AI Research or AI Engineering? The Importance of Correctly Categorizing AI Projects

A high proportion of businesses have undertaken an AI project in some form, with research from McKinsey revealing that 55% of organizations actively deployed an AI solution last year.

However, in order to fully understand what an AI strategy can achieve in a given time period, business leaders must consider which category it falls into. Companies often confuse AI research and AI engineering projects, leading to the misalignment of objectives and doubts about the credibility of AI for the long-term future of the business.

So, how can organizations develop an AI project that best aligns with the needs of their business, while setting the right expectations?

Defining a research AI project

The primary goal of a research AI project is to explore, investigate or discover new network architectures, algorithms or technologies. It aims to advance the understanding of a particular problem or domain and solve what has not been solved so far, while examining how pre-existing technology works under new conditions.

Weekly Update on AI ML: How to Streamline Web Data Processing With AI in 2024

Research projects often involve experimentation, a literature review, hypothesis testing, and an analysis of findings. The typical structure follows this pattern: define the problem, check existing literature, formulate a hypothesis, develop the conditions to test the hypothesis, complete testing, draw conclusions and begin further tests.

The most important aspect of the research project is that it ends with a report that summarizes everything that has been done, highlighting what works and what potential the technology has. Based on such a report, organizations can decide to build a product using the innovation that has been tested.

In certain cases, the project will conclude with an admission that the hypothesis tested can’t be achieved – this doesn’t necessarily mean anything has been done wrong, it can simply be an outcome of the research.

Research projects often involve higher levels of uncertainty and risk due to the exploration of unknown territories and the possibility of failure. There’s no guarantee of immediate practical outcomes and the undertaking may not always lead to a successful solution.

Defining an engineering AI project

An AI engineering project focuses on the design, development, implementation and maintenance of AI systems, based on established principles, specifications and requirements. These projects may include tasks such as data collection and pre-processing, model selection and training, software engineering, deployment, monitoring and ongoing optimization.

The main objective of this type of endeavor is to create and integrate AI-powered systems or applications that solve specific problems or enhance existing processes, without the need for research. At the project’s conclusion there should be evidence of a tangible use case, such as deploying the technology developed in a mobile application.

As a result, an engineering project is typically more focused on delivering granular results within specific constraints, such as time, budget and functionality. While there are still risks involved, they are often more manageable and predictable compared to research projects.

It may turn out that the technology does not work, for example if there is not enough data to train an appropriate model. However, the engineering team is not expected to come up with a solution to fix the issue or to determine precisely why there is a problem – a research project should then begin to investigate the outcome.

Putting the differences into practice

Companies often initiate a research AI project with expectations mirroring an engineering venture and assume that hiring data scientists will deliver a fully functional model within a specified timeframe.

In such instances, managers start to doubt the endeavor when the model inevitably falls short of initial targets, questioning whether the failure stems from the project’s feasibility, the team’s competence or the model’s suitability for the particular use case. In attempts to resolve the issue and demonstrate progress to the leadership team, developers frequently overlook the fundamental distinction: they are engaged in research, not engineering.

The intended outcome is not a functional model ready for production, but rather a comprehensive report outlining what works and how it operates. Based on this information they can proceed to building solution architecture and start an engineering project designed to make the solution production-ready.

In contrast, there are situations where a lot of engineering work needs to be done, but the project is treated as research. Such a situation can occur when the research that needs to be conducted is demanding in terms of resources, infrastructure, and the complexity of the problem.

It is necessary to prepare appropriate research infrastructure to ensure results can be easily reproduced, data can be provided quickly and is of high quality, new ideas can be implemented at a quick pace, and testing is reliable. The construction of such infrastructure is not usually a suitable task for researchers, who are often unaware of how demanding an undertaking this can be.

A situation can also arise where research seems to be progressing, but the results from previous tests are constantly being questioned, making it difficult to carry out further tests because the system is unstable. This highlights exactly why aligning expectations and understanding the nature of the project is crucial for its success.

Where businesses will fail is if they underestimate the initial phase of the project. It’s crucial that they avoid diving headfirst into hands-on work and instead pause to consider which personnel are best suited to the needs of the project, what they hope to achieve, and the different methods they will have to employ. This will make the type of task they are faced with abundantly clear, allowing them to plan accordingly and align their expectations with the project type.

Recommended: Fujitsu SX Survey Reveals New Data-driven Sustainability Goals

Make a value judgment 

Without first identifying whether an AI project falls into the category of research or engineering, developers cannot accurately determine what they hope to achieve. It’s rare that a project aligns perfectly with either definition, but putting measures in place to understand where it lies on the scale is the first step to setting realistic expectations.

[To share your insights with us as part of editorial or sponsored content, please write to sghosh@martechseries.com]