Embracing the AI Revolution – A Shift from Skepticism to Conviction

When Khilan Haria, Chief Product Officer at Razorpay, addressed over 60 participants from 30 different companies during a recent ‘AI Build Series’ webinar hosted by Elevation Capital, he confessed that just a year prior, he viewed AI with uncertainty, questioning if it was "a fad or a real big innovation."

This initial skepticism quickly evolved: by August-September, it became a sense of fear, culminating in complete conviction by October-November that AI was "very, very big" – and that he was "already behind." This journey from initial doubt to urgent adoption mirrors the path many organizations are currently navigating as they grapple with the implications of AI for product development and organizational structure.

The initial state at Razorpay reflected a familiar pattern: pockets of experimentation scattered across the organization. Engineers were tinkering with GitHub Copilot, marketing teams were early adopters of Midjourney, and various individuals were exploring ChatGPT and Claude. However, this disconnected exploration lacked strategic direction until leadership recognized AI as a fundamental shift that would reshape not only their tools but also their entire approach to building products.

The Fundamental Shift: Reimagining Product Development in the AI Era

The Traditional Product Building Cycle and Its Inefficiencies

Razorpay's traditional product development cycle followed a familiar pattern that many organizations will recognize: roadmapping and conceptualization flowing through the traditional troika of design, engineering, and product teams, with each handoff introducing what Khilan termed "collaboration taxes." The inefficiency was stark – even simple changes like updating the primary colour of a city on the website could stretch from one hour of actual work to three days of process.

This cycle applied universally across all complexity levels (Very Low, Low, Medium, High). Every change, regardless of scale, had to navigate through multiple teams and approval layers, creating bottlenecks that frustrated teams and slowed innovation.

The Vision of the "Full-Stack Builder"

The first transformative hypothesis Razorpay tested was whether product managers could become "full-stack builders" – taking ownership of low-complexity initiatives from ideation through to production deployment. Using tools like Replit and Lovable, the team began experimenting with a radical new approach where PMs could handle prototyping, market validation, and even live deployment for certain categories of work.

Early Validation Results

• Four initiatives successfully went live in the "very low to low complexity" category
• Most implementations used Replit as the primary tool
• Both internal tooling and customer-facing products benefited

Notable Example - Partner Marketplace

The early results were promising. One notable example was a partner marketplace – a full customer-facing product that allowed merchants to find and connect with Razorpay's partners. Rather than spending months on prototypes and research, the team validated this offering in production, with real customers.

This wasn't limited to customer-facing products. Internal tooling and admin console configurations also benefited from this approach, demonstrating that the full-stack builder model could apply across various use cases.

Mandating AI Fluency Across the Organization

Razorpay's approach to building AI fluency drew inspiration from the mobile revolution. Khilan drew a parallel to the mobile revolution, recalling how companies in that era provided employees with free mobile phones to accelerate mobile-first innovation. The logic was simple: if employees became heavy users of the technology themselves, they would innovate faster in applying it to customer-facing solutions.

Following this principle, Razorpay mandated that all team members use AI in their day-to-day work wherever possible to build familiarity and comfort with AI capabilities. The hypothesis was that only through personal experience would teams develop the intuition needed to reimagine customer-facing products in an AI-native way.

Organizational Reimagination: Flattening Hierarchies and Building AI Competencies

The Bold Decision to Eliminate Mid-Management

After months of experimentation and validation, Razorpay made two critical decisions that would fundamentally reshape their product org. The first was eliminating the middle management layer entirely from the product team.

The Group Product Manager (GPM) role was removed, flattening the hierarchy from three layers to two. Senior Directors became first-level leaders, while VPs and SVPs became second-level leaders. The people leader to individual contributor ratio shifted dramatically from 4:1 to 7:1 on average, with some leaders directly managing teams of 10-11 product managers.

This decision wasn't without controversy. Khilan candidly acknowledged that "a lot of our mid managers didn't like it, and it created a bit of chaos in the short run." The company was transparent about the changes, conducting multiple sessions with affected managers and offering transitions to senior individual contributor roles. Some chose to leave for organizations that maintained traditional hierarchies, recognizing that not everyone thrives in or desires an IC-focused role.

Managing the Cultural Transformation

The cultural change management process revealed several critical insights. The hardest part wasn't convincing the mid-managers whose roles were eliminated – it was convincing senior leaders who would now manage much larger teams directly. These leaders waited two to three months to see if "the goodness of AI" would materialize to support their expanded responsibilities.

Khilan's hypothesis to these leaders was that AI could handle "v1" of many managerial tasks – providing 80% of PRD reviews, offering initial mentorship and guidance, and automating routine check-ins. When 50% of leaders bought into this vision, they became evangelists who helped convince the rest.

