Companies across all industries are exploring the opportunities AI holds for their business models. While identifying AI use cases is challenging, actually implementing them adds a whole new layer of complexity.
In this process, one of the most fundamental questions you have to answer is whether to build or buy. As we’ll see in this post, the answer to this question is rarely a simple “either / or” decision.
Why buying AI as a service (AIaaS) seems to be the obvious choice
When talking about AI, most people think of the latest consumer applications, such as interacting with Amazon Echo or Google Duplex calling a restaurant to make a reservation.
Partly due to these well-known applications, you might naturally consider Amazon and Google as potential partners when planning your own AI investments.
In fact, your intuition wouldn’t be wrong, as there are several good reasons to buy AI as a service from a large AI service provider.
In general, building AI solutions requires four ingredients, which, in theory, can all be offered by external third-party AI service providers:
- Algorithms
- Data
- Computing power
- People
Algorithms:
Advances in AI are still largely driven by an academic community that’s very open about publishing and sharing algorithms (and even code).
Since many of the most sophisticated algorithms are available for free nowadays, algorithms do not play a major role in the build or buy decision process. Here’s a good GitHub repo that details many of the open-source algorithms out there, for many programming languages:
Data:
As long as most AI applications rely on supervised learning, large data pools are not only a prerequisite for the development of AI solutions, but the quality and amount of data oftentimes determines the accuracy and, thus, the commercial benefit of the respective AI model.
In some cases, AI service providers have managed to pool own and third-party data to a breadth and depth difficult to match for any individual enterprise. Examples where individual companies would find it difficult to develop better performing proprietary solutions include image classification or face detection.
Computing power:
Prior to the cloud era, computing power presented a significant barrier to entry. If you wanted to develop your proprietary AI solution, you had to build your own supercomputer — a time-consuming and costly endeavor.
Today, you can simply leverage the global infrastructure of AI service providers (Amazon Web Services and Microsoft being the biggest). It’s cheaper, more flexible, and you end up with reliable, scalable and globally available computing power.
People:
Despite recent growth, the academic community working on machine learning — the powerhouse of modern AI — has been quite small. As a consequence, only few seasoned professionals work in the field. The resulting AI talent war has been fought with escalating salary offers and has led to a concentration of AI expertise in a few companies such as Google and Facebook.
Due to this concentration of talent—as well as the difficulty and costs of assembling a top-notch AI team—you may find yourself better off collaborating with a large service provider.
Looking at these four ingredients for building AI solutions, it sounds like we can come up with only one conclusion: Buy! AI service providers can offer a cost-efficient and globally scalable infrastructure, more accurate models due to their access to more data, and they’re also able to provide you with cutting-edge AI expertise. So why should you even bother building your own proprietary AI solutions?
Why outsourcing does not work for all cases
While the decision to buy AI solutions seems obvious, there are some use-cases where relying on an external service provider is not a feasible option. All dominant AI service providers, for example, are American companies.
While all of them offer local cloud solutions, companies operating with particularly sensitive customer data may not be willing to use their services due to privacy concerns.
Similarly, you should be hesitant to rely on AI service providers when trying to create or maintain a competitive advantage based on your own proprietary datasets.
The value of AI models mainly lies within the data used to train, test, and calibrate the model, so carelessly sharing this source of value with third parties should be avoided.
Both of these examples show that relying on third party solutions alone will not be an option for most companies. The following is a basic framework that will help you decide strategically when to buy and when to build AI solutions in-house.
How to know when to build AI solutions in-house
There are a few key exceptions to keep in mind when considering hiring your own experts to build in-house AI solutions. As a rule of thumb, you can treat outsourcing as the default option unless you can affirm at least one of the following two statements:
- You have exclusive or preferred data access to relevant data pools
- You see a strategic business need in owning the AI model (i.e. it’s paramount to maintaining or creating competitive advantage)
The following figure illustrates how you can approach the build-or-buy decision, broken down into three main cases.
Build proprietary AI solutions:
There are almost no plug-and-play solutions in AI—new architectures need to be tailored to specific prediction problems, or existing architectures have to be calibrated to target datasets. If your company not only has exclusive access to valuable data, but also your business model relies on this data, building an in-house AI model is likely to be your preferred strategy.
For instance, the business models of global recruitment service companies like Hays or Michael Page rely on the hypothesis that they enable companies to efficiently hire better candidates.
Through their network, they can attract more high-caliber applicants from which they select the right employees based on their experiences.
Picking the right candidates is clearly a key competitive advantage, while having access to countless successful (and unsuccessful) hiring decisions gives them exclusive access to a dataset that external AI service providers can’t match.
If a global recruitment company intends to build an AI model that learns to predict whether a candidate will be a good fit for an open job position, developing an in-house AI model should be the strategy of choice.
Leverage partners to co-develop AI solutions:
Case 2 consists of two scenarios that naturally complement each other. If you have exclusive access to data but don’t use this data (fully) in your core business, there might be an opportunity to co-develop AI solutions with a strong partner (2A).
Conversely, if you’ve identified a clear strategic business need but lack the required data, you should also evaluate new ways of cooperation (2B).
For example, insurance companies would like to offer personalized insurance products at the point of need. However, they’re having a hard time properly personalizing policies and correctly identifying these points of need.
Retailers, on the other hand, usually have large datasets on customer transactions. If an insurance company knew that a customer recently started purchasing dog food, it could offer the customer a personalized dog surgery insurance policy at the right time.
The failure of today’s insurance companies to offer their products at the point of need is one reason why the industry still fears that a giant like Amazon could enter the market. In general, data pooling between companies does not need to take place on a customer data level, where GDPR rightfully protects privacy rights of the individual. But pooling can also be beneficial to all collaborators with regards to more general industry data (prices, volumes etc.).
Especially in light of cases 2A and 2B, where it’s non-trivial for companies to decide how to act, the importance of a data strategy becomes evident.
Companies should strategically define their future data requirements, while simultaneously planning how to get access to the required data pools (e.g. through acquisition or cooperation).
Buy external AI solutions:
Whenever you’ve identified neither a strategic business need nor exclusive/preferred access to relevant data, buying external AI solutions should be the default decision.
Correctly classifying invoices and automating invoice processing, for example, is not a competitive advantage for most companies. Thus, buying an external AI solution to reap the benefits from automation can improve efficiency without risking giving away valuable data assets.
Consult with experts before deciding on a build-or-buy strategy
Answering the build-or-buy question will always be a case-by-case decision. The three cases introduced above should give some guidance on when it is worthwhile to invest additional time to evaluate whether building a proprietary AI solution seems feasible.
Even if you decide to build and own an AI solution, you should consult external experts (including AI service providers).
Building an AI model in-house is not a decision for or against third party providers, but it allows you to sit and stay in the driver’s seat with regards to key processes that constitute the competitive advantage of your company.
Discuss this post on Hacker News.
Comments 0 Responses