India’s artificial intelligence (AI) ambitions have led to a surge in demand for software developers, data scientists and machine learning engineers.
But as more companies deploy AI across banking, healthcare, customer service, and other sectors, a new talent shortage may be emerging: the lack of experts who can test these systems to ensure they are accurate, secure, and reliable before they are delivered to users.
Harry Rao, Founder and CEO of TestGrid, believes that discussions about India’s AI workforce are primarily focused on building AI systems, overlooking the equally important task of evaluating AI systems.
“When you hear about India’s AI talent needs, the conversation often centers around software developers, data scientists, and machine learning (ML) engineers. That focus is understandable,” Rao said.
He pointed to NASSCOM’s 2023 State of Data Science and AI Skills in India report, which predicted that the demand for AI and data science professionals in the Indian technology sector will increase from about 629,000 in 2022 to about 1.13 million by 2026.
“But building AI systems is only part of the workforce challenge,” Rao says.
“The second requirement is becoming harder to ignore. Organizations need experts who can determine whether a system is accurate, secure, fair, and reliable enough for real-world use.”
According to Rao, adoption of AI by companies is outpacing their ability to evaluate AI, creating what could be India’s next tech talent gap.
“What we’re seeing across corporate quality teams is that AI adoption is outpacing many companies’ AI evaluation capabilities. India, in turn, may be on the verge of facing a new talent shortage in AI quality engineering,” he said.
Why testing AI is different from testing software
Traditional software testing follows a relatively predictable process. When the user performs an action, the software returns the expected result and the tester verifies whether it meets the predefined requirements.
However, AI systems behave differently.
The same prompt can generate different responses, and while those answers may seem convincing, they may contain factual inaccuracies or unsubstantiated claims. Performance can also change over time as models are updated, source data evolves, and user behavior changes.
This means quality teams can no longer rely solely on traditional testing and test automation practices.
Instead, you need to assess whether your training and evaluation data is representative, current, and relevant, whether your AI models treat similar users consistently, whether attackers can manipulate prompts to bypass safeguards or leak sensitive information, when decisions should be escalated to humans, and how your organization will detect performance degradation after deployment.
“All these questions cannot be answered at the final stage of interface testing,” Rao said.
“Quality engineering must cover the entire AI application, including data, model behavior, interfaces, safety measures, and post-release monitoring.”
India’s AI challenge goes beyond models
India has unique challenges when it comes to AI quality assurance due to language diversity, regional differences, and varying levels of digital infrastructure.
According to Rao, organizations need test cases that reflect how people actually use AI systems, rather than ideal scenarios.
For example, financial assistants must understand requests that combine Hindi and English, distinguish between balance inquiries and fraudulent transaction complaints, and accurately interpret everyday language even when users misspell words or describe problems informally.
Healthcare applications require even greater scrutiny.
AI-powered medical summary systems can generate sophisticated reports while unintentionally omitting allergies, dosage changes, or other clinically important findings. Evaluating such systems requires medical expertise, clearly defined safety standards, and a process that allows uncertain cases to be escalated to qualified experts.
Speech-based AI systems also need to be tested across regional accents, speech patterns, background noise, and different speech conditions. Consumer applications must continue to work reliably on affordable smartphones, outdated operating systems, and unstable mobile networks.
Why test automation alone is no longer enough
Test automation remains an important part of software quality engineering, but AI applications require a broader range of skills, Rao said.
Traditional quality engineering functions such as software reliability, performance testing, release quality, and test automation will continue to be important. However, AI quality engineering now requires experts to evaluate model accuracy, identify unsupported claims, assess bias across different populations, test adversarial prompts, discover security vulnerabilities, define human escalation rules, and monitor system performance after deployment.
According to Rao, this extends quality engineering beyond software testing to include data literacy, cybersecurity awareness, AI model evaluation, and domain expertise.
“AI quality engineering combines software testing with data literacy, security awareness, model evaluation, and domain judgment, capabilities that organizations may struggle to find,” he said.
The next big AI job may not be coding
Rao believes companies should start building these capabilities before AI systems start failing in production.
“That’s why we shouldn’t wait for AI systems to fail in production before building these capabilities,” he said.
He added that high-quality engineers should be involved much earlier in the AI development lifecycle and help define what the system should do, where it might cause harm, and how to measure acceptable performance.
He argued that simply hiring traditional software testers and adding “AI” to their job descriptions is not enough.
Organizations must decide which capabilities can be developed internally and where expertise in data, security, or AI model evaluation is needed. At the same time, universities and training institutions need to start treating AI assessment as a dedicated discipline rather than an extension of traditional software testing.
“India will continue to need people who can build AI systems. But deploying AI systems responsibly will also depend on experts who can test AI behavior, identify its limitations, and decide how much trust it deserves,” Rao said.
– end
