Anyone who is even remotely connected with AI or has completed even minimal research will be aware of the following three alarming stats:
It was estimated that 85% of AI projects will fail and deliver erroneous outcomes through 2022. (Widely sighted Gartner report from 2018)
70% of companies report minimal or no impact from AI. (2019 MIT SMR-BCG Artificial Intelligence Global Executive Study)
87% of data science projects never make it into production. (Venturebeat article)
While some of these numbers might not be as high today as they were back when the research was published, the problem of wide ranging failure of AI projects persists even today. One of the reasons for this organizations tend to treat AI projects just like any other IT project like say application development.
The problem is AI projects are fundamentally different from any software engineering projects and hence have to be dealt with differently. In this blog, I outline what are these five differences that must be considered while planning for any AI projects.
- Fundamentally, AI projects are Data Science projects where there is no known end state. We don’t really know if the solution to the problem even exists or not. On the other hand, IT projects are engineering projects where we know the specifications and required end state. Naturally, it is easier to reach a known goal than discovering an unknown destination which may or may not exist. This is the fundamental difference between AI projects and IT development projects that most organizations miss.
- Normal projects need only a single code repository while for AI projects you need 3 repositories.
AI projects need
- A code repository,
- A model registry which contains details of which models you have tried
- Data version control system for keeping track of all the data you use for models and experiments
This makes managing AI projects very different from traditional projects.
- Going back to the fundamental difference between traditional IT projects and AI projects, because AI projects involve a lot of experimentation and unknowns, sprints don’t work for AI projects. The right way to manage AI projects is Kanban. Kanban, which literally meansbillboard in Japanese, started as an inventory control system for Toyota to minimize work in progress and to match the supply with demand, helps as it helps team visually track the progress and quickly adapt to changes.
- Another point of differentiation is that AI projects need specialized collaboration tools. Good machine learning models require rapid iteration, and ML engineers need the right tooling to accelerate their projects. This is not the case with traditional IT projects where standard collaboration tools can do.
- Last but not the least validation and testing of traditional IT projects is fairly straightforward as requirements and expectations are clear. In AI projects, data validation and testing are complicated as even improperly labelled data can have a large impact on the effectiveness of AI algorithms.
Many firms and teams commit the mistake of treating AI projects the same way as they do other IT projects, this results in project failures and low or no ROI. If one factors in these key differences between AI projects and traditional IT projects during the planning stage itself, a lot of future issues can be nipped in the bud.