Back to the Wild West: Why is implementing AI models so difficult?
The rise of artificial intelligence is unstoppable. Not only are businesses appealing to AI more frequently than before to boost their digital formation, they also have the data, resources and access to the technology and partners they need to make this possible. Although many are raring to get started they nevertheless get stranded even before things are actually up and running. Fuelled by impatience, they lose track of their original course, or run into obstacles which cause their proof of concepts to constantly be shuffled to the bottom of the proverbial stack: the road to implementation suddenly appears much longer than originally assumed.
More and more businesses are becoming aware of the potential offered by AI. Artificial intelligence offers outstanding possibilities for promoting the innovative aspects of your business, but the benefits it can bring are even more important. A return on investment is a must. There are three ways to do this:
You can create revenue streams by selling enriched data or creating marketing models such as cross-selling and up-selling.
You can optimize your operational efficiency by making use of for instance predictive maintenance or by automating your manual processes.
Machine learning and AI can be used to detect fraud and eliminate risks, thus reducing costs.
Nevertheless, some businesses are still not interested in AI, for example because they do not have the data. Those who have heaps of data are sometimes tempted to immediately get to work, ambitiously and rashly drawing up all sorts of models. Unfortunately, things can go wrong during this phase and, in the end, nothing is implemented into the way of working.
The most important reason for this is because they are working from a technical perspective and are forgetting to listen to their business needs. There is a case, a problem, and the data team sets to work trying to resolve this problem by using artificial intelligence. However, it simply doesn't work like that. AI cannot transform your business until you start developing things from a business perspective, instead of from IT. Even the most promising proof of concept (PoC) is not of any value until it has been implemented. When the marketing department of a dealer builds a model to predict which car a prospective customer will buy, this will only lead to a revenue increase if it is actually used by the sales department.
Not involving the end user is a frequently made rookie mistake. You can't simply embrace AI as a panacea, immediately set up a gigantic infrastructure without a promising foundation and hope for instant success.
In many cases, a proof of concept is developed by someone with insufficient knowledge of programming. A data scientist is good at working with data by definition, but this does not mean that this person will write the best code. Many data scientists go into data science with a background in business intelligence and approach problems in a rather ad hoc manner, using things that may work in a PoC phase but are not suitable for implementation.
They will develop a PoC without taking into account the infrastructure, for example, in which everything is developed in the cloud, while the customer wants to run the model on-premise. When the time has come to deliver the results and the code the customer is often served a plate of spaghetti code that he can't unravel or implement.
When it comes down to AI there are a lot of similarities between our modern business landscape and the Gold Rush in the Wild West era. Everyone is jumping on the AI bandwagon and is starting to diligently mine their own data, in search of promised gold. There are no standards, and neither is there a golden rule that dictates how to integrate machine learning or AI projects into existing systems.
Even companies with gigantic teams that have already implemented several models are stuck in their way of working and continue to look for a framework. One could say that they have created a bottleneck, which makes it quite difficult for them to hand over projects.
A good beginning is half the work done
So, how do you respond wisely to these challenges? The successful implementation of an AI model starts with a well-considered launch. Right from the very beginning, the end user and the business must be involved in selecting the right cases. Will you choose a case with a great deal of visibility or will you optimize a process that may not be visible in the company, but with a high yield? It is important to bring these considerations into balance. They must deliver a successful and promising proof of concept that convinces the end user to free up sufficient budget and actually implement the models. In addition to this, on-premise support is indispensable from the relevant IT department, so that the way of working can be adequately coordinated to the local technical systems.
Above all, you need a framework for a design approach with standardized parameters. Which process will you choose? How can you involve the business optimally? What are the deliverables? Which data sources will you use, and which will be your coding standards?
When you base your solution on model environments like these you will be able to roll out projects and implement them much faster, work in a standardized manner and nevertheless integrate various solutions in the same way. Not only that, you will be able to provide functionality that makes it easier to roll out models via an API, guarantee scalability (taking into account the readability of the code) and be able to smoothly facilitate migrations.
This structured way of working has been transformed at Dataroots into a coded framework that serves as a common thread running through every project. Called Methodology as Code, this offers the team and the customer a strong foothold which immediately guarantees that the delivered projects will satisfy a specific quality standard and guarantee their error-free implementation, whether on-premise or in the cloud.