The AI team at Elevation cumulatively spent over 90,000 minutes this year, roughly 1500 meetings, talking to founders, enterprise buyers, and industry veterans. That’s almost certainly more time than we spent with our own partners, and it tells you everything about how busy (and important) 2025 really was.

This piece is a reflection of what those conversations taught us across founders, enterprises, and the underlying technology stack. Voice moved from promise to production, creative tools are getting rebuilt from first principles, infrastructure quietly became the hardest and most valuable layer, and enterprise adoption forced a rethink of how software is built and sold. This piece reflects on what changed this year, where AI actually delivered outcomes, and the patterns we believe will define the next generation of AI building from India for the globe.

India Is Finding Its Voice

Voice AI has always been well-suited to India because of its mobile-first population, deep language diversity, and the sheer scale of customer interactions that still happen over phone calls. Based on conversations with over 30+ enterprises across new-age commerce, banking, manufacturing, and logistics, it is clear that Voice AI has moved from promise to reality this year. The key shift has been in the underlying technology: faster and more accurate speech models, combined with low-latency LLMs, have enabled systems to move beyond brittle IVRs to real, natural conversations. The opportunity spans the stack: at the model layer (e.g., Sarvam, Murf AI, Maya Research, Navana Tech, etc.), there’s growing demand for Indian-language, accent-aware systems that can compete on both quality and cost; at the application layer (e.g., Greylabs, Ringg AI, Synthio Labs, Hunar, Squadstack, Arrowhead, etc.), the strongest pull comes from workflow-embedded use cases where success is measured in business outcomes, not call minutes. As a result, enterprises are now taking Voice AI out of pilots and into production, particularly in high-volume sectors such as BFSI, telecom, and retail. In practice, several clear patterns are emerging across large Indian enterprises -

• Right now, outbound voice AI is delivering far stronger ROI than inbound use cases. For example, a major fashion commerce player is seeing more value from AI-driven outbound remarketing than from handling its 75,000 daily inbound calls, while a medicine delivery company reports five times higher conversion rates when voice bots proactively call customers for repeat orders compared to traditional push notifications. Inbound calls, by contrast, often involve complex problem-solving that AI still cannot handle reliably.
• At the same time, generic demos matter far less than real-world results. Most enterprises test two to three vendors in parallel, and the winner is chosen based on conversion and performance within a specific workflow, not just technical sophistication.
• Integration complexity is another critical factor. Point solutions tend to break at scale, while companies that invest in deep API integrations into enterprise systems perform significantly better. This has pushed the category toward service-oriented models, since once a solution is deeply embedded into operations, switching costs become high. Successful players are therefore expanding horizontally within the same enterprise across multiple use cases rather than trying to expand rapidly across industries.

Overall, we are in the bundling phase of Voice AI. Large enterprises increasingly want the entire Voice AI stack bundled together and prefer to manage orchestration in-house, directly stitching together models. This approach gives them tighter control over quality, pricing, and compliance. In high-volume, high-stakes environments, Voice AI is becoming core internal infrastructure. Durable value in this space will ultimately be created either by owning differentiated, India-specific voice models or by building deeply verticalized applications that enterprises trust with their most critical, highest-volume workflows.

The Creative Suite, Reimagined

Multimodality in creative AI crossed a real threshold in 2025. Text, images, audio, and video can now be generated within a single creative loop, dramatically expanding what individuals and small teams can produce. That said, while generation across modalities is now possible, stitching them together seamlessly, while maintaining control, consistency, and editability across an end-to-end workflow, remains an open challenge. However, we see this as a temporary constraint: the underlying models and tooling have matured enough that 2026 is likely to be the year when multimodal creation becomes far more composable, controllable, and production-grade. These are some trends we’re seeing in the creative space -

