Capturing the promise of artificial intelligence (AI) will require that companies understand that their current software support infrastructure is inadequate. Right now – to the extent that AI exists in organizations – it exists in small models in distinct units. These models are being created using unique data sets, targeted toward particular outcomes. However, taking AI to scale and putting it in operationally critical places will require a more robust approach organizationally and therein lies the management challenge.
For example, today if you shop on Amazon, the company’s AI models will give you three related items to purchase – the risk of their being wrong is small to both the organization and the individual consumer. However, when you are running or building gas pipelines or chemical manufacturing plants, the risks and dangers become much more significant if AI models in those places are wrong.
As we move AI models from small department level models to models that predict and decide in a way that removes important manual work (automating), we change the requirements of the model. Models will need to be more real-time, consistently accurate and available. In short, they will need to act like large-scale enterprise resource software commonly known as tier one software.
The challenge for most organizations and their senior leadership will be scaling up from small, bespoke department models to enterprise-wide models. If companies just take all the small models at the department level and put them into production across the business, you could end up with multiple models doing similar things. Also, because models are, in essence, small pieces of software, without a strategy, you end up with a lot of custom software that needs to be supported and maintained. If this software is automating core business decision areas, this becomes a serious problem.
Furthermore, models are like software on steroids –not only do you need to know that if you ping it, you will get a response, you need to decide whether you should use the response that you get. However, that ability to know whether you should trust the response lies in details of the data. For example, is it the same as the data the model was built on?
Finally, unlike software, for a model to be really good at solving a specific problem, it may have limited generalizability. As a result, there will be strategic decisions and tradeoffs to be made about how when and where to train the model to solve a specific problem exactly and where to generalize and maybe lose some accuracy.
For that reason, the long-term care and feeding of the models themselves will be critical. This represents a new and significant management challenge as senior executives struggle with the new technology.
Like other revolutionary technology advancements in the enterprise, the business value will only be realized if companies look beyond the technology, to the integration, maintenance and quality process required to scale business impact. Teams will need to develop tools and standards and adhere to appropriate testing and integration processes and develop governance structures that oversee these disciplines. In other words, the use of AI will need to become another well-managed business procedure as it matures in order to have impact enterprise-wide.