When we start talking about AI and AI solutions, it intrigues many, but then they lose interest at the moment we start discussing the technical details.
Let me try to keep this piece a more light-hearted to highlight what are the top 5 considerations I keep in mind when I try to find workable and deployable AI solutions for a problem that a business or an end-user face.
Please bear in mind, some of these approaches I picked up over the past 16 years from outstanding software designers as the best practices of when they design AI solutions, and some I had to improvise myself just by facing unexpected issues in my projects.
Many of such projects were the pioneer projects in companies we were trying to do for the first time. I usually focus more on the bigger picture than an immediate challenge a developer’s faces while developing the AI solution.
I am translating the considerations I put into an AI / ML solution in the form of these five questions that I ask myself when I build a solution/application that includes an AI / ML focus component.
- Is the use case ripe for an AI / ML solution?
- What benefit the use case is targeting, and what’s the ROI?
- How many source systems are involved in sourcing the data?
- What business process(es) these models would facilitate?
- Who is the target audience of the results coming out of the models?
Is the use case ripe for an AI / ML solution?
People confuse the need for AI and ML models with general analytics use cases. I have come across many projects or proposed projects where the defined rules-based analytics is already in place and is working fine. Still, the product owner is smitten with the lure of using AI or ML in the product and push the case for expanding the application with an ML model.
What they don’t realise is that the data needed for an ML model to work with accuracy or desired precision is far more arduous than a general batch driven, rule-based analytics application, or even a generic statistical model.
It becomes regrettable when people commit the cardinal sin of using a general mathematical model in their form and call it a machine learning or an AI Model.
What benefit the use case is targeting, and what’s the ROI?
Business benefit or ROI is a primary driver for many of the AI solution projects to get accepted by the management. If we’re not focussing on the right benefits, then the whole endeavor and the effort behind it is a waste.
In many companies, I have seen tech teams get overzealous with the new tech capabilities available at hand, be it elastic nature of compute cycles.
These new open-source libraries can make AI solutions easy, or they are just experiencing a FOMO moment where the rest of the industry is declaring wins by jumping to new tech stack, and they don’t want to be the laggards.
To join the industry in the mad rush of concurring the technological summit, sometimes tech teams drive the use cases hard without understanding the real ROI of the use case.
You can not use a sword to sew a shirt.
Use cases sometimes can be met with simplistic and readily achievable solutions than overhauling the whole tech operations of the company.
How many source systems are involved in sourcing the data?
Deploying an AI application is a simple task, but getting the whole process of data ingestion, to its consumption, to realise the use case’s real potential is tricky.
The biggest challenge for any data science project is the data that you are using. More the source systems involved, the more complex the data quality and consumption puzzles are.
Each data source introduces specific data quality issues, data assumptions, and basic data interpretation rules.
Such as date format may differ in different systems, and you may have to build pre-processing layers to make the entire data follow a standard set of norms.
Microservices architecture, like calls based on restful APIs and clean data onboarding processes, can help maintain good hygiene for these data-driven applications.
What business process(es) these models would facilitate?
One thing many product owners don’t realise is that how this new ML model will fit into an existing business process (if we are expanding a current capability with AI features)?
How would a new business process be built around the new product or service we’re going to develop under this ML model-driven application?
Some of the commonly missed points are:
- What maker/checker approvals are needed for the new results to be qualified as authentic.
- How to flag the false positives and for future auto filtering of these scenarios?
- How will the feedback loop function work to embed the learning from past mistakes?
- How will the explainer module be built if the use case needs the capability to explain the results?
- Who will approve of the overall accuracy/precision levels of the model?
Who is the target audience of the results coming out of these models?
The target audience of the results coming from these models drives how the presentation layer of the application would be built and deployed.
If it is another team of developers then the output can be just an API call that will provide the resulting scores to the other application, but if the stakeholders who needed the results were senior management then a whole dashboard layer has to be developed, and this will add effort and time.
In many cases, data scientists build machine learning models on their laptops, feed it with partial data from production systems, and declare victory on solving the business-critical problem.
Sometimes they pump the results in excel and employ creative charts of dashboards from excel and claim victory, proclaiming that they solved the critical business problem.
There is a lot of work still pending with the application, which includes piping of the real production data in either batch or real-time feed, and sometimes people choose to skip the whole QA process when delivering AI solutions.
Should we use AI solutions?
AI solutions are as good as we can use them to deliver value. Having a model that has simulated the problem and provided potential predictions for future trends is not enough. How this will translate into actions, workflows, and decisions is equally important to take the value out of it.
Remember, AI is just a tool to get the ends meet for business to find that sweet spot of operations or to find new business avenues.
As they say, an apparatus is as useful as the blacksmith, so focus on your business problem more than just the tool.
There could be more such parameters that might be critical for an AI / ML solution to work, but for my answers to the above mentioned five questions are crucial to know what I can offer as a solution to my stakeholders.
