The Challenges and Triumphs of AI Implementation: An Interview with Vishu Singhal
Artificial intelligence (AI) has transitioned from the realm of scientific experimentation to being a core component in the operations of many industries. However, the path to achieving meaningful business value through AI is often fraught with obstacles. In a conversation with Vishu Singhal, Partner at Artefact, we delve into the challenges that most AI initiatives face and the role of data in these projects.
Determining the Fitness of a Company’s Data for AI Initiatives
When asked about the criteria to assess whether a company’s data is suitable for AI projects, Vishu Singhal emphasizes a fitness-for-purpose approach. He explains that data doesn’t need to be flawless, but it must be representative and accessible. Singhal identifies three dimensions to assess data suitability: signal-to-noise ratio, operational relevance, and lineage. The signal-to-noise ratio evaluates if the data contains the necessary patterns to predict the desired outcome. Operational relevance considers if the data reflects the current reality and can be refreshed and monitored at scale. Lineage refers to understanding the origins of the data, which is fundamental to gaining executive trust in AI outputs.
Pitfalls Preventing AI Projects from Delivering Business Value
According to Singhal, the major obstacles to successful AI implementation are structural rather than technical. Data silos, or “fragmented truth” as he calls it, and the “prototype trap” where models are built on clean, manually prepared data that doesn’t reflect the real-world inconsistencies of live data, are among the most significant barriers. Another major hurdle is the lack of business change enablement. Singhal points out that even accurate AI models fail to generate value if the teams aren’t prepared or incentivized to act on the insights provided by AI.
Translating Business Questions into Solvable Data Problems
Singhal suggests a systematic approach to translating business questions into solvable data problems. He stresses the importance of clarity on the decision to be changed, the responsible person, and the potential implications of the decision, before commencing with any modelling. This clarity helps to reverse-engineer the data problem, reframing it from a predictive challenge to a more actionable one.
The Role of Data Quality, Governance, and Ownership in AI Deployments
Singhal underscores the strategic importance of data quality, governance, and ownership in successful AI deployments. He asserts that high-performing organizations treat data products like business products. These organizations have clear accountability for critical datasets, and their governance is light but explicit. In the context of AI, governance fosters trust, a critical factor for AI adoption.
Metrics for Measuring AI Success
Model accuracy isn’t the most significant metric at the executive level, according to Singhal. Instead, decision impact, adoption, operational stability, and value durability are the metrics that truly matter. He emphasizes that AI success isn’t a one-time event but a managed asset that must adapt as markets and behaviours change.
Working with Limited or Biased Data
Working with limited or biased data requires a careful approach. Singhal suggests asking these three questions: Is the bias understood and measurable? Is the downside of being wrong acceptable? Is there a path to learning? The goal, he says, is not to achieve perfection but to make responsible progress aligned with business reality.
In conclusion, Vishal Singhal’s insights offer a substantial understanding of the complexities involved in AI implementation. He reminds us that AI is not a standalone technology problem but a business transformation challenge that requires clarity, accountability, and execution. This interview provides valuable perspectives for organizations looking to leverage data for AI-driven impact.
Source: Here