• Creative software was historically built for experts using complex tools like Photoshop, Premiere Pro, or Blender, limiting creation to trained professionals. AI-native products such as Runway, Midjourney, Pika, CapCut, etc. now let non-creators generate high-quality design, video, and audio through simple prompts and edits. This has massively expanded the creative TAM, enabling solo founders, indie creators, and small businesses to do work that once required agencies or teams.
• Traditional interfaces optimized for control - timelines, layers, keyframes - but were slow and inaccessible to most users. As AI handles execution, tools like ChatGPT, Runway, and Figma AI let users create and edit through natural language, while node-based systems like ComfyUI offer modular control for advanced workflows. The result is faster iteration, easier collaboration, and more composable multimodal creation.
• Open source has reshaped the creative AI landscape. Open source video models such as Kling, Qwen, and Seedance are moving faster than their closed counterparts, often surpassing models like Veo and Sora on quality. This has lowered barriers for builders and shifted power toward product teams rather than platforms.

These shifts are already reshaping how creative products are built. One of our portfolio companies (in stealth) is reimagining designer workflows from the ground up, revamping its existing vector data layer to a code-based data layer. For founders, this opens up a fundamentally different opportunity. Creative AI products can be global from day one, self-serve by default, and distributed through creators rather than enterprise buyers. For Indian founders in particular, this is a rare chance to build category-defining products that sell themselves by tapping into open ecosystems, community-led adoption, and the accelerating demand for faster and more expressive creative output.

Dev Infra: The Quiet Layer Powering Everything

As AI systems moved from demos to production this year, the bottleneck shifted from models to infrastructure. Building reliable, enterprise-grade AI required far more than calling an API - it demanded tools for orchestration, observability, evaluation, security, and integration across an increasingly fragmented stack. Some trends we have been excited about in the Dev Infra space are:

• As AI systems move from single-turn generation to executing multi-hour or multi-day tasks such as end-to-end video production in Runway, research-to-code pipelines in Cursor, or sales ops agents in Clay and HubSpot, short context windows become a hard bottleneck. The real constraint is no longer raw intelligence but the ability to retain decisions, constraints, and state across steps without drift. This is driving demand for hybrid architectures that combine long context windows with explicit memory layers (e.g., retrieval, scratchpads, and task graphs) rather than relying on prompts alone.
• As AI systems are embedded into high-stakes workflows such as code generation in GitHub Copilot, fraud detection, or customer support automation, failures become expensive and hard to diagnose without visibility. Teams are adopting new tools now to track model behavior, prompt drift, tool usage, and outcome quality in real time. Evals are no longer research artifacts; they are production guardrails that determine whether AI systems can be trusted, iterated on, and scaled.
• Agent-to-agent (A2A) communication has shifted from agents chatting to systems coordinating. Early A2A relied on natural-language back-and-forth, which broke down in longer workflows due to unclear ownership and conflicting assumptions. The real bottleneck now is authority and state management - strong systems separate roles, communicate via structured artifacts, and coordinate asynchronously. We believe that the next wave of A2A will win by constraining interactions, treating coordination as an infra problem rather than a reasoning one.

We became increasingly bullish on this layer because it’s where real complexity lives, and we started seeing the best builders building in this space. The teams that made progress weren’t necessarily those with the best models, but those that could manage prompts, tools, agents, data flows, and failures at scale. We continued our partnership with UnifyApps (Unified Agentic AI Platform), Composio (Tool Calling), and Maxim AI (AI Evals and Observability) in the infra space, as well as partnered with another company in the governance and observability space (unannounced).

We believe the winners in this space will look less like traditional SaaS and more like core infrastructure - deeply embedded, hard to replace, and invisible when they work well.

What 2025 Taught Us About AI Company Building

An MIT study famously pointed out that nearly 95% of AI pilots failed to make it into production, a statistic that caused some flurry, but we had our own realisation of what it means by seeing many companies on the ground. The reason has rarely been model performance. It has everything to do with messy data, brittle workflows, compliance constraints, and unclear ownership. Unlike SaaS, where value could be delivered with minimal context, AI demands a deep understanding of how work actually happens inside organizations. As a result, generic, off-the-shelf solutions struggle to survive contact with enterprise reality.

