Scaling AI isn’t a piece of cake, but there are strategies and concepts that can smooth the way. Keep reading to discover their potential and how to maximize it.
As Artificial Intelligence becomes part of most (if not every) industry, concerns over scaling it into production haunt organizations. If you wish to leave this bedeviled journey behind, read this article and put our advice into practice.
Scaling AI: what is it?
Artificial Intelligence (AI) is here to level up our performance. As a result, it brings more effective decision-making, facilitates customer interactions, improves operations, increases quality, and many others.
In an article published by Harvard Business Review, scaling AI refers “to how deeply and widely AI is integrated into an organization’s core product or service and business processes.” In their paper, Carnegie Mellon University states that “scalable AI is defined as the ability of algorithms, data, models, and infrastructure to operate at the size, speed, and complexity required for the mission.”
Bringing the benefits of AI to all fields of organizations is essential to promote business growth and get higher returns from investments. However, applying AI to one or two areas is easier than putting it into practice in the entire infrastructure.
There are enterprises that are already ahead in this journey, achieving success and advancing after many trials and failures. The secret potion hasn’t been found, but these best practices will facilitate the job.
1 – Keep in mind the bond between data and AI
Data is the best AI ally and should be the center of all business interactions. More often than not, treating data is a complex task requiring time, people, and software. Investing in good data management is worth gold.
Besides providing organizations with insights about their own performance, internal and external data generates quality predictive analytics. They are essential for AI, as it runs based on the input they give.
Due to data management, governance, and quality complexity, new companies have emerged as data labelers (such as SCALE AI and SAMA). Their job is to collect, analyze and classify data, leaving it “clean” for use.
Nonetheless, building a data-driven culture inside the organization is essential to take full advantage of it. Teams shall prepare to use it to reach business goals and measure performance.
2- Use Machine Learning Operations (MLOps)
Operationalizing AI is likely to bring about some bitter-sweet moments. As a result, organizations evolved from DevOps and established Machine Learning Operations (MLOps) with the view to standardize and automate processes.
Since scaling AI involves many teams, from data scientists to software engineers (we will elaborate on this next), applying standardized processes is vital. Thanks to them, teams monitor processes and recommend steps to follow on every occasion, reducing uncertainty and time-consuming tasks.
Building teams is another key feature of MLOps. Instead of having them work in separate departments, team up their skills. The organization of these teams should fall in accordance with two models:
- Pod model: this model puts together a small team of software engineers, data scientists, and data engineers. As a result, this is better for fast implementation but can create knowledge silos.
- Center of Excellence (CoE): here, all data scientists group up and get allocated to teams according to needs and resources available. Implementation can take longer, but data and knowledge are more accurate. Red Hat, a company that provides open source solutions (such as cloud, Kubernetes, Linux), formed CoE that helped customers on AI/ML missions.
For the execution of MLOps, it is fundamental to consider the existing AI infrastructure and tools. Once again, teams should analyze what they already have and evolve from there. Tools that work in the cloud and on-premises are a solution worth considering.
3 – Invest when the conditions for success are there
Organizations sitting in front of AI show that continuously investing in it, even when the winds blow in other directions, is crucial. Scaling AI demands efforts in terms of people and processes, and those who stop financing them will fall short. Innovation, research, and development are the top areas where the focus should be at.
4 – Take one step at a time
Growing big and strong takes time. One of the initial pieces of advice is to look at the organization and define which areas are more important. After that, the focus must be on developing and bringing it to the next level. For example, one could consider customer service as one of the first requiring attention. Therefore, teams and tools end up directed to this domain. Analytics come in handy to determine customers’ behaviors and define approaches to future contacts or sales.
When taking small steps at a time, remember that repetitive tasks/releases can actually lead to faster execution.
Difficulties in scaling AI
Putting AI into practice requires skill and persistence. According to IBM, “90 percent of companies have difficulty scaling AI across their enterprise”, which means that many give up in the early stages.
The difficulties are:
- Data management: the volume of data required to work efficiently has led traditionally used relational databases to struggle. Being so, more flexible and scalable solutions like NoSQL databases emerged. Securing data is also a pressing problem that can have tremendous effects on a brand’s reputation if it is not done correctly. Here, the question is not only with sensitive information but also with non-sensitive data that can be exposed and, thanks to AI, be linked to users and customers.
- Technological demands: computer processing and storage for AI are very demanding. Algorithms are constantly working, combining information, analyzing data, and bringing new inputs. The technological structure to handle this must be big enough, fast, and reliable. Thus, an important share of the budget is allocated to this.
- Qualified people: the goal of AI is to help and boost peoples’ performance. Nevertheless, it requires important skills and qualifications to take the most out of it. Hiring qualified workers can be hard, so organizations must also consider upskilling existing personnel. In addition, employees must understand that this is only possible if a consistent system is applied. Leaders play a vital role in ensuring that steps are taken seriously and consistently.
- Deal with the unknown: uncertainty will always be present. Even with the best AI in place, unpredictable situations will challenge organizations. It might lead to restructuring business processes and new investments, but all organizations face these drawbacks. The secret is how we turn them into opportunities and evolve from there.
These problems contribute to the findings of IBM, which stated that “by the end of 2022, we estimate that just one out of four large companies will have moved beyond pilots to operational AI”.
Proven concepts and steps for scaling AI
Knowing where and how to start the implementation of AI, or in other words, moving from proof-of-concept to proof-of-point is one of the biggest challenges.
In its report, IBM considers an approach called “AI engineering and operations.” The emphasis is on 4 areas: design, deployment, monitoring, and embedding. The goal is to create user-friendly experiences that are fast and effective while analyzing all models and decisions made by AI.
Furthermore, they also elaborated on the importance of NLP (Natural Language Processing) as a way to include unstructured data and solve one of the main difficulties. NLP provides the opportunity to understand what people are communicating in blogs, social media, and articles and add these to databases and statistics, as it was done in the US Open in 2020.
Steps to follow:
To avoid pitfalls, the course of action to follow for those in the early stages of AI and for those that have started is:
- Take the first step: start by analyzing the data available, data management, and governance. Then, jump into action even if it is just small steps. Laying the basis for scaling is vital. To do so, enumerate all tasks and trials made, and don’t worry if you don’t succeed in the first or second experiment: patience and time are best allies.
- Make use of metrics and engineering principles: working with teams and sharing knowledge is a good start. As in engineering, take notes from all steps and measure risks and success to build new paths.
- Establish strong leaders: leaders have a fundamental role in guiding partners in achieving goals. However, customers and stakeholders play a vital role in the process, so consideration at all stages is a requirement.
- Document everything and monitor all steps: once again, in order to evaluate performance and move forward in scaling, registering every step, every conclusion, and every KPI is of vital importance. After this, the created models undergo testing to understand their scalability.
- Partner with others: startups and academics can be an interesting way to boost trust and transparency. If specialized in areas that we don’t master, they can provide knowledge, participating in different stages of the process.
These steps will guarantee that AI leaves theory behind and becomes a practice.
Take off into the future.
Despite the challenges of scaling AI, we have reached a point where AI is decisive for business growth. When conceiving software and/or a system, enterprises must immediately think about how it will allow AI to evolve.
Making data collection properly, by labeling it in the correct databases, is also central to avoiding security problems and facilitating machine learning.
The journey to scale AI and maximize its potential isn’t easy, but if you look further away, you will see it was worth the effort.