AI engineer (artificial intelligence engineer)

An AI engineer, or artificial intelligence engineer, designs and develops AI solutions, then integrates them into the company's applications. Their work may include data preparation, model selection, evaluation, deployment and maintenance. They turn a user need into a feature whose results and limitations can be examined.

The role extends beyond generative AI. Depending on the project, it may involve prediction, computer vision, language processing or robotics. These are possible specialisations: a position does not require expertise in all of them. The work also depends on what the company wants to develop itself and which services it can integrate.

The engineer works with architects, data specialists and developers. Maintenance may form part of the role or be shared with operations specialists.

AI engineer, machine learning engineer and data engineer: who does what?

• The AI engineer connects the need, the technical solution and its integration. Their role may cover several stages, from initial experiments to monitoring the application.
• The ML engineer also works on modelling, integration and production. The two job titles overlap: neither title alone establishes who trains the models or who operates them.
• The data engineer works on ingestion pipelines and making data available. The preparation needed for a model must be divided between them and the AI engineer, without implicitly assigning the entire data platform to the AI engineer.

The same job title can involve different technical responsibilities.

An AI feature must be evaluated against its intended use. A convincing result in a demonstration does not, on its own, show how robust it will be in other situations. Part of the engineer's work is to choose validation data, performance measures and tests that make limitations visible.

Relating results to product constraints

Choosing a model also means considering the resources needed to run it. The quality achieved, response time and computing resources must be examined together. For the company, these trade-offs determine the conditions under which the product will operate and which decisions require approval from those responsible for the product or infrastructure.

Fictional example: a company is testing a system that identifies defects in images. Before choosing an approach, ask which defects the system detects and which it misses. Also ask how representative the test images are of those the system will need to process. An overall score cannot replace this examination of errors.

Organising technical dependencies

Data preparation includes addressing gaps, normalisation and examining anomalies. If making data available is a problem, clarify how the work is divided with the data engineer. Hiring an AI specialist does not remove the need to determine who supplies the data and who corrects its defects.

After deployment, performance monitoring and reproducible updates make it possible to examine how results change. Maintenance responsibilities should therefore be planned when defining the role.

For generative AI, another risk must be made explicit: a plausible answer may be inaccurate. Evaluation must take account of how that answer will be used. The CNIL highlights this risk, and no test or choice of tool can promise to eliminate it.

• Translate user needs into technical objectives and verifiable success criteria.
• Choose an AI architecture compatible with the available applications, data and resources.
• Prepare the data needed for models by addressing gaps, anomalies and differences in representation.
• Develop or integrate the models and services needed for the intended feature.
• Compare approaches using validation data and justify the trade-offs made.
• Test the solution's robustness in situations representative of its use.
• Integrate the feature into applications with the developers and architects involved.
• Monitor performance after deployment and prepare reproducible updates with those responsible for operations.
• Document technical choices, results and limitations for the teams that use or maintain the solution.

Useful routes into the role combine technical foundations with practical projects. The study by Apec and CESI describes education in computer science or applied mathematics, supplemented by practice, self-directed learning or research. These routes provide reference points without establishing a mandatory qualification for every AI engineer position.

The knowledge and skills needed depend on the work assigned. A modelling role draws on an understanding of data, statistics and evaluation methods. A role focused mainly on integration also requires programming and the ability to make a service work within an application. The mathematical depth expected should match the decisions the person will need to make.

A candidate with a software development background may have strong programming skills while needing to develop their understanding of models and validation. Someone whose background is more focused on data or research may need further learning in integration and maintenance. Existing skills and support needs vary with experience: the title of a qualification alone does not establish how independently someone can work.

Specialisation offers another route. The French Mastère Spécialisé in artificial intelligence listed by France compétences has its own admission requirements and provision for validation of prior experience. This example is neither a required route nor a promise of rapid retraining.

Keeping up with developments relevant to ongoing projects helps maintain skills.

GetPro's method is based on an assessment framework, open questions supported by examples, and reference checks. The guidance below helps adapt this foundation to the AI engineer role.

1. Define the criteria that will guide the decision

GetPro distinguishes verifiable aspects of a candidate's background from skills to examine in an interview. An assessment method is assigned to each criterion, and the framework then guides the discussions.

For your position, select capabilities that match the decisions assigned: preparing data, integrating a solution, evaluating its results or monitoring its operation. Specify what requires independent work and where support can be provided.

Link each capability to evidence that can be examined. An explanation of an architecture, an error analysis or an exercise may reveal different aspects. Avoid a list of tools that says nothing about what the candidate will need to achieve.

2. Examine a completed project and the candidate's personal contribution

GetPro interviews use open questions and concrete examples to explore the key criteria in depth.

Ask the candidate to describe a project from the initial need to the result achieved. Have them clarify their decisions, the work done by colleagues and the constraints they had to address.

If the project's code cannot be shared, ask the candidate to explain their personal contribution within the limits of what they can disclose. You can ask for a simplified diagram or an explanation of their trade-offs, or offer an exercise using data they are authorised to use. Look for consistency between the need, the decisions and the results described. These approaches are suggestions for adapting the assessment: they do not require access to a former employer's code or confidential data.

Then explore a particular choice: why this model, this validation data or this metric? Look for an explanation that connects the choice to the intended use. A list of technologies without supporting reasoning warrants further discussion.

A positive sign is the ability to describe a limitation and how it influenced the work. An answer that credits the candidate with the entire result without clarifying individual contributions calls for further questions.

3. Set an exercise close to the expected work

Fictional example: provide test data that is authorised for use and ask the candidate to compare two approaches to a classification task. Have them explain the errors observed, the choice of measures and the checks still needed.

Adapt the task to the position. For an integration role, you might favour an application diagram and a testing strategy. For a modelling role, explore data preparation and validation in greater depth.

Examine the approach as closely as the result. Someone who identifies the exercise's limitations provides useful information. A score presented without an explanation of the data or errors is not enough.

If your company lacks the necessary expertise, have the technical choices examined by a specialist who can discuss them critically. The managers concerned retain responsibility for defining business requirements, while the specialist examines the technical choices.

4. Assess communication and coordination

Ask for a presentation aimed at a non-technical audience. The candidate should be able to explain what the system does, what remains uncertain and which decision the results can help inform.

For a position involving coordination, explore how work was divided and how technical disagreements were handled. If the position includes people management, ask for an example of clarifying objectives and supporting a team member.

Distinguish precision of explanation from ease of delivery. A simple explanation that preserves the limitations of the result is a useful sign.

5. Compare observations with references

GetPro's reference checks place skills in the context of working relationships and seek specific examples. They complement the assessment of strengths and areas of concern identified in interviews, without automatically validating them.

For this role, ask for details of the candidate's contribution to the solution and the quality of their interactions with other teams. Cross-check the observations against your framework. A general judgement without an example leaves a question open rather than confirming a skill.