Over the 9 years we've been building and deploying AI solutions for clients, we've formed a deep understanding of the sheer variability in an organisation's needs and the outcomes they desire from automating their business operations with AI.

And, there is added complexity because speed, accuracy and cost can have several meanings in AI so it is important that everyone is on the same page from the start.

Bringing a successful AI project into life has two main fields of quality that we consider at Deeper Insights. One is the engineering side and the other is in terms of the quality of the models that we apply.

For example, we can talk about speed in terms of building the solution (how many months it takes) or speed of the predictions of the model.

More complex models can take longer to make predictions but usually have better accuracy. Lighter models can be very fast, but the quality is not as high.

AI models do have constraints. For example, one typically has to choose a model from several families of models, and each has its own strengths and weaknesses. Some models can be simplified so much that they give an output almost instantly (fast inference). Other models need to quickly adapt/learn to fast changing environments which can result in lower accuracy too. Finally, models can be made interpretable by humans, and when that is the case the accuracy can be affected.

Some examples of the decisions can be between:

• accuracy vs speed of inference
• accuracy vs speed of learning
• accuracy vs explainability

A successful AI project cannot be scalable and maintainable if there is not good software infrastructure to support it. The software delivery process is bounded by 3 main constraints: speed, scope and cost. Projects are usually requested to be built with high speed, high quality, or low cost. High speed can require a bigger team (which results in higher costs) or to lower the quality (cut some corners e.g. no scalability). If you want a low cost infrastructure, you need to be ready to give away speed or quality of the approach.

Deeper Insights work with all stakeholders to clearly define the scope of the project that has maximum impact in the client's business. Having a clearly defined scope upfront is a great way to have our teams aligned with the business goals.

Both infrastructure and AI model constraints are linked to the business needs. Very fast models can be required for Fraud Detection for credit card transactions. Building a fast infrastructure with lower quality can be useful for prototyping.

To summarise, the quality of Machine Learning models does not rely only on the choice of the models. It strongly relates to the quality of data as well.

See related articles:

Starting your AI project begins with data preparedness