AI Pitfalls Series – Traditional vs AI Projects

[Mattab]

A Journey Not A Destination

This blog is part of an ongoing series exploring some of the common challenges facing organisational AI projects. You can read the first one here.

“Build it and they will come.” – When it comes to AI, will they? There are many challenges which must be considered before building and integrating AI. Some are common across traditional software development, whereas some are unique to AI development itself. Analysis of these shared issues are well documented; Here are five examples which represent the greatest overlap between traditional and AI projects

1. Not starting with a real business problem is often the primary reason why software development projects fail. AI or otherwise, without the right starting point it is easy for an endeavour to lose track of what the overarching objective is.
2. Once the reason has been identified, it is important to build use cases and set the scope to put the right foundations in place for success. By doing so, you can avoid projects becoming unachievable as they gradually get more complex by trying to achieve too much.
3. “Good Data”, as opposed to Big Data, is the key to unlocking AI in a cost effective way. AI is not the tool to fix every problem, therefore when it is used inappropriately it can become expensive and complex.
4. Getting the right skills to identify, shape, build, test, deploy and support projects can be challenging (AI or otherwise).
5. Lastly, Integration. Without this, any system will become an island that is limited to the resources it has and disconnected from the outside world.

However, while these shared issues are well understood and mitigatable, there are several important ‘pitfalls’ to consider that are largely exclusive to AI:

Interesting science projects:

If an AI is built in the lab, it will live and die in the lab. Build AI projects against real business problems, using real business data and integrate it into business processes and pipelines. AI is not new, so you don’t need to prove the concept of AI (it’s already been proven!). Instead, prove AI against your business challenges on an environment that can scale and, importantly, has the capability to handle the most sensitive of corporate assets: data. If it can’t, your project runs the risk of locking data away in either organisational or data silosand blocking productivity. To circumnavigate this, aggregate and make the sum of the parts greater than the whole.

Governance:

Governance of technology innovation is a fine balance; too much and innovation dies, too little and you are just dabbling in science projects (see point 1).  Coordination is key and must be done at an organisational level to add value to help initiatives going, while providing the benefits proper governance offers.

Platform & Core Competence:

Unlike many software development projects, traditional IT infrastructure and the supporting software landscape is ill-equipped for modern AI workloads when it comes to both ‘training’ (building the models) and ‘inference’ (running the models). Having the right platforms is key to acceleration. As well, getting core competence in this space can be as challenging as new hardware; GPU acceleration, software management, experiment management, clustering, model control are all among the paradigm shifts that come with getting AI projects moving at pace. Without these competencies, projects will get bogged down in the technical complexity of integration, learning new technology, security, and model management, stopping team AI working collaboratively.

In my next article I will explore some of the key design decisions around buying, building and a hybrid approach. Stay tuned!