With AI agents from frontier AI companies going rogue, the next big challenge in agentic AI is evaluating whether an agent is actually carrying out the prescribed task. This gives global capability centres (GCCs) in India an impetus as they take greater ownership of global product development.
US-based market research company Fuel Cycle has opened a 50-person GCC in Navi Mumbai to build its AI platform. The company plans to use the India centre to build AI, engineering, and product capabilities for its decision intelligence platform over the next three to four years.
Daryush Laqab, Chief Product and AI Officer at Fuel Cycle and former NVIDIA and Google AI leader, tells AIM that the centre was set up primarily to build the engineering team needed for the company's AI ambitions.
“Fuel Cycle is about market research. I joined Fuel Cycle about three months ago with the intention of building an AI that would orchestrate and run the market research activities. What that all means is that I need a good, talented, rock-solid engineering team,” Laqab mentions.
AI Agents are Harder to Evaluate
Fuel Cycle is building an AI system designed to orchestrate parts of the market research process, from identifying research requirements and consumer groups to designing studies and interpreting results.
But Laqab says evaluating such a system is fundamentally different from testing a standalone AI model.
“Some of them are agentic, some are not, but they are building workflows. And right now, each one of those workflows needs to be manually tested, manually verified, and manually reasoned,” he explains.
The gap arises from a lack of an end-to-end evaluation framework for AI applications.
“But I feel like the market as a whole doesn't have a unified framework for looking at the end-to-end workflow. We have that framework for software, but not for AI applications,” Laqab adds.
For agentic workflows, checking for hallucinations or accuracy is only part of the problem. Enterprises also need to determine whether the agent followed the right process, captured the necessary information, and produced the intended outcome.
“When we are looking at the performance of an agentic flow, we need to look at it holistically, and that is where some of the challenges are,” he states.
Fuel Cycle is primarily using Anthropic models for its main application, while also experimenting with OpenAI and SpaceXAI models.
Laqab says the company can use one model to perform a task and another to evaluate its output. Fuel Cycle does not currently have a formal partnership with Anthropic, although Laqab says such a partnership is being considered.
The approach reflects a broader shift in enterprise AI—from evaluating individual model responses to testing entire agentic workflows. This is essential as 62% of enterprises actively experiment with AI agents, with 23% scaling agents in at least one business function, according to McKinsey’s State of AI report.
Beyond Engineering
The Navi Mumbai GCC will initially focus on AI engineers and specialists working on agents, prompts, and evaluation. Fuel Cycle also plans to hire forward-deployed engineers and forward-deployed researchers as the product moves closer to customers.
The company is initially relying on lateral hiring but plans to expand into campus recruitment.
“I really feel like the new talent that comes out of universities has the AI mindset. They have the mindset of developing with A,” Laqab remarks.
For now, Fuel Cycle plans to keep the GCC at around 50 engineers, effectively making it a nano GCC. But Laqab says the centre could expand as the AI product gains customers.
“Ultimately, it comes down to building the product, building the AI. Ship it to a few customers, validate it, learn how many to adjust it, bring it back,” he adds.
The company chose Navi Mumbai partly because its existing Indian engineering team was already concentrated there. “The nucleus of the team that we had in India was actually around Navi Mumbai, so it just made sense for us to actually bring that team together.”
For Fuel Cycle, the immediate bet is not on building a large GCC, but on using a small India team to solve a much larger problem: how to reliably test AI agents once they move beyond generating answers and start executing workflows.
Your reaction