As Indian enterprises move from AI experimentation to real-world deployment, the differentiator will not be access to models alone. In this interview with CIONOW, Balaji Raghunathan, Data & AI Engineering Business Unit Leader at Sigmoid, explains why dependable AI outcomes require strong data foundations, robust architecture, disciplined token economics and governance that keeps pace with autonomous systems.
While the hype of AI in Indian enterprises has persisted, are tangible business results now becoming a reality?
It is a mix. New-age businesses, especially fintechs, e-commerce players and digital-native companies, are ahead in embedding AI directly into workflows. They have already built digital infrastructure, accumulated usable data and developed an analytics-led decision culture. AI is compounding those existing strengths.
In traditional enterprises, adoption is more uneven. There are established use cases in manufacturing-where AI assists frontline staff in identifying problems, identifying resolutions and expert support more rapidly. But there are more significant enterprise use cases developing in such areas as marketing, which includes segmentation, loyalty, attribution and personalization. These functions had both data and digital practices in place before GenAI arrived.
The important distinction is this: AI does not create maturity from scratch. It amplifies what is already there. Good practices compound, but poor data and weak processes compound too.
What can traditional enterprises learn from more mature AI adopters?
The lesson is not simply to deploy a model. It is to create a management discipline around data adoption and data quality.
I have seen FMCG leaders begin by asking sales teams a basic question: are you using the dashboards? But the conversation soon matured in subsequent weekly review meetings. Instead of measuring dashboard logins, leadership began asking: “Revenue is declining in this region. What does the data say, and what action are you taking?”
That shift matters. It moves the organisation from adoption theatre to decision accountability. It also exposes process gaps. Often, teams blame data quality, but the real issue is that the underlying business data was not updated consistently. The technology may be ready; the operating discipline may not be.
Most traditional companies have many of the data platform and analytics skills that can be utilized by an AI strategy. Typically what slows them down is governance, risk appetite, and high sensitivity to regulatory compliance needs to implement the insights they already have.
How should CIOs split AI budgets between data engineering and models?
There is no universal ratio because it depends on whether the organisation is building on an existing environment or starting afresh. In a greenfield initiative, data engineering can account for roughly 20–30 percent of the initial implementation spend. If data quality is poor, it can rise towards 40 percent.
Integration can consume an equally significant share. Enterprise AI has to connect to systems of record such as ERP, CRM, supply-chain and operational platforms. The model is only one part of the equation.
After deployment, the economics can reverse. In agentic systems, model and inference costs can rise quickly, particularly when agents repeatedly call models in autonomous loops. Some organisations exhaust token budgets far faster than expected because the system was not designed for cost control.
CIOs therefore need to invest not only in models, but in the architecture around the model: memory, caching, knowledge management, tool use and guardrails. This is the “harness” that makes an AI system useful, grounded and economically viable.
Will AI reduce the need for software engineers?
AI can reduce demand for repetitive support and maintenance work, such as routine ticket resolution or narrowly defined code fixes. But it increases the premium on deeper engineering capabilities.
Vibe coding can be useful for proofs of concept. It can help teams express an idea quickly and produce an initial version. But production-grade systems require strong architecture, systems engineering and design thinking. Teams must understand the “what” and “why” before AI can accelerate the “how.”
This is exactly where forward-deployed engineers and Agentic AI Engineers, fill a void. These are the specialists who bring a complex enterprise system into an artificial intelligence framework, operating within an organization’s existing structures with certainty to guarantee it works in the final, production setting. The future will not reward generic coding volume; it will reward engineering judgment.
Does GenAI eliminate the need for structured data?
No. GenAI has become valuable because it can work with the large volume of unstructured enterprise information like documents, images, audio, video and free-form text that traditional analytics could not easily use.
But real enterprise use cases require both structured and unstructured information. Consider a product-support scenario: an AI system may need to relate maintenance manuals and service notes with structured records on products, components, customers and incidents. Without those connections, it can summarize information, but it cannot reliably support decisions.
This is why data quality remains critical. There is also a requirement to have a holistic view across systems, a standard method for identifying entities and a semantic layer to give context to the data. Governed data is required, but AI ready data must extend to machines and agents relationships and context to the meaning of data.
How much data quality do you need to use AI?
Depends on your autonomy level. For Human-in-the-loop, AI suggests and a person makes the final call. You might be able to use as low as 75 percent of data quality for relatively low-stakes use cases.
The AI in cases involving ‘Human-on-the-loop’ will try to automatically process the transaction, but will escalate if something goes out of ordinary procedures. These applications generally need a higher level of accuracy of around 85-90%.
In cases pertaining to autonomous systems where the AI handles tasks with little or no human oversight, it requires the strongest level of accuracy possible as well as controls and oversight of the process.
The pitfall is thinking that everything must be fully autonomous from the start. Consider the right balance between automation and human involvement first and as accuracy and overall performance improves you can consider moving towards greater autonomy.
What is tokenomics, and why should CIOs care?
Tokenomics is becoming an important discipline for managing AI spend. It is not just about tracking how much an organisation spends on a model. It is about designing AI systems to deliver the required outcome at an economically viable cost.
The cost of an AI deployment is influenced by several factors such as the number and type of model calls, the size and frequency of prompts and responses, whether agents operate in repeated or autonomous loops, how efficiently the system uses memory, caching and knowledge retrieval, and the combination of cloud, SaaS, on-premises infrastructure and AI services.
Many organisations begin with a fixed enterprise tier or user licence and assume costs are contained. But providers are increasingly moving toward usage-linked pricing, particularly as high-volume users consume more compute. That makes cost visibility and disciplined architecture essential.
The CIO-CFO conversation must therefore move beyond “What does this AI tool cost?” to “What business value does this AI workflow deliver per unit of spend?”
What will define enterprise AI over the next 12 months?
The market is shifting from proofs of concept to deployment and productionisation. As enterprises have tested, piloted, and experimented withGenAI applications in the past two years, the upcoming period marks its transformation from a proof-of-concept into something that can run reliably and scale within a productive enterprise. GenAI productionization is however just the beginning of the journey.
With ever-more omnipresent agents, enterprises must address an array of potential pitfalls such as the risks associated with unplanned model usage, unforeseen cost impacts, exposure of company data, the generation of substandard quality content, and proactive decision-making that might deviate from company-wide strategy.
That is why AI governance and agentic governance will become central. Without them, enterprises may increase AI spending without generating proportionate value.
What should AI governance cover?
It will be up to the governance frameworks to determine safety, responsibility, the financial controls and fitness for the purpose of AI. In the short term, CIOs need to determine:
Which decisions will AI be allowed to recommend, act on, or escalate to a human.
Set cost guardrails for model use, agent loops and token consumption.
Monitor model performance, output quality and exceptions continuously.
Design and embed a sense of ownership across technology, data, risk, security, finance and business teams.
Implement AI ready data foundations, a semantic layer, ontology and/or Knowledge Graphs where applicable.
Measure value through operational and business outcomes, not pilot activity or usage volume.
The next phase of AI will not be won by the organisations running the most pilots. It will be won by those that can govern AI well enough to deploy it responsibly, control its economics and turn it into measurable business value.