The transition was deliberately phased rather than sudden. While new GPM positions were immediately frozen, existing GPMs were given time to transition, with some remaining in place for three to six months as the organization adjusted to its new structure. The key message emphasized repeatedly was that this wasn't a cost-cutting measure but a strategic reimagination of how product teams should operate in the AI era.

Building AI as a Core Competency

The second critical decision was adding AI proficiency as a core competency for all product managers. This manifested in two dimensions:

• Day-to-day PM work with AI: Teams were expected to use AI tools for strategy authoring, discovery, customer research, PRD writing, and go-to-market planning. The company provided enterprise licenses for Claude, ChatGPT, Midjourney, and other tools, actively monitoring usage to guide adoption.
• PMs as full-stack builders: Product managers were expected to develop the ability to take low-complexity initiatives directly to production, bypassing traditional development cycles entirely for appropriate use cases.

The Infrastructure of Learning

Razorpay's approach to upskilling was multifaceted and intentionally designed to accommodate different learning styles and comfort levels with technology.

• Early Experimentation Phase: Teams were allowed to explore various AI models and tools without mandating specific choices. Engineers tried GPT's research model, some explored DeepSeek, others used Claude Sonnet, while teams experimented with Lovable, Bolt, and Vercel. The key enabler was establishing tech compliance frameworks that allowed safe experimentation with paid enterprise plans ensuring data residency and preventing company data from being used for training.
• "AI Show and Tell" Sessions: Every Friday from 4-5 PM, team members who had achieved notable results with AI would present to the organization. These sessions exceeded expectations – the first attracted 200 attendees, and consistently maintained over 100 participants for months. Khilan personally sponsored these initially before transitioning them to the L&D team.
• Leadership Hackathons: Fifty senior leaders participated in mandatory two-day hackathons where they built AI solutions hands-on. The message was clear: "leave your work... this is not optional." These sessions provided basic coaching on tools like Replit, Claude, and Lovable. Remarkably, a non-technical team won one hackathon by building a live voice agent using Eleven Labs and other tools, proving that "AI is not really that hard."
• Internal Slack Channels: Two channels became central to the transformation:
  • "AI Bulletin" for sharing live AI projects with implementation details.
  • "AI Help" for support requests, personally monitored by Khilan for the first three months.
• Streamlined Procurement: The traditional 30-60 day procurement process for new tools was compressed to 7-15 days through direct CEO involvement and a dedicated program manager for AI tool procurement.

Practical AI Implementations and Use Cases: From Automation to Customer Experience

Internal Tools:

1. Automated Product Requirement Document (PRD) Reviewer: An AI agent that reviews product specs, giving red/amber/green ratings and detailed section-by-section feedback, augmenting human reviewers.
   1. Impact: Replaced 80% of first-line manager review work, allowing PMs to iterate independently before human review.
2. Automated Weekly Check-ins: An internal tool that aggregates data from various sources (DevRev, Tableau, Slack, Google Drive, OKR trackers) to generate exec summaries for team check-ins.
   1. Impact: Reduced multi-hour weekly task to 30 minutes.
3. Strategy Reviewer: An AI tool on Claude that reviews strategy documents, providing overall ratings, summaries of strengths and areas for improvement, and specific recommendations.
   1. Impact: Automated the basic feedback Khilan previously provided personally, reducing intimidation for PM1s and PM2s seeking guidance.
4. Resume Parser: An AI tool to filter out non-relevant resumes, score candidates, and identify those to move to interview queues.
   1. Impact: Manages 300-400 applications per posting efficiently
5. AI Interview Agent: A fully voice-interactive agent that conducts initial rounds for product thinking and problem-solving, providing feedback and scores to move candidates to human rounds.
   1. Impact: Has interviewed 400 candidates, reducing PM interview load from 3/week to 1/week as requested.

Customer-Facing Products:

1. Customer Service Automation: An agentic AI system resolving thousands of customer support tickets daily, offering deterministic, consistent, fast, and cost-efficient support.
   1. Evolution: First version had "poor quality signals compared to humans," leading to V2/V3 iterations that balanced probabilistic and deterministic approaches
2. Merchant Year Snapshot: Automated year-end summary for merchants. Generates insights on highest order, average order value. Compares merchant performance to peers. Delivers 10-20 personalized insights via email.
   1. Impact: Created throwaway seasonal product quickly without traditional development resources.
3. MCP Server: API wrapper enabling programmatic access to Razorpay functions such as link creation, report generation, and other dashboard functionalities.
4. Customer Dashboard Copilot: An AI-powered conversational agent within the dashboard allowing customers to navigate UI, download reports, and interact conversationally, aiming for a "fully invisible UI" or WhatsApp as a primary UI for smaller customers.

Overcoming the Last Mile: From POC to Production

The Production Challenge

Moving from proof-of-concept to production emerged as one of the most significant challenges – a theme that resonated strongly with webinar participants. Razorpay's approach was pragmatic and iterative.

