AI projects don’t fit neatly into traditional IT project management. AI development is continuous, data-centric and iterative, making it difficult to manage with traditional project methods.
Project management software uses AI to detect scheduling issues and resource constraints for traditional IT projects, but managing AI projects requires more than AI-assisted planning.
Organizations such as the Project Management Institute are beginning to define a project management methodology for AI, listing six different AI project management phases — business understanding, data understanding, data preparation, model development, model evaluation and model operationalization. Even so, CIOs still have little practical guidance on managing AI projects.
The question facing CIOs and project managers remains: What changes should be made to traditional project management methodology to accommodate the unique nature of AI?
Let’s take them one by one.
How AI projects change IT and business roles
AI system development is continuous and iterative. AI projects are also centered more on data than on applications. If the data that applications operate on isn’t good, the results won’t be, either. This changes the project dynamics in IT, as the primary go-to people for AI projects become data specialists, not application developers.
For business users, this change in dynamics also presents challenges, because it’s the subject-matter experts in end-user functions who must determine whether the data is correct. This forces end users into more active roles in IT-oriented projects than they are accustomed to.
Finally, tracking AI project progress can be frustrating because AI projects are evolutionary in nature and may never actually end, at least not in the conventional sense.
Choosing an AI model strategy
Clearly understanding the business use case for an AI system and the outcomes the company expects is the first step in AI project management. Determining the appropriate model development strategy for the AI project is the next challenge.
AI model development can be challenging because IT and users struggle to understand what it entails. IBM defines an AI model as “a program that has been trained on a set of data to recognize certain patterns or make certain decisions without further human intervention.”
However, depending on the business purpose of an AI system, the approach to model development can vary.
-
The model can be a set of algorithms that are programmatically defined to operate on a set of data by querying that data with specific questions.
-
Or it could incorporate elements of machine learning that are either highly supervised or not supervised at all.
-
Companies also have the choice of using prebuilt foundation models that roughly address the business issues they are trying to solve, with the option of customizing these prebuilt AI models for their own particular use cases.
To determine the best AI model, companies need an in-depth understanding of the business use case they are addressing. If the goal is to inject AI into what-if scenarios and financial forecasting based upon data the company already has under management, a set of algorithms for standard queries could fit the bill. If the company wants to improve cancer diagnosis, it might want its AI diagnostics system to look outward as well as internally at data, “learning” from worldwide symptoms and data to embellish what is known locally. If a company wants AI to assist it in an area where it lacks expertise (e.g., customer service), it can purchase a foundational customer service AI system that comes preconfigured with data and algorithms, which the company can customize over time to its own needs.
Infrastructure requirements and AI readiness
The other preparatory step that should be taken before any AI project gets the go-ahead is to assess staff expertise and IT infrastructure readiness.
If the existing IT infrastructure isn’t sufficiently robust to support AI data and processing, one option is to host the AI system in the cloud, where resources can be scaled, assuming budget dollars support this.
The bigger question concerns IT and end-user readiness for AI.
On the IT side, data analysts already have a strong background in data cleanup, preparation and data technologies like extract, transform and load. Data analysts know how to normalize data so it can move across systems through APIs and seamlessly exist in hybrid data repositories.
However, on the AI model development side, there is bound to be a gap between what IT developers and end users know and what AI model development demands. AI model development requires skills in algorithm development and even in statistical analysis. Data scientists have these skills, but IT developers might not.
Then, there is the AI model training itself. On the user side, model training must be done by subject-matter experts — and that training must be vigilant and ongoing, so the AI model and its outcomes don’t drift and lose contextual accuracy over time.
The only way to ensure quality and continuing evolution of an AI model is for end users and IT to work closely together over the long haul. This is a departure from traditional IT project management, which at some point declares a project over and has everyone go their own way.
Deploy AI gradually
When it’s time to bring AI into production, the No. 1 goal should be to automate selected business workflow steps rather than the entire workflow. This helps ensure the initial success of an AI project deployment because, as business workflows are altered, human job responsibilities also change. This can upset users and derail project progress and trust. Proceeding at a digestible rate of workflow change is the best course for AI project success.
There is an additional rationale for gradual workflow change in AI projects: It is vital to have humans in the loop because the AI might not be right. AI systems can produce unreliable results when trained on skewed or biased data and can also hallucinate. “I recently fully automated an AI failover system in my data center,” one CIO told me, “but when it comes to activating an actual failover, I still want to be the one looking at the data and pushing the button.”
The bottom line for managing AI projects
The project management methodology for AI projects is still evolving and few project management software systems address AI’s unique demands. This places the burden of AI project management squarely on the shoulders of CIOs and project managers.
We do know several things:
Accountability is just as important for AI projects as it is for traditional IT projects. Someone must be in charge and ready to call the shots. That person should consistently communicate with team members and C-level management on project progress.
AI projects are not like traditional IT projects. In fact, these projects might not end until the business use cases for them expire. Project members and the C-level need to accept this reality upfront.
It’s best to proceed at a gradual pace with AI projects. People are learning as they are doing, which calls for care. AI projects should target small, tightly constructed business use cases with clear and achievable goals.
Project task schedules should also include tasks for IT and end-user education and training. CIOs and their teams should accept that initial AI projects might be a mixed bag of successes and failures that you learn from — and upper management should share that understanding.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.