As organizations push to scale AI, the meaning of “robust” infrastructure has evolved. Beyond uptime and security, robustness now demands adaptability, modularity, and rapid iteration.
To explore this shift from multiple angles, the AIM Global Council roundtable brought together voices who operate at very different speeds. This roundtable brought together three enterprise leaders and three startup founders building AI-native products. The goal was to explore one big tension: how do you move fast without breaking systems that need to last?
Around the table were Anand Mahalingam Ph.D – Vice President – Data Science at Digit Insurance, Abdul Rahman Janoo – CEO of Tericsoft and SuperEngineerAI, Dr. Saurabh Pramanick – Data Governance Officer (DGO) – Data Governance Council (DGC) at Bank Muscat, Paddy (Padmanabh) Padiyar – Executive Director – (Head) Transformation at OCBC, Sarthak Sharma – Co-Founder at Flaunt and Srishti Arora – CEO & Co-founder at Agriforetell collectively representing both the enterprise and startup ends of the AI ecosystem. The session was moderated by Anshika Mathews, Global Media Lead at AIM Media House.
The significance of this conversation lies in the pilot-to-production divide that dominates 2025. Enterprises are hindered by fragmented systems and delayed feedback, while startups risk creating solutions misaligned with scale or compliance. Both sides face the same question: can agility and stability coexist, or do you have to choose?
Robustness = Composability + Evolvability
The roundtable discussion redefined robustness. It’s not about systems that never fail. It’s about systems that evolve, reconfigure, and recover fast when failure happens. As Sarthak Sharma from Flaunt put it, “Robustness today isn’t about being unbreakable. It’s about systems that evolve continuously and recover quickly when things inevitably break.” This reflects reality. AI systems face constant change. Models update, data pipelines shift, and workload demands fluctuate unpredictably.
Startups emphasized speed and adaptability as core criteria. With limited resources and small teams, they need infrastructure that supports rapid iteration and frequent model retraining. Composable architecture works because it’s built from small, well-defined components. You can reconfigure without destabilizing entire systems. The focus is on recovering quickly and improving continuously, not preventing every possible breakdown.
Enterprises prioritized predictability. They need reliable uptime, consistent latency, strong security controls, and alignment with regulatory standards. Their evaluation process involves multiple stakeholders. Infrastructure, security, architecture, and governance teams each assess robustness through different lenses. In highly regulated industries, expectations extend further. They need full audit trails and compliance documentation.
Despite different starting points, both groups found common ground. Enterprises often find startup solutions lack necessary governance frameworks. This prompts revisions that strengthen both security and operational resilience. Modern MLops practices bridge the gap. They incorporate performance monitoring, drift detection, automated recovery, and adversarial safeguards. This delivers both agility and stability. Composable, well-governed architectures emerged as the practical solution that satisfies competing demands for flexibility and control.
Data Maturity Is the Single Biggest Multiplier
Participants emphasized that 90% of AI success depends on data quality, metadata, and stewardship. Not model selection. Clean, well-organized data prevents hallucinations and unreliable outputs. Poor data sabotages even the most sophisticated models.
The most critical enabler is a single source of truth. Projection-based approaches that read from the source, rather than duplicating data, eliminate version drift and misinterpretation. Organizations with mature data warehouses, lakes, glossaries, and metadata catalogs onboard AI use cases significantly faster. Their consolidated datasets allow models to plug in and generate insights immediately.
Less mature organizations face a dilemma. Building full data maturity takes five to six years. But AI can’t wait. As Dr. Saurabh Pramanick from Bank Muscat put it, “You can’t wait five years to perfect your data maturity; AI will evolve faster than your architecture. You have to build both in parallel.” They bypass this by adopting data virtualization or semantic layers. These create unified views on top of distributed sources while governance improvements continue in parallel.
Both groups agree on one principle. Data chaos stems from missing ownership. Abdul Rahman Janoo from Super Engineer AI noted, “If no one owns the KPI, data quality drops fast. That’s when your AI output starts hallucinating.” Assigning KPI or outcome owners preserves trustworthiness.
Startups lack dedicated governance teams but compensate with lightweight connectors and targeted datasets that deliver clean inputs quickly. Enterprises have legacy infrastructure but struggle with fragmented ownership and regulatory complexity. Both prefer connectors over ad-hoc copying to maintain consistency. Use case prioritization depends on data readiness. Some projects must pause until foundational quality improves.
