Anthropic partners with Pratham to deploy AI assessment tool across Indian schools
Claude-powered formative feedback system targets exam preparation gaps in underserved communities.
Anthropic has partnered with Pratham Education Foundation to deploy a new AI-powered formative assessment tool across Indian schools, marking its first strategic AI lab partnership with one of the country’s largest education nonprofits.
The collaboration centers on Pratham’s Anytime Testing Machine (ATM), a system powered by Anthropic’s Claude model that generates curriculum-aligned questions, grades handwritten responses, and delivers personalized feedback. The tool is being piloted with 1,500 students across 20 schools and adapted for more than 5,000 learners in Pratham’s Second Chance program, which supports women preparing for India’s grade 10 board exam.
The partnership comes as Anthropic deepens its presence in India, now its second-largest market for Claude.ai. Nearly half of Claude usage in India involves computer and mathematical tasks, reflecting a developer ecosystem building production-grade AI systems at scale.
Addressing the grading bottleneck
Pratham has spent three decades working to close learning gaps across India. Yet even with large-scale programs validated through randomized controlled trials, including Teaching at the Right Level, one operational constraint persisted: grading.
In many classrooms, teachers manage 60 or more students, limiting their ability to provide individualized feedback on practice exams. The challenge is particularly acute in the Second Chance program, which serves women who left formal schooling and lack access to trained subject-matter instructors.
As Nishant Baghel, Director of Technology Innovations at Pratham and a visiting scientist at MIT Media Lab, says: “These women didn’t get a chance to practice the exam enough times because we didn't have enough people to grade all those answers.”
The Anytime Testing Machine was designed to address that constraint. Students write answers by hand, photograph them, and upload the images. The system converts the images to text and uses Claude to evaluate responses against structured rubrics, generating feedback on content, accuracy, and expression.
Pratham selected Claude after evaluating multiple models. Sravana Chandra, AI Lead at Pratham, says: “Claude did consistently well across tasks including question generation, checking the quality of generated questions, grading, and feedback. This, along with Anthropic's focus on safety and responsible AI, led us to choose Claude over other LLMs.”
Iteration and accuracy gains
Anthropic and Pratham teams met weekly over several months to calibrate the grading pipeline. According to Pratham, grading accuracy initially measured at around 30 percent against expert-reviewed benchmarks. Through iterative prompt engineering and evaluation design, that figure improved to roughly 80 percent alignment with subject-matter experts.
Chandra explains: “We implemented an LLM-as-a-judge framework, benchmarking model evaluations against internally developed golden datasets that were manually validated by subject matter experts.”
On question generation, the system now delivers 90 percent accuracy aligned to Bloom’s Taxonomy. The Claude-powered ATM has completed more than 1,500 student assessments across 20 schools, with plans to expand to 100 schools by the end of 2026. The Second Chance program, which currently serves 15,000 women preparing for the grade 10 exam, plans to migrate fully to the Claude-powered pipeline by the end of 2026.
Multilingual capability has also been central to deployment. Chandra says: “Given Claude’s linguistic capabilities, it was readily able to handle feedback generation in a mix of Hindi and English, using English terms where relevant (such as for scientific terms) while keeping most of the text in Hindi.”
Keeping teachers at the center
A core principle of the design is that teachers review and can refine AI-generated feedback before it reaches students. Chandra notes: “Teachers feel empowered because they remain the final evaluators who validate AI feedback before it reaches the student.”
Baghel adds: “Human agency is something we focus on. Instead of asking what AI can do without teachers, we ask: how can AI help our teachers and students exactly where they're stuck?”
Madhav Chavan, Co-founder of Pratham, says: “AI tools like Claude give us a way to reimagine learning for students who do not have access to advanced educational resources. In addition to providing personalized support to understand the textbook, the ATM innovation will help children verify and authenticate their knowledge beyond the textbooks.”
Chavan frames the longer-term ambition in structural terms. “Instead of asking children questions about the curriculum, ask them what they know,” he says. “If we flip that, we can flip the education system from being a filtration mechanism to one that offers differentiated pathways based on a child's interests and background knowledge. Before AI, this was not possible.”
Anthropic and Pratham are now expanding the partnership to support Tech in TaRL (Teaching at the Right Level), with a randomized controlled trial planned for several thousand students. The organizations are also exploring digital public infrastructure initiatives, including knowledge graphs, and potential expansion to Kenya, Rwanda, and other regions in the Global South.
The collaboration aligns with Anthropic’s broader India strategy. The company recently opened a Bengaluru office and announced partnerships across enterprise, education, and agriculture. India is now the second-largest market for Claude.ai, with educational and instructional tasks accounting for 12 percent of usage.
ETIH Innovation Awards 2026
The ETIH Innovation Awards 2026 are now open and recognize education technology organizations delivering measurable impact across K–12, higher education, and lifelong learning. The awards are open to entries from the UK, the Americas, and internationally, with submissions assessed on evidence of outcomes and real-world application.
