The pace of AI is rapid, and for many organisations a detailed understanding and ability to develop the required AI services to address their business challenges is not a core strength. This is one of the reasons why vendors are introducing AI capabilities wrapped by a services model, let’s call this AI-as-a-Service (or AIaaS).
2 predominant types of as-a-Service models available:
- Use Case specific AI services capabilities
A use case specific as-a-Service example could be image recognition or speech to text conversion. These capabilities often generalise well across use cases and therefore can be useful to many organisations. This also means the vendors don’t need to train their algorithms multiple times for multiple clients, thereby increasing their development and maintenance costs. Use-case specific AI capabilities typically have significant investment from large vendors such as Amazon, Microsoft, IBM, Google, although there are other smaller players such as rev.ai entering the market.
If the capabilities match your use case well, then these capabilities can be very cost effective and quick to start working with. It can be possible to develop a proof of concept or demo offering internally using a 3rd party AIaaS capability in hours.
Due to the significant investment incurred by these companies, as new features are being made available continually, different vendors may provide the best capabilities over time. Traditional IT development is often hard coupled to external dependencies making change complex and time consuming. To provide flexibility in a fast-moving landscape, think carefully about how you architect and develop applications so that you can switch vendors quickly, should the need arise. To support this, think carefully about the commercials around AIaaS, ensure that there is flexibility in the use of services whether you believe you need them now or not.
- A platform as-a-Service where you can run your own machine learning models
These capabilities can be used where your use-case is unique to your business or perhaps industry, requires custom or private datasets, or requires a model that is not readily available off the shelf, and you don’t want to go through the process of setting up the required infrastructure (compute instances with GPUs, storage, networking, etc.)
required to train and run AI models. Examples of this kind of service are Google Colab, AWS Sagemaker and Azure Machine Learning. They provide a set of underlying data pipeline and machine learning capabilities that you can use to develop, train and run your own models on.
With many AI/ML use cases, an understanding of the data is as important as what machine learning algorithm you use, and therefore it’s also important to ensure that you consider the implications relating to data residency, security controls (such as VPC configuration, encryption, etc), and for large datasets the cost and time associated with transferring them from local storage to the cloud.
General AI services considerations
There are many startups in the AI services space. When looking for a partner, consider the following points. This list is not exhaustive, but will give you and your organisation a starting point for evaluating the marketplace:
- How robust is their business model? Many startups are getting significant investment purely because of AI, rather than because they have a solid business model or long-term customer agreements.
- How robust is their algorithm(s), and do they generalise to use cases beyond those that they demo? Algorithms can be made to look very clever when they are trained to a very specific dataset, but when deployed to the real world the algorithm can prove to be overfitted to the demo use-case and doesn’t work in the wider world.
- Should the unforeseen happen and the company closes, how do you ensure you have access to your data, and can you also get access to their underlying models so that you can continue with your business. This is good practice with any software, product or service vendor, and should already be baked into your existing commercial agreements.
Other areas to consider with AI startups, especially around legislation such as GDPR, is what controls they have around any data, do they require your original, labelled data to train, and do they use data from different organisations to make a more robust model. The answers to these need to be mixed with your own organisations policies and appetite for risk.
Using AI as a service can definitely give organisations a good starting point and to be able to develop applications more quickly. The key thing is whether the use cases provided, and the risks associated with the vendor, align with your organisations.