AI Research Scientist vs ML Engineer vs Applied Scientist: Name Your Convention First
The three titles describe company conventions rather than fixed job definitions. Three questions about your own company name the convention you already run, and production code volume separates the roles more reliably than any other dimension.
The handoff between modeling and production decides how many of the three a team staffs. A correct hire landing inside the wrong convention resigns within the year. ISG Partners fills all three roles through recruiters who calibrate on the work rather than the title.
Which Naming Convention Does Your Company Run?
AI job titles operate as local dialects rather than definitions, so the same title names different work at different companies.
One person holds all three titles across a career while doing the same work. One large employer uses the research scientist title for a doctorate plus a publication record. A second employer uses the same title for a modeler who declines to write production code. Neither company is wrong. Each runs a dialect.
The consequence lands on the buyer. A requisition copied from a company running one dialect, into a company running another, sources the wrong pool. The mismatch surfaces at offer stage.
Three questions name your own dialect. Ask them about your company, never about the market.
Who does modeling work here today, and what does the title say?
Does that person write code that runs in production?
Does anyone here publish?
The answers sort into three patterns.
Publishing with low production code means your dialect treats research as a separate discipline. Research scientist means researcher.
No publishing, with production code coming from the modeling seat, means your dialect is applied. Applied scientist means a modeler who ships.
No modeling seat at all means your dialect is engineering. Every AI role gets called machine learning engineer.
Translate before you copy. Describe the work the hire owns in the first quarter, then attach whichever title your dialect uses for that work.
What Separates the Three AI Roles?
The three roles separate by the volume of production code the work requires, and every other difference follows from that volume.
Most comparisons draw a line from theory to production. A line describes three roles without deciding between them.
Production code volume decides better.
A research scientist owns problems with no published solution. The work runs experiments against benchmarks and produces methods, papers, and proof-of-concept code. Production readiness sits outside the remit.
An applied scientist owns known methods against your data. The work adapts published approaches, designs evaluation, and produces a model measured against a stated business metric. Some production code, rarely the serving layer.
A machine learning engineer owns the systems around models. The work builds training pipelines, deployment, monitoring, and retraining, and produces infrastructure that survives traffic.
One pool consequence decides how wide a search runs. Candidates cross between adjacent roles readily. Crossing into research, or out of research, stays rare. A search spanning two adjacent roles reaches a wider pool. A search spanning research and engineering reaches almost nobody. The compensation spread across AI and engineering roles runs widest at the research end for the same reason.
The table below sets the three roles against the dimensions a requisition needs.
| Dimension | Research Scientist | Applied Scientist | Machine Learning Engineer |
|---|---|---|---|
| What the role owns | Problems with no published solution | Known methods against your data | Systems that train, serve, and monitor |
| What the role produces | Methods, papers, proof-of-concept code | Validated models and evaluation results | Pipelines, endpoints, monitoring |
| Production code volume | Low | Moderate | High |
| Typical pool | Research groups and labs | Modeling and data science paths | Software engineering paths |
| What failure looks like | The method never works | The model never moves the metric | The system goes down |
| Hire when | Nobody has solved the problem | The method exists and your data is untested | The model works and the system does not |
Does an Applied Scientist Write Production Code?
The applied scientist question has no market-wide answer, and asking the question is how a buyer discovers which dialect their own company runs.
Authoritative sources split on the answer. One says the role rarely writes production code. Another says the role does both, and states outright that the answer flips between two named employers.
The split has a cause. The applied scientist title sits exactly where two dialects meet. A research-first company uses the title for the person who carries research toward product. An engineering-first company uses the same title for a modeler who ships. Both usages are established and neither yields.
The practical consequence arrives in the pipeline. A requisition using the title without describing the code expectation attracts two incompatible candidate populations. Half the shortlist expects to own a service. The other half expects to hand a notebook to somebody else. Both halves interview well against a job description that never named the difference.
Write the expectation into the requisition. Name the repository the hire commits to. Name the review process. State whether the hire owns anything that pages at night.
Applied science gets chosen deliberately in one situation. The method exists, the data is yours, and nobody knows yet whether the method holds on your data.
How Do the Newer AI Titles Map On?
Seven newer AI titles map onto the same three roles, and none of them names a fourth discipline.
An AI engineer builds product features on existing models. Closest to engineering, furthest from research.
An LLM engineer is an AI engineer with model-side depth in fine-tuning, retrieval, and evaluation.
An applied AI engineer is an AI engineer under a different label. Company convention, not a separate discipline.
