The Agile AI for the Not So Huge Business
You can do more with less resources, if you use the right kind of AI for you. (Warning!: do not read if you have more than 100 data scientists in your organization.)
Zehra Cataltepe[1]
Highlight
While millions of dollars are spent in AI by big companies allowing them to have an unfair advantage, smaller companies can now avoid the mistakes made by those big players and benefit from more agile AI systems.
Some approaches to AI implementation feel like killing a fly with a sledgehammer. I also remember Shel Silverstein’s poem “… If you’re a bird, be an early early bird — But if you’re a worm, sleep late.” While there are hawks, eagles, turkeys, helicopters and planes of each industry, there are also other flying members of the same industry with a smaller market share and their own times and ways of flying. Those companies (or departments such as Commercial or Life compared to Personal Auto for example) just do not have the resources, such as armies of data scientists (See Figure below), huge amounts of data, computational and human power to maintain lots of deployed models, to undergo huge AI implementations. And guess what, they don’t need to!
Just because you have some budget, you should not waste your (actually, your customer’s) money on implementations that don’t fit your organization. It would be a pity!
So, what is the right kind of AI approach for those organizations?
- Agile NoCode AI
In order to deal with the lack of hundred people data scientist teams, we propose NoCode AutoML, supporting agile and collaborative business teams and also supporting MLOps so that, not experiments, but working models can beeasily deployed, monitored, and, if not right, torn apart or updated very easily! Business is a complex matter, especially in small and medium sized companies, without including the business users in every step of the operation, it is impossible to have the right kind of AI model. And let me be clear, by business user I mean anybody that is not considered part of a cost center in an organization, anybody that is responsible for with her decisions to increase or decrease revenue for the company. Such complexity requires the AI solutions to be understandable (by those business users) and I don’t mean the pdf kind of ‘understandable,’ but the excel kind of ‘understandable,’ interactive and actionable. We have seen wonderful actuaries and underwriters add features, revise training sets and updates on the models and increase retention rates by more than 8% in only 3 months with this kind of AI! Those people had scrum meetings (sometimes a bit longer than the brief meetings recommended) every day and worked with, understood, criticized and made better their own AI models. The whole team learned and used AI as much as they needed to operate. Just as you don’t need to be a mechanic to drive a car, you don’t have to be a data scientist to use AI.
2. Human in the Loop
In order to deal with the lack of data problem, we do not suggest that you go ahead and create huge data lakes or oceans! Make sure that you understand the value of each feature that you have or need to have by identifying the necessary business KPIs and the business KPIs achieved by the models. Acquire only the data that are beneficial or necessary for your AI models today or in the near future. What if you still do not have enough data? Well, do you have business users with domain knowledge, energy and will to help? If so, then you must use (business) understandable AI that can learn from those front line employees (this is also called human in the loop or semi supervised learning). These hybrid systems can learn from both human and data! Suddenly, even if you are not a large company with a dedicated large AI budget, you can have good and agile models driven by your race car driver business experts. For example, in the insurance field you can allow fraud or insurance experts to be iron-humans[2] with the ability to direct the AI machines for the maximum benefit of the organization. We have seen 77% reduction in labeled data requirements when labels requested by the AI were provided to such a system. This could mean your organization can quadruple its amount of data on a value terms! The business user can also help AI on what kind of data to not use, allowing huge reductions in costs and sometimes responsibilities of that data.
3. Continuous Learning
Data changes so fast, specially today. So, the usual (batch learning) practice is to deploy a trained and validated model, keep monitoring and when it goes bad, train another model and replace it. When business conditions and data change, it is likely to affect not just one model (your retention, risk, fraud and demand models might (and will) all be affected. This means your army of data scientists (if you have them) will have to run from one model to another trying to fix the models (this reminds me of the movie ‘Brazil’ with all the complex and malfunctioning pipes). We propose continuously learning and continuously validated (by the business users and automated processes) models and actually ensembles of them. As soon as the deployed model starts changing, you can proactively validate and deploy a better model, or for certain industries and use cases, you can keep training the deployed model itself.
Energy Efficiency
Thanks to the continuous learning technology’s ability to update easily with every instance, the computational requirements for the continuous learning systems can be 100 times[3] less than the batch learning systems! This means you don’t necessarily need tons of computational power. This is like when you have the affordable electric powered car available for you today, insisting on using a gas-guzzler. For God’s sake, we polluted this great planet enough, let us not repeat the same mistake we did with cars, with AI.
Conclusion
To close on our parallel, you were being told to buy an expensive supercharged SUV and drive it through the desert, obliging you to be your own mechanic, or even worse, have two support service truck vehicles with an army of data scientists and consultants. My suggestion is that you drive your own racing car on the right racing track, with the right-sized engine that you choose, and fueled by electricity, which is energy efficient.
In short, the AI you need is a lot more reachable, accessible and useful than you think. And after a couple of practices with your AI and the improvements based on it, “Oh, The Places You Will Go!”.
And I don’t have to have a dream, this dream-like reality is here at your service, today.
[1] Views and opinions are mine.
[2] My dream is to have the marketing budget to have Robert Downey, Jr. to drive one of our AI models. ;)
[3] The use case is a stock market prediction problem, updating models every second.
