Vivriti Capital is investing heavily in FinOps and AI governance to manage rising token consumption, optimise model usage and maintain regulatory compliance as enterprise AI adoption scales across lending, operations and software development.
As artificial intelligence adoption accelerates across financial services, organisations are increasingly balancing productivity gains with governance, compliance and cost management. Vivriti Capital is applying AI across enterprise lending, portfolio monitoring, software development and operational workflows, while simultaneously investing in FinOps, data residency controls and AI governance frameworks to ensure responsible deployment at scale. The company is also exploring how AI can enhance credit assessment, automate monitoring processes and improve software development productivity. Sushantam Mohan, Chief Data Officer, Vivriti Capital, speaks with FE Futech.
Balancing GNPA, NNPA and Scaling up Disbursements: If disbursements scale rapidly but NPAs rise, the data function is considered ineffective. If risk remains controlled but growth slows because underwriting is too conservative or manual, that is also viewed negatively. How are you employing AI to manage GNPA, NNPA and scale up disbursements without compromising on asset quality?
This can be viewed in two parts.
The first is retail lending, where AI has been well entrenched for years. Earlier, it was largely referred to as machine learning, but the underlying objective remains the same. AI is used through origination scorecards that help lenders decide whether to approve or decline a loan and determine the appropriate credit limit. The idea is not to avoid risk, but to manage it within a predefined risk appetite.
In retail lending, AI enables lenders to analyse thousands of attributes simultaneously, helping them make more informed decisions than what a purely policy-driven approach can achieve. Beyond origination, collection models also play a critical role. Asset quality is influenced not only by the borrowers you onboard but also by collection efficiency. Models such as propensity-to-pay help identify which customers require greater collection focus and which are likely to repay without intervention. Together, origination and collection scorecards help optimise both disbursements and asset quality.
However, Vivriti’s primary focus is enterprise lending, and this is where we see a significant opportunity for AI.
Enterprise lending involves large-ticket loans, with minimum ticket sizes of around Rs 5 crore. Credit decisions are not straight-through processes; they involve experienced professionals such as chartered accountants, credit analysts and finance specialists who evaluate financial statements, business plans and sector outlooks before forming a lending view.
A significant portion of this process remains manual across the industry today, and that is what we are working to address. We are evaluating how AI agents can replicate parts of the work performed by chartered accountants and risk analysts. This includes analysing financial statements, assessing industry trends, building projections and generating insights that support credit decisions.
The second area is portfolio monitoring. Once a loan has been disbursed, we collect large volumes of information on a periodic basis, including bank statements, inventory statements, profit-and-loss accounts and balance sheets. Monitoring this information manually is time-consuming. We are increasingly using AI to automate parts of this process and generate early warning signals that may not be immediately visible through traditional reviews.
Retail lending is already a relatively mature use case for AI. Our focus is on applying similar capabilities to enterprise lending, where AI can help automate analysis, strengthen monitoring frameworks and improve the efficiency and consistency of credit assessment while retaining human oversight in decision-making.
What kind of underwriting discipline has the company been able to cultivate after using AI? What kind of AUM growth has the company experienced, and can that growth be attributed to AI?
At this stage, most of our AI initiatives remain a work in progress. We do not start AI projects for the sake of experimentation; every initiative is tied to a specific business objective and measurable outcome.
One example is credit assessment in enterprise lending. A credit assessment that may traditionally take four weeks is something we are evaluating whether AI can help compress to approximately one week. If a risk manager is currently able to evaluate six cases a month, the objective is to increase that throughput significantly without compromising quality. That can have a direct impact on business growth by improving the speed at which opportunities move through the underwriting funnel.
The second area is standardisation of analysis. Credit and risk teams typically comprise professionals with varying levels of experience. AI can help create a more consistent analytical framework, enabling junior team members to perform at a level closer to experienced analysts. This improves the quality and consistency of underwriting while also creating operational efficiencies.
Beyond underwriting, we are also applying AI to business development and operational workflows. For example, our business teams evaluate thousands of companies to identify potential lending opportunities across products such as term loans, working capital financing and project finance. Many parts of this process are repetitive and well suited to automation.
We have made significant progress in enabling business users to build their own AI-driven workflows through what we call a DIY, or “Do It Yourself”, framework. Users can create agents that automate routine activities without requiring support from the technology team. Tasks such as monitoring information sources, gathering data and distributing reports can now be orchestrated far more efficiently. Even a saving of two to three hours per employee can translate into substantial productivity gains across an organisation of our size.
That said, we are still in the early stages of measuring the business impact of many of these initiatives. Most of them have been launched within the last six to twelve months, and the economics of AI continue to evolve as organisations learn to manage usage and associated costs.
