Blog: Artificial Intelligence/ Machine Learning solutions: Build v/s Buy
One of the decision moments faced by leaders to solve a business challenge using Artificial Intelligence / Machine Learning solution is Build v/s Buy dilemma i.e. whether to develop the solution from scratch (either using internal teams or using external consultants) or to buy an existing off-the-shelf solution.
There are multiple cases where organisations tried to develop the solution when they should not have and failed to execute, eventually scrapping the use case or even instilling distrust about feasibility of ML in the organisation. Similarly, there are many cases where organizations purchased the ML solutions without much due diligence on different perspectives and failed to adopt the solution successfully or later replacing that. Because of nature of the ML / AI solutions, usual Build v/s Buy decision criteria or the intuition of experienced technology leaders used for other technology solutions won’t be appropriate here. I present six factors (not always mutually exclusive) and the associated framework that helps AI/ML leaders make the Build v/s Buy decision. Note this doesn’t address the choice between developing AI/ML solutions internally and outsourcing the development. This rather helps choose between fresh development and purchasing off-the-shelf customizable AI/ML solutions.
Applicability to Core Functions: AI / ML user cases are highly beneficial when applied to core function in the value chain. This will help solve very specific pain points in key process areas of the core functions. More the specificity, higher the adoption, usability and hence the benefits. For example, using anomaly detection techniques for automated monitoring of raw material composition to achieve consistent clinker quality is at the core of cement manufacturing companies and therefore should be developed in-house. Similarly, a company whose core business revolves around detection of bone fracture in x-ray images rather develop an advanced deep-learning based image processing solution in-house. On the other hand, sentiment analysis for a tissue manufacturing company, however important, is not applicable to core function of the organisation.
Developing these specific use cases and the full solution requires intricate knowledge of the internal processes. Also, these solutions needs to be tightly integrated with existing systems (v/s using a standalone product) for seamless and continued adoption. Therefore, these solutions that are applicable to core functions should be considered as key candidates for internal development.
Ability to create Competitive Edge: AI/ML solutions can be table stakes or bridge existing competitive gap or create competitive edge. AI/ML solutions that create competitive edge needs to be developed from scratch in-house. Netflix wouldn’t have any competitive edge by purchasing a product recommendation platform from the market. Netflix attained strong competitive edge because of gathering unique data and developing specialized models. On the other hand, Natural Language Processing based document matching solution for a CPG company is table stake and is a candidate for buying from the market.
Need to scale to other regions / business units: Some of the AI/ML solutions would be to solve a very specific problem in a region or a business unit. On the other hand, many AI/ML solutions once deployed for a region or a business unit need to be scaled to other regions / business units. Unlike normal technology solutions, scaling AI/ML solutions will require different kinds customizations and rework including retraining and optimizing the models with new data. Such solutions, if purchased from market, can be costly, complex and time consuming to scale.
Dependency on external network data: Few ML/AI solutions benefits from using private data aggregated from different sources. For example, a solution for early detection of Alzheimer’s disease might be using brain scan data aggregated from different hospitals. These solutions benefit from network effect. A healthcare company should rather purchase that solution than trying to develop inhouse unless it also has access to similar data. Similarly, ML solution that matches resumes to job description might benefit from millions of data points collected across organizations.
Company specific data for modelling: An AI/ML solution that is trained with non-company specific data can be used in the organization without retraining the models. Therefore, purchasing such solution from the market makes sense to save the overall cost and also reduces complexity of creating new models. For example, ML solution that analyses product reviews on eCommerce sites uses publicly available review comments on sites such as Amazon and Target does not require company specific work except for filtering review data specific to company products and thus can rather be purchased.
Dependency on complex algorithms for performance: Models are the heart of AI/ML solutions. If the performance of the solution can be regularly improved by finetuning the algorithms, such solutions should be developed inhouse. This allows the company to periodically experiment with new algorithms and improve the performance. For example, many of the forecasting solutions available in the market use traditional timeseries based models. However, forecasting accuracy can be improved in many cases using advanced techniques such as LSTM. Despite many breakthroughs so far, Artificial Intelligence domain is still in the nascent stage and useful inventions are ahead of us. Organisations need the flexibility for adopting to latest innovations in the domain which can only happen if the organisation can control and modify the algorithms behind the solutions.
Use case 1- Oil and Gas company that wants to simulate reservoir performance: Undoubtedly reservoir efficiency is at the core of Upstream operations and that of overall value chain for Oil and Gas companies. Even a small percentage improvement in reservoir efficiency can have big influence on the bottom line. This solution can benefit from network effect too i.e. using data collected from reservoirs across many companies. However, appropriate use of internal data and effective algorithms can help create strong competitive edge. Overall, the factors tilt towards internal development of the solution.
Use case 2 – Toys manufacturing company looking to develop a Natural Language Processing (NLP) based solution that extracts relevant insights from product reviews in eCommerce sites. While this solution is important to understand the consumer preferences, it is not applicable to the core business and also doesn’t not create a competitive edge (table stake). Company can certainly scale this solution to all regions where it operates. This solution uses publicly available data and does not use company specific data. Overall, the company should rather choose to purchase an off-the-shelf product review analysis solution from the market than attempting to develop internally.
Use case 3 – Credit card company that needs to develop a real-time fraud detection solution. Fraud detection is certainly a core domain in the whole value chain for the credit card company. Better fraud detection creates highly competitive advantage position and the lack of it increases risk for the company. This solution requires company specific data and specialized algorithms for better efficiency. While network data can further improve the accuracy of the model and fraud detection, overall this is a candidate for inhouse development.
Use case 4 – CPG company that needs trade promotion optimisation (TPO) solution. Trade promotion Management and Optimization is very critical component of the business for CPG companies and these companies spend 15% to 25% of total sales on Trade Promotion. While this solution isn’t part of manufacturing process which is the core function in the value chain, Trade Promotion is indeed very key and impactful area of the value chain. Using very effective algorithms to simulate and optimize the trade promotion can hugely impact the top line and also create competitive gap. Therefore, these solutions need to be developed and owned internally.
Above factors inherently consider leaders’ dilemma of resource allocation. Utopian organisation with unlimited resources obviously can choose to purchase from the market or develop from scratch and yet may not go wrong. However, in realistic scenarios judicious choice using the above framework will help create long term sustainable success with Artificial Intelligence / Machine Learning Solutions.