A research engineer builds the training infrastructure and evaluation harnesses that research requires.
ML systems engineer, inference engineer, and AI infrastructure engineer name platform work closer to performance engineering than to modeling.
Member of technical staff is a lab title covering research and engineering together. Read the work.
Founding AI engineer names a stage rather than a discipline.
One rule covers all seven. Read the responsibilities section before the title. The responsibilities section gets written by someone who knows the work. The title gets written by someone who knows the ladder. Once the role is defined, how we source machine learning engineers covers the search itself.
Where Does the Handoff Sit?
The number of AI roles a team staffs depends on where work leaves one pair of hands and enters another.
Specialization creates a boundary. A modeler builds in a notebook. An engineer converts the notebook into a service. Context dies in the gap, iteration slows, and maintenance cost rises.
Two structures follow from that boundary.
A team with no handoff needs one person who spans modeling and production. That person exists, arrives rarely, and prices accordingly.
A team with a real handoff staffs both sides or accepts the loss across the gap. Staffing one side and assuming the other side arrives free is the most common structural error in early AI teams.
The second error runs the other direction. Research capacity arrives before a research agenda exists, and the capacity sits idle while the product needs models shipped. Most early teams need the applied and engineering end first.
Decide the handoff. Then decide the headcount. Then decide the titles. Reversing that order produces a requisition nobody inside the company agrees on.
What Breaks When the Convention Mismatches?
A candidate who matches the role and misses the dialect resigns inside the first year, and the search reads as successful until the resignation lands.
The pattern hides well. The requisition was correct. The shortlist was correct. The offer landed and the candidate accepted. The expectations sitting underneath the title belonged to a different dialect.
Three directional failures follow, each distinct.
A research hire in an applied seat produces careful work against a fast metric, then leaves for a lab.
An engineering hire in a research seat produces a wrapper around somebody else's paper and calls the wrapper a result.
An applied hire in an engineering seat builds infrastructure that works until traffic arrives.
Every one of the three reads as a performance problem from outside the team. None of the three is a performance problem. Strong hires fail inside the first ninety days for structural reasons more often than capability reasons, and convention mismatch is one of them.
Prevention costs one paragraph. Name the work in the requisition. Evaluate the same work in the interview. ISG Partners writes the first-quarter ownership statement during calibration, before sourcing opens.
When Does None of the Three Fit?
Four situations resolve without any of the three roles, and each one arrives disguised as an AI hiring problem.
The work is product engineering on existing models. A strong backend engineer with applied AI depth covers the requirement. None of the three titles applies.
The bottleneck is data quality. A data engineer fixes the input. A modeler hired into that problem models noise.
No evaluation harness exists. Nobody currently tells whether a model works. Build the harness before hiring the person who builds the model.
The need is one scoped research question. A consultant or an academic collaboration answers the question and stops. Neither justifies permanent capacity, and neither fits an embedded engagement.
Which Role Does Your Plan Need?
Naming your dialect, your handoff, and your first-quarter ownership converts an AI headcount request into a search that closes.
Three moves decide the outcome. Translate the dialect. Locate the handoff. Describe the work rather than borrowing a title.
The same discipline applies to a single senior AI search and to sustained hiring across modeling and platform at once.
ISG Partners calibrates every AI search on the work rather than the title. Reporting covers time-to-fill, cost per hire, offer acceptance, and 90-day retention every month. Start with which hiring plans fit an embedded engagement, then book a discovery call and bring the requisition exactly as written.
Common Questions About AI Hiring Titles
Which of the three roles does a first AI hire cover?
Most first AI hires cover the applied or engineering end. Research capacity ahead of a research agenda sits idle. Name the work the hire owns in the first quarter, then match the title to that work.
Does an applied scientist write production code?
The answer depends on the company. Research-first companies use the title for modeling with little production code. Engineering-first companies use the same title for a modeler who ships. Write the expectation into the requisition.
Why do the same three titles mean different work at different companies?
AI job titles operate as local dialects rather than definitions. Each company names the same range of work differently, so a requisition copied across companies sources the wrong candidate pool.
What breaks when the wrong role gets hired?
The hire underperforms or resigns for structural reasons. A research hire in an applied seat leaves. An engineering hire in a research seat ships a wrapper around somebody else's paper.
Is an AI engineer the same as a machine learning engineer?
Not quite. An AI engineer builds product features on existing models. A machine learning engineer builds and operates the systems that train, serve, and monitor models trained on your own data.