AI is no longer confined to experimentation labs or innovation teams. CIOs and technology leaders are making strategic decisions about how to deploy AI in environments. This means teams must make decisions in real time, with operational constraints and real consequences. As organizations transition from AI experiments to AI adoption, there is a need for the same attention given to infrastructure as to the models.
In this episode of the Analytics Insight Podcast, Priya Diyalani speaks with Shrish Anand Lal, Executive Director and Chief Business Officer at NewStreetTech, about why enterprises need to focus on AI infrastructure and architecture rather than relying only on AI models. The discussion explores the limitations of model-first strategies, the core building blocks of enterprise AI infrastructure, the evolution of cloud and hybrid deployment, AI ROI, autonomous agents, AI platform strategy, and the importance of deterministic execution and governance. Here are the key excerpts:
I believe the AI model is only about 20% of the problem when we talk about enterprise adoption. The other 80%, however, includes all of the above. At NewStreetTech, we classify the problems in terms of FITS: fear, inertia, trust problems, and surprise. None of the above problems can be addressed simply by adopting a better AI model.
As per the research from NASSCOM, 65% of companies in India have adopted pilots for AI, but only 15% have made it to the production stage. I think the model has to be applied at the design phase, whereas the execution phase has to be deterministic. The AI model is becoming a commodity very quickly, but the AI architecture would be a competitive advantage.
There are four key building blocks that companies need to pay attention to. First of all, it is separation. AI should operate at the design and configuration stage, while execution should remain deterministic. These two layers must not interact.
Secondly, it is governance by design. Maker-checker workflows, audit trails, version control, and role-based access cannot be features that organizations add later. This is the foundation. Thirdly, it is orchestration. Implementation of one AI platform for all purposes is the wrong approach. At MyFix.ai, we use more than 36 agents of configuration and 14 execution engines, which are deterministic.
Finally, it is deployment flexibility. The company has to have a choice of on-premises deployment, cloud deployment, or hybrid deployment. It will help regulated organizations to deploy their AI platform into their environment.
AI computing costs are falling very fast, and this is definitely good for its adoption. But in my view, the actual revolution is going beyond computing into architecture. It is the transformation from AI as a service to AI as an infrastructure.
Organizations need to be resilient through multiple providers to avoid dependency on one AI provider. Certain workloads may need to be kept on-premise, while some others can go into cloud environments. This architecture transformation is even more critical for regulated industries. Where the whole system is left up to the AI, it creates a black box with many questions unanswered.
Two years ago, the question in many boardrooms was whether organizations should adopt or invest in AI. The question for today, on the other hand, is why AI has not been ready for production yet in their company. This alone goes a long way in highlighting the progress that has been made regarding the debate.
What has been debated as far as the ROI of AI is concerned has moved from the cost of implementing AI to the cost of not implementing AI. Enterprises are beginning to worry about the cost of failing to implement AI and lagging behind their competition. But those companies that are really making an ROI out of AI started small.
AI agents are not chatbots. They are specialized workers that handle specific enterprise concerns. One agent may configure a user interface, another may set up a validation rule, another may build a workflow, while another handles compliance.
At MyFix.ai, we have more than 37 specialized configuration agents and 14 deterministic execution engines. The agents build the system, humans approve it, and the engines run it. That is the fundamental difference between building enterprise-grade AI infrastructure and simply adding more computing power. The future will not be one AI system doing everything. It will be specialized agents working through orchestration, trust levels, cross-validation, and human oversight.