'The Data Is Already There': The Case for AI Layering to Avoid Reinventing the Wheel

In the corporate rush to adopt AI, companies often create more problems than they solve. By forcing employees to learn clunky new platforms that disrupt daily work, leaders inadvertently create barriers to adoption. Teams may then view new tools as either magical or broken, with little nuance in between. A more mature approach is emerging, however, one that makes enterprise AI "invisible" by embedding automation directly into the tools people already use.

This is a strategy championed by leaders like Ravi Kiran, a seasoned engineering leader and VP of Engineering at the AI firm KnackLabs. With a career focused on developing production-ready AI, he argues that the path to real productivity and ROI lies in a non-disruptive approach that delivers intelligence directly into existing systems.

“If employees don't adopt a new solution, you will never see the ROI," Kiran says. "The focus should be on building an agentic layer on top of the legacy systems they already use to help them automate parts of their workflow and make decisions faster." Such a bespoke approach ensures the AI fits a user's existing workflow, which maximizes acceptance. As messy, siloed data derails many enterprise AI initiatives, however, that task becomes more challenging.

• The messy data myth: To get around this common point of failure, this model creates surgical access to existing data, avoiding massive, multi-year data cleaning initiatives. "Most enterprises are already on legacy ERP systems," Kiran explains. "The data is already there. The key is to build an API layer for AI to communicate properly with those systems."

• Lighting up the ledgers: The real data problem, Kiran notes, is often internal company documentation. "Many employees don't even know this information exists because it's siloed and only a few people have access. For these kinds of documents, the first step is to digitize them before moving forward."

By unlocking this "dark data," valuable unstructured information becomes accessible for AI processing, which expands what the technology can achieve. Employee resistance shrinks when expectations are managed from the outset and systems are designed to augment a user's capabilities.

• A simple extension: In one project, Kiran describes how a layered AI system helped the Indian Army expedite the processing of hundreds of letters. "By installing a small Chrome extension in the browser, they could do their regular work normally," Kiran reports. "The extension reads the data, sends it to the on-premise LLM, gets the response, and fills in the summary and metadata." In this example, a non-disruptive tool delivered measurable gains without requiring users to learn a new application.

Beyond the immediate outcome, the project’s success demonstrated the framework’s resilience. Navigating client constraints, such as switching from a preferred open-source model to a public one due to geopolitical concerns, proved the formula’s adaptability. It also serves as a repeatable model for other clients in high-security, on-premise environments.

Ultimately, any framework's success must be measured in concrete business terms, and for Kiran, the most important metric is human hours saved. By layering AI, he says a team's productivity can double. The result is a clear and compelling return on investment.

This strategy also sidesteps the psychological trap of seeing AI as either perfect or useless, a key reason many tech workers remain skeptical of AI. "People are at two extremes: they either think AI is a magical wand or they think it's rubbish," Kiran concludes. "The goal is to get them in the middle by helping them think through the use cases that can be solved with AI. In a ten-step process, for example, at least six steps can be automated, but the other four still need human supervision." Transparent and pragmatic communication is essential for building trust, demystifying AI, and positioning it as a powerful augmentation tool rather than a replacement.