Pharma supply chains need clean data before AI, says Cosette India head

Cosette Pharmaceuticals is a US-based, fully integrated pharmaceutical company with a topicals and dermatology portfolio and a growing branded pharmaceuticals business, backed by a 100-plus year heritage and more than 350 employees across manufacturing sites in New Jersey and North Carolina. The company is backed by Avista Capital Partners.

Girish Bassavaraju leads Cosette's India operations as Country Head, Global Business Operations. He built the India team from scratch in 18 months, growing it from zero to 65 people across nine to ten departments, including IT, SAP, finance, regulatory affairs, pharmacovigilance and data analytics. He sits on the company's leadership team, with additional responsibility for M&A support, portfolio management and digital transformation across India and the US.

Bassavaraju discusses AI adoption in pharma operations, the real state of forecasting accuracy, why Cosette started its digital transformation with finance, and whether India's GCC talent pool gives the country a lasting edge.

You recently spoke about the industry moving from reactive firefighting to real-time visibility. What's driving that shift right now?
Pharma is a highly regulated industry, unlike an FMCG product, because people consume what we make. Patient safety comes first, with no compromise. Regulatory expectations are the second factor, largely driven by the US FDA, with some variation across European, Canadian and Japanese markets. Cost is a third factor, because without controlling it, an organisation cannot run. Real-time dashboards help leaders understand what is happening in the market and in internal operations at the same time, rather than discovering gaps after the fact. AI is what makes that visibility possible on a day-to-day basis, across people, operations, forecasts and inventory.

Where is AI delivering real value today versus where it's still mostly talked about?
AI itself is not new. What has changed in the last four to five years is that it brings together data that used to live in different systems and different languages into one company-level dashboard covering business, operational, people, quality and safety KPIs. That means a single missed KPI, such as attrition moving from 15% to 25%, immediately shows its impact elsewhere, such as on revenue. Stock visibility has also changed. Manufacturing teams can now see inventory levels directly instead of relying on emails or weekly meetings. I have also noticed a generational gap: people in their first few years out of graduation adapt to this data far faster than senior colleagues. To build confidence with leadership, we built case studies showing work that used to take ten hours being completed by AI in one hour. That is often what convinces sceptical leaders to invest.

How much can pharma actually trust AI forecasting today?
If you are forecasting one product, it is relatively simple, because experienced people already understand the trends. It becomes far harder once you are managing hundreds of SKUs across multiple geographies. That is where AI performs better than any individual, because it can stitch together patterns across that scale, including external market signals, not just internal data. But I would not recommend trusting AI completely. Use it as a first level of assessment, then apply your own experience on top of it. Today, accuracy is around 50%. It will likely move to 60, 70, 80% over the next three to four years as adoption matures, similar to how trust in UPI payments built up over time.

You've pointed to data foundations as a gap. Where does that show up most?
Resistance to change is actually not the biggest issue any more, because consumer AI tools like ChatGPT and Copilot have normalised the idea. The real challenge is that data is scattered, sitting in Excel sheets, physical notebooks, email and shared drives, sometimes without a proper ERP or even Office 365 in place. Cost is a second factor. Cloud licences can run to $200 to $300 a month, and tools like Copilot cost around $100 a month, which is expensive at an organisational level even if it feels cheap per user. A third factor is noise. Constant headlines about AI make leaders cautious, so companies often default to low-risk use cases like drafting emails or analysing spreadsheets rather than tackling core data at the SAP level.

How should companies sequence this? Fix the data first, or let AI adoption force the clean-up?
There is no perfect day to start, so somewhere you need to begin. Over the last two years at Cosette, we started with finance, because that data is the most robust and already reviewed by auditors. We have cleaned up around 50 to 60% of our finance data and built smart workflows around budgets, inventories and receivables and payables at the product and customer level. That gave our leadership confidence that this approach works. Manufacturing and production data is the logical second area, followed by the customer database for demand forecasting and engagement.

What is the toughest area for AI to handle?
Market intelligence. Pharma companies cannot always predict what the market will actually consume. During COVID-19, many organisations built capacity for COVID-related products, and that demand simply disappeared. We have seen something similar with semaglutide-based weight-loss drugs. A small number of companies built billion-dollar businesses on that molecule, and several generic manufacturers assumed similarly huge demand in India, investing heavily in production and marketing. Sales did not follow. Predicting market behaviour, rather than internal operations, remains the hardest problem for AI to solve.

What does a control tower give a company that a traditional ERP or dashboard doesn't?
A control tower brings an analytics team together in one place to look at demand, inventory, production, quality, suppliers, logistics and customer complaints simultaneously, rather than each function working in isolation. That gives far better visibility and the ability to anticipate problems, such as failing to deliver to a customer and facing a penalty. Consumers have come to expect real-time tracking from platforms like Amazon or Blinkit. Pharma delivery is more complex, because a batch can fail, raw materials can be delayed, or there can be a natural or political disruption. Having all of that information in one place allows a company to build contingency options rather than being caught off guard.

Does India's talent pool change what's achievable here versus a similar setup elsewhere?
This has been building for 20 to 25 years, starting with the IT talent built by companies like Infosys, TCS and Wipro. Engineering colleges have expanded well beyond the major cities into tier 2 and tier 3 towns, and the quality of that talent has improved. Cost is also a factor: an engineer starting in the US might cost $60,000 to $70,000, compared to roughly ₹6 to ₹10 lakh in India, an eight-to-tenfold difference. Add strong English proficiency and a willingness to work flexible hours, and you get the recipe behind India's GCC boom. Cosette's India operation is itself a GCC that has moved from a cost play into becoming more of an innovation centre. That said, the US still leads on fundamental research and innovation. India and China have the edge in large-scale operations and manufacturing.

Where does AI fit into cold chain and export logistics specifically?
Every pharmaceutical product has a temperature limit, and that limit does not change as a shipment moves from, say, Bangalore to Rajasthan, where temperatures can be 15 to 20 degrees higher. Logistics has to become far more data-driven to manage that, alongside route planning and the choice between air and sea freight. Over the last six to twelve months, we have also seen shipments rerouted because of the conflict in Iran, with more companies preferring air freight over sea freight despite the cost. AI is increasingly used to track temperature, route changes and shipment status in real time, and that data also feeds into managing insurance premiums, since greater uncertainty in transit pushes those costs up.

Three years from now, what will an AI-enabled pharma operations function look like compared to today?
Awareness is building fastest at the leadership level, where some leaders now worry that not becoming AI-comfortable could make their own role redundant. Repetitive tasks will increasingly be automated, freeing people to focus on higher-value work. I also expect employees at every level to become more aligned with organisational goals, because information that used to sit only with senior management will be far more visible across the organisation. Decisions will increasingly be based on data rather than gut feeling or experience alone.

What's the biggest misconception about AI in pharma that you'd want to correct?
That AI will not replace people. It can, and I am not going to deny that. Unless you build competitive skills around it, it may replace you. I think of it the same way as the shift from a basic phone to a smartphone, or how C and C++ programming used to be a premium skill and is now a baseline requirement. The question is not whether this transition happens, but whether people treat it as an opportunity to grow or resist it until their skills become obsolete.

Is anything Cosette itself is exploring on this front that you can share?
We are not trying to be first movers for the sake of it, but we are open to learning and investing. Over the last two years, we have built digital transformation across our commercial function, analytics, customer data and finance, combining AI tools with industry best practices. We have also recruited a Chief Information Officer specifically to build this capability. We run internal hackathons where good ideas get sponsored if they make sense within budget, and we try to build a culture where failure is treated as part of the process, not something to avoid.