Build vs Buy — 3-Step Guide to Choosing Your AI Path
Ilia Badeev, Head of Data Science, Trevolution Group
For modern companies, being at the forefront of technological innovation is crucial. One missed opportunity, and the competition races past, taking your customers with them. Companies sprint towards innovation, as if their survival depends on it — because most of the time, it does. 2024 will firmly be the year of AI, and many more years to come as well. It is the new frontier of innovation, but should you buy into the innovation or build it yourself?
2023 only scratched the surface of what AI is capable of, with large language models, natural language processing, predictive analysis, and more. This year, AI will move from the news headlines to our daily lives, being implemented in everything from search engines and cellphone processors to customer service chatbots and voice assistants. An ever-deeper integration is inevitable as companies build their own or train their bought AIs.
Estimates by PwC suggest that AI technology could generate $15.7 trillion in revenue by 2030. And the AI market is set to grow by 26% in 2025, according to Tractica. So it only makes sense that, according to Gartner, 80% of large enterprises will use internal AI platforms by 2026.
Is building an internal AI realistic for most companies?
It is more than realistic because there aren’t just two options: completely in-house or ready-made. Like with anything, there are numerous options in between.
For analogy, if you want to make a pizza, you have several options: you can order a ready-made one; you can go to the supermarket, buy the ingredients, and cook it at home; or you could start by planting wheat and tomatoes to harvest and make a pizza. The last option seems absurd, but if you are a large company, it might be the most suitable for you.
The same applies to AI. You can take a fully ready product (which requires minimal expertise and maintenance), or you can start from scratch developing your mathematical algorithm (which requires maximum expertise). Of course, there are hundreds of additional options between these two extremes. Any AI platform (and AI product) consists of a large number of components (blocks), and a company can decide whether to take an already fully assembled constructor, just some of its pieces (and make the missing parts themselves), or make every necessary block on their own.
The question then arises: where on this scale does an AI become “internal”? Yes, 80% say they will use internal AI, but does internal just mean accessible within the internal infrastructure? How many of its “blocks” must be in the platform for it to become internal?
What are the choices?
Since there are many “middle-ground” options, any company can choose their AI version suitable for their budget, time, and expertise. In practice, I expect that most companies will use ready-made out-of-the-box AI solutions from major providers: AWS, GCP, Azure, with a minimal level of customization for internal needs.
Companies that want customization will not do everything from scratch. Those wanting to use their internal AI will take a base model trained by a provider and further train it on their data. This is the optimal balance between time, cost, and the necessary expertise for development and customization of the final result.
Of course, one can go a step “deeper” and take an empty model (de facto a mathematical algorithm) and train it from scratch. This requires more time, money, and expertise but also provides practically the maximum level of final customization. Each company can choose its path.
Why customize at all?
Why do some want customization? Simply because a ready-made product cannot be a “jack of all trades” that will perfectly fit everyone. Every company has its business processes, often differing from other companies. Accordingly, to apply AI, they need to either change their business processes or adapt AI to their processes. And again, there are many middle-ground options. Each option will cost its money, time, and expertise. Each has its pros and cons. Companies need to make a choice.
Companies choosing a high level of customization will be able to perfectly integrate AI into their existing business processes. They will create their own unique systems perfectly suited for their business, their way of doing things, and their data.
And, when it comes to AI, data, of course, is key. Commercial AI is built to show good results on any data. You, however, can make AI that shows perfect results on your data. For example, you’re making AI for analyzing audio calls. Ready providers give good results on any calls, but your AI (trained on your data) can give perfect results specifically on your calls.
When discussing bought-vs-built systems, we cannot forget about costs. The cost of developing your AI can be significantly more expensive outright in terms of money and time. But the “operation” of a customized AI down the line can be cheaper than a ready “out-of-the-box” one.
So, how to choose then?
It’s all about resources. To develop a truly customized AI, you need expertise. You need time. You need money. Almost always, you will need your data (and it’s not always easy/possible to prepare). You need a clear understanding of what you want from internal AI.
Most importantly, as in any AI project, the result is not guaranteed. AI models are non-deterministic, and the final outcome depends on many factors: the formulation and understanding of the problem, the quality of the dataset, data processing, hardware, final integration, and usage. But even if you have all these components perfected, the final result is still unpredictable and may not be accurate enough for your use case. Thus, you could spend months developing an internal platform, which turns out to be worse than existing ones (I will touch on how to deal with AI disappointments in my next articles).
If you have the resources, data, and you are not totally risk-averse, a customized AI, either partially or built completely from scratch, will undoubtedly serve your business, its needs, employees, and customers better than a generic store-bought version.
There will not be a one-size-fits-all situation when it comes to AI. And the AI that is best at one particular aspect or business or data may be completely unfit for others. So, to make sure you are about to make the best possible choice of AI, here’s a simple three step guide:
- Assess your business need. This is the most important step, and you should include as many stakeholders as possible in deciding why exactly do you need AI, what purpose will it serve and how will it help your business grow. The notion of AI can get ‘buzzwordy’, so don’t let the hype sway you and truly understand your situation and what would the AI be solving. This will in turn bring the most value to your customers.
- Run an audit of your data. It is the make-it-or-break-it key component of a successfully implemented AI. If you have bad, incomplete or flawed data (or data gathering methods), the AI’s results will reflect that. Take stock of your data and its funnels and remedy any potential problems before implementing any AI solutions.
- Seek help. AI is still a new frontier, full of potential pitfalls, so seeking the help of data scientists and AI consultants — AI consultancy, I predict, will be an emerging subsector in the years to come — is only logical. Explain your needs, present your data to them, and they’ll be able to guide you in the best possible direction of your AI journey.
And as a parting note — businesses nowadays chase AI and basically try to emulate each other in the way they implement their respective AIs (yes, sometimes plural). But as I have written before — despite the monstrous workload it can manage, AI is still just a tool. And to be successful in the age of AI, as in any age before that, you have to be creative and come up with novel and interesting ways to use the AI to solve the problems no one has thought of before.
Ilia Badeev is a data science expert, currently leading as Head of Data Science at Trevolution Group. His prior experience includes the role of Principal Data Scientist, developing innovative machine learning solutions like an automatic document verification system.
During Badeev’s tenure at Peregrine Technologies in Berlin, he implemented a range of projects from neural networks for traffic sign detection to sophisticated data analytics web services. His technical acumen spans across Python, TensorFlow, AWS, and Kubernetes, demonstrating his proficiency in handling complex data-driven challenges.
Earlier, at factory market in Berlin, he made notable contributions by developing ML models for retail analytics and transaction analysis, handling massive data sets. Badeev’s career is marked by his expertise in machine learning and data analytics, applied across diverse industries and challenging environments.
