality bar
Shared failure
Nobody owns the eval set after launch

A concrete example

A bank wants to answer questions about loan policy. The research team proposes a retrieval method and reports a ranking gain on a public benchmark. The ML team trains a classifier that labels query intent and reports F1 on a held-out set. The data team models the policy corpus and defines freshness SLAs for the source documents. The AI engineer builds the assistant: chunking, hybrid retrieval, reranking, a citation contract, an abstain path, a p95 latency budget and the eval set the bank's compliance team signs off on.

Each trade did its job. The project still fails if the research method was never ported to the production retriever, if the classifier's label set drifts from real queries, or if the freshness SLA excludes the one document that changed last week. All three are seam failures, and none of them are visible in any single team's metrics.

Trade-offs

Hire forYou getYou give upFails when
Research scientistNovel methods, principled ablationsShipping cadence, on-call ownershipThe gap is integration, not method
ML engineerTraining and serving discipline, offline metricsProduct framing, eval design for open-ended outputThe task is generative and has no clean label
Data engineerTrustworthy data, lineage, freshness guaranteesModelling judgement about user-facing behaviourThe requirement is answer quality, not table quality
AI engineerEnd-to-end delivery, evaluation, operationsDeep specialism in one layerThe problem genuinely needs new modelling work
One generalist for all fourSpeed and low coordination costDepth, and a single point of failureThe system reaches production scale or a regulated bar
Trade-off mapInteractive

Good for

Novel methods, principled ablations

Cost

Shipping cadence, on-call ownership

Fails when

The gap is integration, not method

  1. The gap is integration, not method
  2. The task is generative and has no clean label
  3. The requirement is answer quality, not table quality
  4. The problem genuinely needs new modelling work
  5. The system reaches production scale or a regulated bar

Figure. Each choice with what it buys, what it costs, and the case where it loses.

What to notice. The fourth column is the decision: if the losing case describes your situation, the choice is wrong.

Text version

Four trade-offs: Research scientist is good for Novel methods, principled ablations, costs Shipping cadence, on-call ownership, and fails when The gap is integration, not method. ML engineer is good for Training and serving discipline, offline metrics, costs Product framing, eval design for open-ended output, and fails when The task is generative and has no clean label. Data engineer is good for Trustworthy data, lineage, freshness guarantees, costs Modelling judgement about user-facing behaviour, and fails when The requirement is answer quality, not table quality. AI engineer is good for End-to-end delivery, evaluation, operations, costs Deep specialism in one layer, and fails when The problem genuinely needs new modelling work. One generalist for all four is good for Speed and low coordination cost, costs Depth, and a single point of failure, and fails when The system reaches production scale or a regulated bar

Failure modes

Hiring the title instead of the owner. The team gains a researcher but still has nobody accountable for the eval suite. Debug: write the ownership table before the job description — component, decision-maker, on-call, approver — and hire against the gaps.

The eval set lives with the wrong team. Product owns the bar, engineering owns the harness, and the two drift until the score no longer means anything. Debug: make the eval set a versioned artefact with a named owner and a review date; treat edits to it as product decisions.

Data contracts stop at the warehouse. The corpus that feeds retrieval is a separate pipeline from the one with freshness tests. Debug: trace one document end to end and check whether its SLA covers chunking, embedding and index refresh.

Research handover without a serving story. A method arrives as a notebook with no latency profile, no failure behaviour and no fallback. Debug: require a latency and cost estimate plus a degraded mode before accepting the handover.

In production

The handoffs need artefacts, not meetings. In practice that means a data contract with freshness and schema guarantees, a versioned eval set with a scoring rubric, a model card or system card describing intended use and limits, and a runbook that says what to do when quality or latency degrades. When those four exist, role boundaries stop mattering; when they do not, reorganising the team will not help.

Common misconceptions

People think: AI engineering is just ML engineering with a new name.

Actually: ML engineering centres on training and offline model metrics. AI engineering centres on composing frozen models into a product, where the hard problems are context, evaluation, latency and failure handling.

People think: Data engineering is upstream plumbing that ends at the warehouse.

Actually: For retrieval systems the data pipeline runs all the way to the index. Freshness, deduplication and chunk-level provenance are part of the product's answer quality.

People think: A research background is the strongest signal for a product AI role.

Actually: Research skill predicts method quality. Product AI work rewards judgement about scope, measurement and operational cost, which is usually built by shipping.

Interview angle

A common opener is "how do you work with data, ML and research teams?" — asked to see whether you understand the seams. Answer with the artefacts you exchanged, not the org chart: the eval set you co-owned, the data contract you negotiated, the serving constraint that shaped the research handover. Then name a case where the boundary was wrong and what you changed.

See tell me about a time you chose a simpler model or architecture and the rest of the behavioural bank.

Sources