The clearest return on investment today is in software development. We are already seeing productivity gains of 30% to 40% in parts of the software development lifecycle. Once requirements are clearly defined, AI can significantly accelerate coding and testing activities, reducing development timelines and improving delivery speed.
For underwriting, business operations and other enterprise use cases, we are still evaluating the long-term impact. As of now, we do not have a definitive number that directly attributes AUM growth or underwriting improvements to AI, but the early indicators on productivity, consistency and speed are encouraging.
You mentioned the growing challenge of token consumption. The cloud industry eventually addressed similar cost-management issues through FinOps. Do you see FinOps emerging as the solution for AI usage as well, and have you already started working on it?
Yes, we have already started working on it, and I believe FinOps will be critical to sustainable AI adoption.
AI vendors are continuously refining pricing structures as they gain a better understanding of consumption patterns and compute economics. Organisations are going through a similar learning curve.
Today, most enterprise AI platforms provide detailed usage consoles that offer visibility into how different models are being used, by whom, and for what purpose. As a technology leader, I can see whether a developer is using a model for planning, coding, debugging or other activities, along with the associated costs.
That visibility is important because not every task requires the most capable or most expensive model. For example, a premium model may be appropriate for planning and architecture decisions, while routine activities such as debugging can often be handled by smaller, lower-cost models. Optimising model selection based on the task is a key part of FinOps.
We are already allocating a meaningful portion of our technology operating expenditure towards AI usage, and it is important that these costs remain under control. With more than 80 developers using these tools, including interns in some cases, governance becomes just as important as cost management.
For us, FinOps and governance go hand in hand. FinOps focuses on tracking costs, understanding usage patterns and identifying optimisation opportunities. Governance focuses on ensuring that the right models are used for the right tasks and that AI resources are deployed responsibly.
That said, I would not claim that the industry has fully solved this problem. Like many organisations, we are still learning. In the future, we may move towards more structured workflows where specific stages of work are automatically routed to designated models, rather than providing unrestricted access to every model. We are evaluating such approaches, but we are not there yet.
Broadly, what kind of AI ecosystem have you built? AI adoption is not just about tools—it also involves data pipelines, governance, risk management, upskilling and stakeholder management. What does your AI ecosystem look like today, and what is the roadmap going forward?
We broadly look at our AI ecosystem through two lenses: non-technology users and technology users.
For non-technology users, the objective is straightforward. We want to improve accuracy, reduce manual errors and significantly increase productivity by automating repetitive tasks. Teams across functions such as operations, finance, business and marketing are being enabled with internally developed and externally sourced AI tools that can help automate parts of their day-to-day work.
Given that we operate in a highly regulated sector, governance is a critical component of this ecosystem. We have built a wrapper around our underlying AI models to ensure compliance with our governance requirements. One of the key considerations is data residency, as customer data cannot move outside India. The wrapper ensures that AI usage remains within a controlled environment and complies with regulatory expectations.
We have also established role-based access controls across AI platforms. Users only have access to capabilities relevant to their functions, and uploads to external AI platforms are restricted. Adoption is therefore taking place within a governed framework rather than through unrestricted usage.
Alongside governance, we are investing heavily in user enablement. We conduct regular training programmes to help employees understand how to use AI tools effectively and build their own no-code workflows and agents. The objective is to make AI accessible to business users without requiring coding expertise.
Our long-term goal for non-technology teams is to move employees up the value chain. Repetitive and administrative tasks should increasingly be handled by AI agents, allowing employees to focus more on analysis, decision-making and problem-solving. Ultimately, we want employees to be as comfortable building AI-driven workflows as they are today using tools such as Excel.
On the technology side, the focus is even more heavily skewed towards governance, efficiency and FinOps. All AI-assisted software development takes place within controlled repositories and environments, ensuring that code quality, security and compliance standards are maintained.
In fact, FinOps is currently one of our largest areas of AI investment. While AI adoption among business users remains important, a significant portion of our focus is on ensuring that AI usage within technology teams is governed, measurable and cost-efficient.
A major area of focus is ensuring that developers use the right model for the right task. Cost optimisation has become an important discipline, and FinOps plays a central role in managing AI expenditure while maintaining productivity gains.
We are also closely tracking software development metrics, including code commits, deployment velocity, bug rates and user acceptance testing outcomes. For us, efficiency is measured by how quickly new requirements can move from development to deployment without compromising quality.
The roadmap is centred on accelerating software delivery while maintaining strong governance controls. Historically, a feature may have taken close to two weeks to develop and deploy. Our objective is to compress that timeline significantly through AI-assisted development, testing and workflow automation.
Ultimately, the ecosystem we are building is not centred on AI tools alone. It combines governance, compliance, data residency, user training, cost management and productivity measurement into a single operating framework. The long-term goal is to make AI a core part of how both business and technology teams operate while ensuring that governance and risk controls remain firmly in place.