In response, successful deployments followed a different playbook. Forward-deployed engineers, highly customizable systems, and deep integrations with existing platforms - CRMs, ERPs, core banking systems, and call centers - became table stakes. This reignited the build-versus-buy debate in a way SaaS never did. While SaaS scaled through uniformity, enterprise AI is scaling through adaptability. The companies making real progress treated AI less like a product they purchased and more like infrastructure they assembled, combining vendor speed with in-house control over critical workflows. This shift is also enabling a new generation of AI-powered services businesses, where custom software is easier than ever to build, and value is delivered through faster turnaround and deeper customization, particularly across ERP, CRM, and HR ecosystems.

Even the way AI/SaaS companies market and address discoverability has fundamentally changed. One of the major changes we see is that marketing will no longer be optimized for only humans, but rather a hybrid between humans (who continue to discover, interpret, and emotionally evaluate brands) and agents (who increasingly intermediate comparison, decision-making, and execution). This dual-audience reality also fundamentally reshapes how customer acquisition works and has brought GEO (Generative Engine Optimization) to the forefront. GEO, at its core, is about ensuring a brand is correctly represented in AI-driven discovery and decision systems to position it in the most relevant way for high-intent users and agents.

Stepping back, these enterprise dynamics mirror a broader change we observed across the ecosystem. We see Indian founders are becoming bolder, no longer content to build incremental tools or replicate existing categories. Instead, they are defining new markets - often at the intersection of AI, services, and infrastructure - where traditional boundaries no longer apply. The India–US corridor has become a genuine advantage in this transition, enabling founders to combine deep technical talent and cost efficiency with global market access and ambition, building globally relevant companies from day one.

Looking ahead, our biggest takeaway from 2025 is that AI is no longer compressing just costs; it is compressing cycles, hierarchies, and time-to-impact. The companies that endure will be those that embrace this reality early: building closer to outcomes, closer to workflows, and closer to where real decisions are made.

New Partnerships & Growing Relationships

This year, we were fortunate to partner with several pioneering companies that are reshaping their respective sectors.

We partnered with Adopt AI as Deepak Anchala, Rahul Bhattacharya, and Anirudh Badam build the platform that transforms everyday applications into intelligent agents responding to natural language commands. Their technology collapses what typically takes product teams 6-9 months of development into less than 24 hours, making it possible for any application to gain powerful AI assistant capabilities without rebuilding from scratch. Read our investment memo.

We led Synthio Labs' $5 million Seed round as Supreet Deshpande, Sahitya Sridhar, and Rajashekar Vasantha build clinical-grade AI infrastructure transforming how life sciences companies engage with clinicians and patients. Their conversational AI platform is purpose-built for the regulatory complexities and compliance demands unique to healthcare and pharma. Synthio's AI Operating System unifies three powerful platforms: Jarvis, the voice AI copilot for field teams; Ather, the multimodal engine powering seamless physician and patient engagement across channels; and Simulation Studio, generating high-fidelity digital twins for research and strategy.

Our partnership with UnifyApps deepened significantly this year as we supported their $50M Series B round. Since leading their $11M Seed and doubling down in their $20M Series A, we've watched Pavitar Singh and team execute with remarkable precision. UnifyApps has evolved from a unified integration platform to a comprehensive AI-native operations layer for enterprises, helping companies move from digital transformation to truly intelligent, autonomous operations. The addition of Ragy Thomas as Co-CEO and Chairman brings invaluable experience to further accelerate their mission of helping enterprises master the AI era.

We deepened our partnership with Composio through their $25M Series A, building on the conviction that led us to lead their Seed round. Soham Ganatra and Karan Vaidya's vision to build the learning infrastructure that transforms AI agents from static tools into systems that actually improve over time has proven prescient. Since our initial investment, Composio has emerged as the go-to platform for AI agent infrastructure, attracting over 100,000 developers and earning 25,000+ stars on GitHub. The platform now supports 10,000+ integrations and serves 200+ startups and enterprises.