The team started with independent applications that didn't require deep integration with existing services or data. The partner marketplace and merchant year snapshot projects exemplified this approach – they could fail without affecting core systems, and their "throw away" nature in terms of code quality was acceptable for validation purposes.

This freed teams to launch quickly without full design system integration, though this caused friction with the front-end design team who pushed for consistency.

Streamlining the Path to Production

Several strategies emerged for accelerating production deployment:

• Security and Compliance Fast-Track: The security approval process was automated for certain categories of applications, with a dedicated team building "one-click" deployment capabilities within secure VPCs. A program manager with direct access to Khilan and the CEO helped unblock procurement and compliance issues.
• The 80/20 Engineering Partnership: Rather than requiring full engineering team involvement, PMs handled 80-90% of the development work through tools like Replit, with engineers contributing the final 10-20% for code review and deployment. Many PMs leveraged "engineer friends" who contributed a few hours of spare time to push projects over the finish line.
• Design System Trade-offs: While initially accepting UI that merely mimicked Razorpay's design without using actual design tokens, the team is now exploring tools like Cursor that can integrate with their Blade design system through MCP, potentially solving the consistency challenge.

The "One-Way vs. Two-Way Door" Decision Framework

Razorpay's decision-making framework proved crucial for maintaining velocity while managing risk. API integrations represent "one-way doors" with strict review processes and multiple layers of approval. Most launches, however, are "two-way doors" where the pod (PM, EM, and designer) is trusted to launch independently, with failure being acceptable.

This framework forces teams to explicitly categorize their decisions and justify their reasoning, creating thoughtful risk assessment without bureaucratic overhead.

The Road Ahead: Building for an AI-Native Future

Key Lessons and Ongoing Challenges

Several critical insights emerged from Razorpay's journey:

• The Probabilistic-Deterministic Balance: Pure AI solutions often hallucinate or provide inconsistent results. Successful implementations combine probabilistic AI capabilities with deterministic rules and checks, especially for customer-facing applications.
• The Importance of Executive Sponsorship: From Khilan personally monitoring the AI Help channel to the CEO's involvement in procurement acceleration, leadership engagement proved essential for cultural transformation.
• Tool Evolution vs. Current Limitations: Many desired capabilities don't yet exist in the market. Razorpay found that while tools are "evolving very rapidly," they often had to build custom solutions or wait for vendors to catch up with their compliance and data residency requirements.
• The Voice Frontier: Voice interaction remains challenging. As Khilan acknowledged, they still lack "great system-driven evaluation for voice-to-text accuracy," relying on manual audits. Response latency and call transfer logic in voice systems remain active areas of development.

Strategic Frameworks for AI Adoption

Razorpay's approach to identifying AI opportunities follows a clear progression:

• Start with "Boring Work": Target tasks that teams don't want to do – interviews, routine reviews, check-ins
• Enable Experimentation with Room for Failure: Accept that not everything will work and plan accordingly
• Distinguish AI-Native from AI-Enabled: Recognize that AI-enabled solutions solving previously impossible problems are valuable even if not fully "native"
• Iterate Strategy Quarterly: AI strategy cannot be annual – Razorpay reviews and adjusts every three months

The Organizational Transformation Continues

Razorpay's transformation extends beyond the product team. While the flattening of hierarchy and AI competency requirements currently apply primarily to product management, each function is determining its own path. Some teams, like customer service, are already deeply transformed, while others are still finding their way.

The company has established dedicated teams for AI exploration – one focused on product hypotheses and another on horizontal AI infrastructure. These "SWAT teams" own the discovery of new capabilities, from exploring how PMs might write full automation tests to investigating the potential for AI-native payment gateways.

Conclusion: Embracing Continuous Evolution

Razorpay's journey from AI skepticism to organizational transformation offers a blueprint for product teams navigating similar transitions. The key isn't just adopting AI tools or automating workflows – it's fundamentally reimagining how product organizations operate in an era where AI can augment or replace traditional management layers.

The path requires courage to make bold organizational changes, patience to allow time for adaptation, and humility to acknowledge that many experiments will fail. It demands investment in learning infrastructure, from show-and-tells to leadership hackathons, and the creation of psychological safety for experimentation.

Most importantly, it requires recognition that this transformation is ongoing. As tools evolve and capabilities expand, organizations must maintain the flexibility to continuously reimagine their structures and processes. The companies that thrive will be those that, like Razorpay, combine strategic vision with pragmatic experimentation, moving decisively when hypotheses are validated while remaining open to course correction when they're not.

The AI-native era isn't coming – it's here. The question isn't whether to transform, but how quickly organizations can evolve while maintaining the stability needed to serve customers and support teams through the change. Razorpay's experience suggests that with clear communication, strong leadership support, and a commitment to continuous learning, even dramatic organizational transformation is not only possible but also necessary for remaining competitive in the AI era.