AI Engineer vs ML Engineer vs Research Scientist: Who Does Your Startup Actually Need?

An early-stage AI company rarely needs all three roles at once. It needs the person who can remove the most important technical constraint between the product today and the next meaningful customer outcome.

That sounds obvious, but “AI engineer”, “machine learning engineer” and “research scientist” are often used interchangeably. The result is a broad job description, an unfocused interview process and candidates who are solving different problems from the one the founder actually has.

The simplest distinction is this: an AI engineer usually turns existing models into a useful product, an ML engineer makes machine-learning systems reliable and scalable, and a research scientist creates or validates new methods. The boundaries overlap, but the centre of gravity is different.

Quick comparison

Role	Primary outcome	Typical strengths	Hire when
AI engineer	A customer-facing AI capability	Software engineering, model APIs, retrieval, agents, evaluation and product judgement	The model exists, but the product does not yet work reliably for users
ML engineer	A dependable production ML system	Training and serving pipelines, data, infrastructure, monitoring and optimisation	Model performance or production scale is the bottleneck
Research scientist	New knowledge, methods or model capability	Experiment design, mathematics, deep learning research and scientific communication	Your defensibility depends on solving a problem established methods cannot

Real teams also employ research engineers, applied scientists and ML infrastructure engineers. The title matters less than the ownership you define.

What an AI engineer should own

For most Seed-stage companies building on foundation models, this is the first specialist AI hire. The work often includes selecting models, building retrieval or agent workflows, creating evaluation sets, controlling latency and cost, and integrating the capability into the product.

The best candidates are not simply good at prompting. They can build production software, understand failure modes and make sensible trade-offs between model quality, speed, cost and user experience. They should be comfortable saying that a deterministic workflow is better than an LLM when the evidence supports it.

Hire an AI engineer when the central question is: “How do we turn available models and our proprietary context into something customers will repeatedly trust?” If your founding team still needs to discover the workflow and ship the first version, broad product judgement is normally more valuable than narrow model-training depth.

What an ML engineer should own

An ML engineer sits closer to the model lifecycle and production system. Depending on the company, they may build data and feature pipelines, train or fine-tune models, deploy inference services, monitor drift and quality, and improve throughput or cost.

This role becomes critical when prototypes work but reliability does not. It is also important for businesses using proprietary models, high-volume predictions or specialised data where evaluation and retraining cannot be managed as occasional engineering tasks.

Do not use “ML engineer” as a more impressive name for a backend engineer who calls an API. Equally, do not reject a strong software engineer because their previous title was not ML Engineer. Assess the systems they personally built and operated.

What a research scientist should own

A research scientist should have a genuine research question. Their job is to design rigorous experiments, test hypotheses and create capability that is not readily available through existing models, tools or published methods.

That may justify a doctorate and publication record, particularly in fields such as computer vision, reinforcement learning, speech, safety or scientific AI. Credentials alone are not the outcome. The candidate must be able to connect research progress to the company’s technical thesis.

Frontier organisations demonstrate how blurred the boundary can become. Anthropic currently advertises combined “Research Engineer / Scientist” roles alongside dedicated research, infrastructure and performance positions. Its alignment role asks candidates to build and run thorough ML experiments and values research-paper experience. The lesson for a startup is not to copy the titles; it is to define whether the work is judged by scientific insight, production performance or customer impact.

Choose the role from the bottleneck

Start with the next 12 months, not the eventual org chart.

• Customers like the demo but cannot rely on it: hire an AI engineer who can own product quality and evaluation.
• Models work in notebooks but fail under real traffic: hire an ML engineer with production systems depth.
• The company cannot win without a technical breakthrough: hire a research scientist or research engineer in the precise domain.
• The data is fragmented or unusable: the first hire may actually be a data or platform engineer.

This is particularly important in competitive markets. CBRE’s 2026 research reports that AI jobs represented 31% of US tech-talent listings in June, up from 11% in 2022. A vague brief enters a crowded market without giving candidates a reason to engage.

A sensible early-stage sequence

A Seed company using third-party models may begin with two strong product engineers, one of whom has applied AI depth. As usage grows, it may add an ML engineer to own evaluation infrastructure, fine-tuning, serving or data quality. A research scientist comes earlier only when original research is central to the product’s advantage.

Our guides to hiring AI engineers in San Francisco and building an AI engineering team in New York explain how local competition affects that sequence.

Build different interview loops

Give an AI engineer a realistic product failure and ask how they would diagnose it. Give an ML engineer a production system and probe data, deployment, observability and cost. Give a research scientist an open technical question and examine experimental rigour, baselines and what evidence would change their view.

Every candidate should still demonstrate ownership, communication and the ability to work with imperfect information. Early-stage companies cannot afford specialists who treat the surrounding product as somebody else’s problem.

How Dataworks can help

Dataworks works with Seed–Series B AI companies across technical and go-to-market hiring. We help founders turn a broad ambition into a precise search: the outcome, technical environment, evidence required and realistic candidate pool.

Before opening three roles, define the constraint you need one person to remove. Speak to Dataworks if you want to pressure-test the brief before taking it to market.

Featured photograph: Mohammad Rahmani on Unsplash.