AI models: Build, Buy, or Borrow

AI models: Build, Buy, or Borrow

One of the most common questions I get from clients is the best approach to building the AI components of a given solution i.e. the core algorithmic model. This question breaks down into the age old tripartite decision framework of Build, Buy, or Borrow. In the case of AI solutions this would be more accurately phrased as, “Do I…

  • build custom models from scratch, or
  • buy an out of the box solution, or
  • borrow solution elements from similar use cases to accelerate development.”

There is a time and a place for building models from scratch, developing models with the help of accelerators, and deploying OOTB analytics solutions, with the best answer depending on the industry, use case, and analytics maturity of the enterprise in question. So then, where do we start?

To begin with we need to start by understanding AI solution opportunity cost. This concept focuses on the differential costs of AI solution adoption depending on the maturity of an organisation. Essentially, late entrants to the market must pay higher costs even when they use OOTB solutions owing to snowballing licencing fees and missing the opportunity to build internal capabilities. This is not a new phenomenon but a common pattern when new technologies are introduced to an industry.

No alt text provided for this image
AI Solution Opportunity Cost

????????????????????????????????????

Credit Risk AI models are a great example as there is a spectrum of solutions from OOTB to in-house built models but regardless the opportunity cost for the new entrant is high even when they use OOTB solutions like FICO or SAS. This will come from ongoing and licencing fees but also critically being limited to the OOTB solution with no flexibility to develop custom insights.

In case of genuinely new AI solutions, building custom models from scratch is doable using open source and generic cloud services. Given the novel area of application business impact from any solution may not be certain. Only where the project concerned is highly experimental in nature, rather than being driven by a need for actual proven business impact, is a pure scratch built model approach recommended.

Building models with accelerators, such as previously trained models or pre-defined data sets, is the natural path to maturity but worth noting that there is no “free lunch” and costs are added as these accelerators bring material benefits. For instance, marketing agencies and credit rating agencies offer very good accelerators whether that be an AI model for prediction and even proven pre-defined datasets that are known to work but these do not come for free.

OOTB models are especially useful where an enterprise has relatively low analytics maturity and where the use cases, they are interested in are already well established and proven. With OOTB solutions effectively providing a “plug and play” environment even non-technical resources can devise meaningful insights.

In summary, there is value in each approach, and it depends on the particular circumstances in question:

- Custom built models, for any high performing company, will always be the secret source & sauce of competitive advantage but for new to analytics company it carries a high risk of failure and long lead time to business value realisation as it will take time to know the know-hows of what mode, what data and how that is used to make business impact.

- Using accelerators is the most common and effective way to build AI/ML models today allowing clients to build on proven solutions and tailor them to their requirements

- OOTB solutions are a great way to provide basic analytics insights for non-technical teams although their application is relatively narrow and limited.


Ultimately a mix of all three approaches will be relevant for the majority of clients as they move through their analytics maturity journey.


Finally I'd just like to thank Huw K. for, as always, being an amazing sounding board to bounce my ideas off.

要查看或添加评论,请登录

Huw Tindall的更多文章

  • The Rise of Open-source LLMs

    The Rise of Open-source LLMs

    Introduction It’s been a year since I wrote my first article on Generative AI so I thought it’d be worth penning my…

  • ChatGPT - we need to talk

    ChatGPT - we need to talk

    Intro Every man and their dog seems to have an opinion on ChatGPT and what it may mean for businesses, and society more…

    4 条评论
  • A Holistic Delivery Approach for AI: Part 1 - Introduction

    A Holistic Delivery Approach for AI: Part 1 - Introduction

    Welcome Welcome to the first in a five-part series where I will be outlining what I call the “Holistic Delivery…

    3 条评论
  • Top 10 Tips for a successful Strategy Project

    Top 10 Tips for a successful Strategy Project

    Top 10 things to consider in a Strategy Project | huw.tindall's Blog ‘Strategy’ is a very broad discipline.

    1 条评论
  • Collaboration and the benefits thereof

    Collaboration and the benefits thereof

    Preamble This short piece is about the benefits of collaboration in the modern workplace with a focus on Financial…

  • Luxury in India - Market Entry Strategy and Insights

    Luxury in India - Market Entry Strategy and Insights

    History repeats As it was for the European spice traders of the 16th century, the luxury market in India is a…

    5 条评论

社区洞察

其他会员也浏览了