Enterprise AI investments have expanded across cloud platforms, data systems, copilots and generative AI tools. Yet many organisations still struggle to turn these investments into measurable business results. Fragmented data, scattered institutional knowledge and weak governance continue to slow AI deployment.
In this episode of Analytics Insight Conversations, host Priya Dialani speaks with Supratik Shankar, Co-Founder of DView, about the infrastructure enterprises need to make AI more useful. The discussion focused on unified knowledge layers, business context, governance, legacy systems and the path from AI experiments to measurable outcomes in AI models. They can also create more consistent results when models work. Here are the key excerpts:
I see enterprises investing heavily in AI models while giving less attention to the infrastructure that makes those models useful. A model brings intelligence, yet it needs access to an organisation’s institutional knowledge to solve enterprise problems effectively. I find fragmented data, scattered business knowledge and broken governance across many organisations. These issues make every new AI application solve the same data and context problems again. Enterprises that build the underlying infrastructure first can deploy AI solutions faster. They can also create more consistent results when models work with trusted business data and clearly defined organisational context.
I see a unified knowledge layer as a bridge between institutional data and the meaning that business users attach to that data. We connect different data sources without moving the underlying information. We then add metadata, business context, KPI definitions, relationships and organisational hierarchies. This gives AI systems a clearer understanding of what enterprise data actually means. It also helps models produce more consistent answers. Enterprises can preserve institutional knowledge in one connected layer instead of leaving it inside spreadsheets, emails, messages or individual stakeholders. That approach can improve trust and make valuable enterprise knowledge available to the right users.
I regularly see three major challenges. First, enterprises struggle with fragmented data across modern and legacy environments. Second, governance must control which information different users can access. Third, organisations often have conflicting definitions for the same KPI. A term such as LTV can carry different meanings across financial institutions and businesses. AI needs to understand which definition applies to a particular user and task. Legacy systems create another challenge since organisations may hold decades of information in older formats. Enterprises need modern data foundations that can support faster intelligence while preserving access to important historical information.
I believe enterprises should start with use cases that have clear economic value. Organisations do not need to solve every possible AI use case at once. They should identify areas where AI can improve revenue, reduce risk, lower operating costs or increase customer volume. Data definitions also need agreement before teams deploy AI solutions. The technology must fit into existing workflows rather than operate as an isolated copilot. Business and technology teams need clear ownership throughout deployment. Finally, organisations must measure results through specific outcomes such as time saved, revenue growth, risk reduction or lower operating expenses.
I see agentic AI moving into complex enterprise workflows where intelligence needs to interact with existing processes. That requires more than a capable model. Enterprises need reliable data infrastructure, a unified knowledge layer and governance that controls access based on roles and responsibilities. Agents also need human intervention where the workflow requires judgement or oversight. I would start with a focused business problem and establish the relevant data, KPI definitions and governance rules. Once those foundations work, organisations can connect AI intelligence with workflows. This approach can create measurable value instead of producing another isolated AI experiment.
Listen to the full discussion on the Analytics Insight Podcast.