6 steps to successfully implement AI project

While enterprises recognize the measurable business benefits of AI adoption, they don’t necessarily see the path to get there. As the Everest Group survey indicates, 3 out of 5 of enterprises fail to adopt AI and don’t achieve meaningful business outcomes. Let’s look for the missing key to harness the full potential of AI implementation.

In many cases, AI projects in enterprises are approached in the same way as other IT implementations, where focusing on selecting an experienced vendor and creating a precisely defined proof of concept are the key milestones. Projects carried out in this way usually end in the PoC phase and never reach deployment. The search for efficient solutions using data science or machine learning requires a different approach. The CRISP-DM (CRoss Industry Standard Process for Data Mining) methodology becomes a helpful tool. CRISP methodology assumes six stages of running AI projects that when repeated over and over again, lead to the right solution. This concept involves an agile approach, where priorities and pathways change as the project development progresses.

Phase 1 – business understanding

Understanding business needs is a key element of data science and the greatest challenge for AI projects. The further success of a project will depend on how deeply the given business use case is understood. The basis is a well-defined business problem that can be described in the language of data science. That is why at deepsense.ai we focus on close contact with business owners. From the very beginning of cooperation with clients, the entire team dedicated to the project participates in the discussion on business needs, possible solutions, and available data. This ensures that all team members have a broader awareness of the purpose, know the limitations, and are able to contribute both technically and conceptually.

Phase 2 – data understanding

On the basis of in-depth knowledge of the business problem that needs to be solved with the use of AI, we can proceed to the data review phase. This phase requires not only special commitment, but also self-confidence because the data received from the client will not always be able to provide valuable answers to a specific business use case. At deepsense.ai, we focus on full transparency and the readiness to modify the project scope if we see that the data will not provide valuable solutions in a given area.

It is difficult to assume in advance how long the data exploration phase will last, but both our experience and intuition are helpful here.

Phase 3 – data preparation

Once we make sure that the data provided by the client meets the business value, we are able to start preparing the data set. A very common challenge is not so much the quality of the data set, but problems related to its labeling. That is why, at deepsense.ai we pay special attention to data labels and always try to support the customer in this area.

Especially in computer vision, data labeling is an extremely subjective task. It sometimes happens that the subject matter experts do not agree with each other on how to label, or they do it in an unsystematic way. An interesting example was a quality assurance project for visual defect detection that we did for one of our clients. Unfortunately, even the experts involved in the project had difficulty in classifying the precise area and type of defects uniformly. At deepsense.si we deal with such situations and solve them based on the specificity of the problem or labeling costs, e.g. by preparing a detailed specification, multiple labeling to obtain consensus, or creating a dedicated application that interactively supports the labeling process.

Phase 4 – modeling

A mixture of thorough experience, efficient exploration, and agile approach is what matters most in this phase. At deepsense.ai we are always focused on optimizing the modeling phase by either proposing a proven approach or – in case of unusual challenges – by exploring many solutions in parallel and their quick selection. The best results are achieved by using a cascade approach, which allows us to spend more time refining the most promising models. We also keep in mind that “the best model is no model” and always try to make things simple whenever possible.

In this phase, we specify milestones and work within a strict regime, so as to provide a solution skeleton in the shortest possible time, the functionalities of which we will then further deepen.

In addition to that, we use neptune.ai – a tool that allows us to monitor experiments and make accurate decisions by tracing all the improvements (such as data preprocessing, model selection and training, validation strategy) and tracking their impact on the end results.

Phase 5 – evaluation

At this stage, we return to the business owners, proposing specific solutions and approaches for selected use cases. It is also the best time to review the project scope from a business perspective. It very often occurs that thanks to the experience and knowledge we have gained, together with the business owners, we discover new possibilities for AI implementation with significant business value.

A good example is a project carried out by deepsense.ai for one of the leading european fashion retailers. As part of our cooperation on online sales prediction, we were able to propose a number of UI solutions for the online shop to better track the customers experience and preferences.

Phase 6 – deployment

From the client’s point of view, the deployment phase is the most crucial; at this stage the proposed solution delivers tangible business results. Here, we try to actively support the client not only during implementation, but also in monitoring and maintenance operations.

We always have a flexible approach to project maintenance. Very often the volume of data to be analyzed increases over time or the project is extended to other departments in the organization and requires, for example, transferring it from on-prem to the cloud.

Summary

In the most advanced AI projects – which therefore have a chance to become the most innovative – there are no ready-made solutions and approaches. The data science team has to try novel methods, take risks and trust their experience. These are the elements we love the most about our work at deepsense.ai.