We participated in Zoca’s $6M round as they build the fundamental infrastructure for local businesses to thrive in an AI-first world. Ashish Verma and Robin Chauhan are transforming how hyperlocal service businesses get discovered, booked, and rebooked with their AI-first growth platform. Their hyperlocal intelligence engine identifies neighborhood-level search trends while a comprehensive suite of AI agents manage the entire growth funnel automatically. In less than a year, Zoca has helped over 1,000 local beauty and wellness businesses generate $10M+ in revenue and book more than 120,000 appointments, while saving them hundreds of hours in marketing tasks.

Podcasts From Our Stable

Our podcast initiatives continued to capture the voices and insights of remarkable founders navigating the AI transformation.

Day One

In our conversation with Deepak Anchala of Adopt AI, he articulated a vision where "humans will not have to adapt to applications. Rather applications will just understand what humans want and get things done." The episode explored why every company's DIY agent efforts will hit a fatigue wall, the transition from prompt-and-response to fully autonomous agents, and why every founder needs to become an influencer.

Soham Ganatra of Composio joined us to discuss building "AI's muscle memory, the action layer beneath intelligence." The conversation unpacked how Composio occupies a unique position in the AI stack between LLM frameworks and real-world applications, creating the bridge that allows LLMs to reliably interact with thousands of tools.

SummitUp

In our "State of AI in 2025" episode, Krishna Mehra and Poorvi Vijay unpacked the latest developments after a year where the landscape was completely rewritten—from sophisticated reasoning models to multi-billion dollar M&As and multimodal interaction becoming standard. The tactical deep-dive covered where Elevation is investing in AI, India's enterprise adoption challenge, the agentic stack revolution, scaling dynamics of small teams, the new talent wars, and building distribution and moats.

Our "India's AI Opportunity" episode explored three key themes where we see immense potential: Infrastructure & Middleware (building essential tools connecting models to enterprise data), Vertical AI (reimagining industry workflows by combining domain expertise with AI), and the Services + Product Play (a unique India advantage in delivering end-to-end AI transformation globally). Unlike cloud adoption which took a decade, enterprises are rapidly deploying AI. And this time, India isn't playing catch-up.

Strengthening The India-US Corridor

As more Indian founders build with a global-first mindset, the Bay Area has become a critical hub for our community. This year, we invested heavily in nurturing this cross-border community through a series of gatherings that brought together founders, technologists, and ecosystem leaders in the Bay Area.

Our Bay Area gathering in June brought together founders building across the India-US corridor for an evening that captured the vibrant energy of this cross-border community. Saurabh Tiwary, VP & GM of Cloud AI at Google, joined Krishna Mehra for a candid fireside sharing insights from scaling AI at Google. The discussions reinforced our conviction: the India-US corridor has evolved from a talent pipeline into a platform for category-defining AI companies.

In partnership with Founders FUNDA, we explored strategies for building enduring companies in fast-paced markets. Panels on "Architecting Scale: Building Startups That Last" and "Reimagining Legacy Systems with AI" featured founders from companies like WisdomAI, Hightouch, and DOSS. Key themes emerged: stay close to customers, keep product marketing founder-led, protect founder-market fit, build moats around passion and empathy, and ship fast always.

Our Bay Area AI CTO/CPO Mixer, co-hosted with Foster Ventures, centered on how companies are building and deploying AI agents across applications, from choosing between copilot versus autonomous approaches to building user trust and transparency in AI systems. The intimate gathering facilitated candid exchanges on different tools and technologies leaders are evaluating.

We celebrated Diwali with 65+ founders in San Francisco, co-hosted with AngelList. The evening brought home a little closer for brilliant founders who are building and thinking AI-first, but away from families and the light and joy that is Diwali. When founders stay beyond the official wrap time, still deep in conversation, you feel what makes this community special.