With the explosion of interest in AI across businesses of all kinds, particularly in traditional industries, the emergence of new platforms and solutions have helped make AI more accessible. The latest versions of frameworks like TensorFlow and PyTorch have made model building more straightforward with clever abstraction hiding unnecessary technical complexity. The barrier to entry continues to decrease over time.

Despite this, it can still feel like working with AI is the exclusive domain of software engineers or data scientists. My experience, however, is that successful implementations are a team effort and require a broad range of skills and backgrounds. It is certainly important to have the relevant technical skills in the team, for reasons outlined below. That said, engineering a model and putting it into production are elements of a much longer process that benefits from a diversity of competencies pitching in together.

This tent is not a big tent and therefore not representative of the right approach to AI. Photo by Ben Duchac.

Non-technical team members can participate in the generation of training data; domain knowledge and familiarity with the business problem are an advantage in this work. Domain experts will generally have a better understanding of ‘ground truth’ than an engineer unfamiliar with the relevant business context. Similarly, when it comes to feature engineering, understanding the downstream business or product challenge is valuable and non-technical domain experts can play a key role in this.

Having a business or product stakeholder in the wider development team is crucial to keep model development aligned with the challenge it aims to solve. Without this, there is a risk of divergence or runaway complexity causing major problems later in the process. I have observed an example of this where the business case called for a simple binary classifier but due to misalignment between business stakeholders and the developing team, the work snowballed into a multi-class model with all of the additional complexity that brought with it. This led to a poorly performing model and ultimately, a suboptimal business outcome.

Even in emerging areas of interest such as AI bias, non-technical team members can play a crucial role. Challenging assumptions, providing test cases and working with technical teams on bias investigation and mitigation call for a diverse range of team members working together. Including this broader group also helps channel the wider organisation’s sometimes distracting enthusiasm around AI into useful output.

AI: Easy to spell, hard to implement correctly. Photo by Anas Alshanti.

Great training data and features alone will not prevent issues like over-utilization of computing resources, overfitting or other model design flaws; it goes without saying that it is necessary to have appropriate technical expertise involved. Apart from ensuring reliable, high quality engineering work, it is vital in tasks such as the validation, verification, measuring and managing performance of models. The more business or product critical the model, the more due diligence needs to be taken in these steps.

Where does this leave the role of the data leader? Think holistically about your AI competencies and keep an open mind when it comes to who can add value. Establish clear roles and responsibilities for team members across a project and nurture an environment for collaboration. Utilize the domain expertise you have within your teams. Keep key stakeholders in the loop and make someone accountable to ensure the project stays focussed on the business challenge at hand. When considering bias, take advantage of the diversity of thought available to you and remain open to new perspectives. If AI is a team effort, be the best coach you can be.

– Colm