The Pilot → Production Gap
AI pilots frequently succeed in controlled settings but fail to scale. The surrounding system remains underdeveloped. User workflows, feedback loops, orchestration, and operational readiness aren’t there. User behavior and real-world input frequency differ sharply from pilot conditions. This exposes gaps only visible at production scale.
A recurring failure mode is feature-driven development. As Paddy Padiyar from OCBC observed, “A lot of AI is built in fragmented ways. Teams build features, not systems and features don’t scale.” These fragmented, isolated implementations don’t scale because they aren’t designed as reusable, composable units embedded in end-to-end processes.
Successful scaling requires deliberate domain prioritization. Enterprises must select two to three priority domains and reimagine high-value workflows completely. Don’t scatter shallow pilots across dozens of disconnected functions. Production demands whole-system engineering. You need MLops, monitoring, versioning, rollback paths, observability, and coordination across product, engineering, risk, compliance, and operations. Not just data science expertise.
Enterprises face organizational barriers. Compliance reviews, regulatory checks, process rigidity, and insufficient senior sponsorship to push pilots forward. Startups can deliver rapid proofs-of-concept but often lack production depth. They’re missing hardened deployment processes, SLA management, and risk alignment. Srishti Arora from Agriforetell highlighted this balance: “Speed is important, but going too deep into infrastructure slows us down because we don’t have big teams to manage it.”
The tension is clear. Enterprises want predictable integration. Startups want quick iteration. Both groups agreed the solution is joint design. Co-build the system around the model from the outset. Have a shared understanding of regulatory, operational, and workflow constraints. Don’t treat the model as standalone technology.
Cultural Tradeoffs
The cultural contrast between startups and enterprises was articulated through a compelling analogy by Anand Mahalingam. He explained “Startups are like speedboats, fast and highly agile but prone to flipping with even a small mistake. Enterprises are like oil tankers, slow to turn, but extremely stable in rough conditions.”
Startups thrive on speed, experimentation, and rapid feedback loops. Enterprises operate through governance, predictable processes, documentation standards, and deliberate risk mitigation. Each benefits by adopting the other’s strengths.
Startups gain sustainability by embedding enterprise disciplines. Reproducibility, compliance rigor, auditability, and productization standards become essential as they scale. Enterprises unlock agility by embracing startup behaviors. Test ideas quickly, iterate without exhaustive planning, and tolerate controlled failure.
Co-building emerged as the most productive model. When startups and enterprises design solutions together, supported by shared APIs, transparent governance, and trust-based data sharing, startups understand regulatory constraints early. Enterprises gain access to innovation without retrofitting. Increasingly, enterprises establish small innovation pods that operate like internal startups. They experiment rapidly while adjacent teams harden and scale proven concepts.
Both sides seek predictable business value and reduced operational risk. Participants agreed that culture, not tooling, determines execution velocity. The solution isn’t choosing speed or stability. It’s deliberately swapping strengths to achieve both.
Conclusion
The roundtable revealed a fundamental contrast. Enterprises prioritize compliance, predictability, end-to-end workflows, and long-term stability within regulated constraints. Startups optimize for speed, experimentation, minimal friction, and rapid feature delivery driven by customer discovery.
But beneath these surface differences lies substantial common ground. Both groups value data quality, clear ownership, composable architectures, and measurable business value. Participants unanimously agreed that models succeed only when supported by reliable data and production-grade systems.
Tension persists around trust and resource allocation. Enterprises demand governance assurances, data residency controls, hardened SLAs, and risk frameworks before deployment. Startups resist heavy upfront engineering and lengthy integration timelines that slow validation cycles. The definition of “robust” itself diverges. Agility for startups, auditability for enterprises. This creates friction in collaboration.
Solutions emerged. Data virtualization and semantic layers bypass multi-year maturity cycles. Shared APIs and transparent governance enable co-building. Domain prioritization focuses resources on high-value workflows. Assigned data owners prevent quality decay. The moderator emphasized that success depends on concise, high-impact priorities executed across startup and enterprise boundaries, not isolated efforts.
The consensus was clear. Robustness requires composable architectures, disciplined data stewardship, and deliberate business prioritization. Looking forward, participants expect hybrid organizational models to become standard practice. Startup-like innovation pods paired with enterprise hardening squads. AI infrastructure will advance not through choosing speed or stability, but by engineering systems capable of delivering both simultaneously.
