From Models to Agents: Why the Hard Part of Scaling AI Isn't Just the Model
Scaling AI strategy has long been framed as a question about models: which one, how large, how capable. The roundtable made clear that this question has quietly stopped being the primary bottleneck. Value appears only once a model is wired into an agent that takes real action inside an enterprise workflow, and that operational shift is where the core execution challenges now live.
What replaced the model-selection dilemma is a set of core infrastructure and governance questions: what it costs to run a system at true operational volume, whether an enterprise can switch models without re-architecting its software stack, who holds genuine operational control over infrastructure and data, whether autonomous agents can be trusted to act without constant human oversight, and how organizations move adoption beyond isolated pilots. These concerns crossed every sector represented in the room.
Cost and unit economics surfaced as immediate priorities. Organizations running billions to tens of billions of tokens daily actively manage token consumption and GPU expenditure as recurring operational costs. This drive for unit economics is prompting a parallel demand for model portability and open-source self-hosting to avoid vendor lock-in. Trust and governance remain a close second concern: leaders across fintech, healthtech, and consumer platforms pointed to the need for stronger internal AI auditing, unified agent harnesses, and standardized evaluation frameworks, particularly for multimodal systems where model cards and benchmarks lag well behind text-based tooling. In safety-critical domains such as healthcare, lending, and space research and technology, human-in-the-loop validation remains non-negotiable. Data sovereignty, meanwhile, was frequently discussed in terms of direct operational control rather than contractual paperwork, extending down to who can physically access a server, route a network path, or switch a system off. And even where the technology and infrastructure are ready, adoption inside the organization was described by participants as one of the hardest parts of the journey, requiring dedicated prototyping teams and disciplined, short-cycle demonstrations to earn the trust of business teams and leadership alike.
Ultimately, the roundtable highlighted that enterprise leaders moving fastest with agentic AI are those addressing compute economics, portability, sovereignty, trust, and adoption as one integrated infrastructure and organizational challenge, not five separate problems.
Free with a work email
What the report finds
3 shifts buyers should know
Open any finding for the analyst's explanation.

In the full report
- Every vendor profile and score
- Strengths and cautions by vendor
- The full-resolution quadrant
- 01Value arises from agents effectively executing tasks within enterprise workflows, not just from the models themselves.
- 02Economic pressures are driving demand for model portability and open-source solutions to avoid vendor lock-in.
- 03Governance frameworks and internal AI auditing mechanisms are lacking, which complicates effective AI deployment across sectors.
Abstract
In an era where artificial intelligence (AI) is increasingly integrated into enterprise workflows, the shift from model selection to agentic execution has become paramount.
This report synthesizes insights from a roundtable discussion among leaders in India's startup ecosystem, covering crucial topics such as token and GPU economics, operational sovereignty, and the importance of governance frameworks. Key findings reveal that the real value from AI arises not merely from models but from agents that can execute actions effectively within business processes. Additionally, concerns about operational control and trust are recognized as essential for widespread adoption. The report underlines that successful deployment in sectors like fintech and healthtech necessitates a flexible, economically sound infrastructure to manage compute needs efficiently while ensuring operational integrity and data sovereignty.
Full report · PDF
Get the vendor profiles, scores and full quadrant
Free with a work email. The PDF arrives in your inbox.
For vendors
Not in this study?
Every PeMa Quadrant starts with an open RFI. See which studies are open and when the next edition of this market starts.
