Agent marketplace one of our large bets, says PayU CPO Manas Mishra

PayU is building AI-powered tools and agent marketplaces as it bets on agentic commerce and payments becoming the next frontier in digital transactions

As artificial intelligence (AI) reshapes digital commerce, India’s payments industry is seeking to capitalise on opportunities emerging at the intersection of agentic commerce and payments, while preserving its longstanding emphasis on trust and security. Manas Mishra, chief product officer (CPO) at PayU and Wibmo, discusses how this transition could unfold, where companies stand to create value and how AI is equipping developers with greater context and productivity, in a video interview with Ajinkya Kawale in Mumbai. Edited excerpts:

 

In an evolving AI world, what are your key focus areas?

 

Our AI strategy is broken into four buckets. First: The role of PayU in the larger trend of commerce shifting to agentic rails, and how we help consumers and merchants. Second: How we democratise access to AI and GenAI (generative AI) technologies for our customers who are banks and merchants. 

 

Third: As we are seeing new trends emerging in AI in terms of fraud patterns or consumer behaviour, the products we have built over the years have to be AI-native ones. Fourth is improving efficiency and productivity of all functions in the organisation — product, engineering, operations, legal, and financial — to solve bottlenecks. 

 

How do you see agentic flows play out for merchants? 

 

When you think of the commerce journey there are two parts: Agentic commerce and agentic payments.  Our systems have been designed with a view of a user who gives explicit consent; all products and processes are designed keeping that in mind. When it’s an AI agent, questions of trust, traceability, accountability, audits, all of them come into picture. This is where we are seeing pilots today, which are one-on-one and tightly coupled. We need an interoperable protocol in the market that can define how this would work. 

 

Once commerce is solved, then comes agentic payments. In the agentic world, that doesn’t translate well and there is friction where a human in the loop is necessary to complete the payment. In the new world, the user could give a mandate to their AI agent, and it then should be able to execute it across different merchants. India has underlying frameworks for mandates, authentication and other guardrails; once they translate into the agentic world, we will be far advanced as a market. It is also a unified payments interface (UPI)-driven market, and the change will probably start there, and then cards would follow. 

 

Companies use different metrics to measure AI productivity, lines of code automated by AI or comparative time for GTM (go to market)... 

 

Different functions and teams track different metrics. It is driven by the output metric of what a function is supposed to do. Today when we are able to design, write the product-requirements document, hand it to engineering, look at the code draft, review it, quality control it; many parts are getting automated depending on the complexity of the code. It can go live within a few hours to a few days where some larger systems take time. Also, what has been the turnaround time to onboard a merchant, address and resolve queries, becomes some of the metrics. 

 

How fast has it become compared to last year? 

 

We run a session every Friday where the CTO (chief technology officer) hosts engineering and product teams. Many times within that two-hour session, we are able to build a product feature that is almost ready to go live. Given that we are a financial institution, there are checks before anything goes into production, but many simple features that would have taken one or two sprints can now be completed within those two hours. More complex features take longer, largely in designing and solutioning rather than actual building.

 

Another thing we are trying to do is democratise knowledge. Earlier, a product manager or engineer carried a lot of tribal knowledge, so even a simple product required dedicated ownership. Now that knowledge is available in our internal documents to everybody. As a result, one person can work across multiple products and context-switch on demand, making product development far more efficient.

 

Mythos, with its ability to detect zero-day vulnerabilities, is also causing industry-wide concern.

 

Cybersecurity has always been a cat-and-mouse game, and AI is changing both sides of that equation. While detecting vulnerabilities — including emerging and zero-day threats — is important, the real challenge is reducing the time between discovery and remediation.

 

At PayU, our focus is on combining AI-driven detection with automated triage, risk prioritisation and remediation workflows so that critical issues can be addressed faster and more consistently. When vulnerabilities are identified — whether internally or externally — we follow a structured and AI driven response process to assess impact, prioritise based on risk, and drive resolution with appropriate urgency.

 

Do you depend on specific frontier models or AI products? This is from the perspective of the sudden suspension of Anthropic’s Fable, a few days back. 

 

We actually have opened up quite a few of these models to our team. Prosus has also built Toqan which is used by a lot of teams at PayU. It is to ensure which is the best tool for a specific-use case as well. 

 

We are experimenting with a bunch of tools. Anything we use in production for an external-facing product has to be protected through specific licences and agreements with the model providers. This is critical because if a customer-facing product depends on an underlying model, that is where a break can happen. Protecting those dependencies is the only way to safeguard ourselves. Internally, however, we are diversified enough to move from one model or provider to another relatively quickly if required.

 

Is there a need to balance AI budgets, productivity and cybersecurity?

 

Today, most companies are focused on enabling their organisations to use AI. We view spending on AI tools for engineers and product managers as a learning investment. If you ask whether the RoI (return on investment) justifies the cost today, the answer is probably no, because we have not yet shipped enough products or achieved the level of efficiency required to justify it. However, organisations like ours see this as a long-term investment. Over the next year, there will be evaluations where these investments should be directed and whether they are justified in specific areas.

 

How would companies like PayU derive value from merchants, customers or payers in general for different agentic use cases like price watch? 

 

It is still early days. Our belief is that value will come from merchants. For users, agentic is more of a convenience play where they shift from one channel to another. They would probably not be willing to pay more for a new agentic channel in particular. PayU has focused on value-added services like providing affordability solutions. This would also mean how do we enable merchants in an AI world to be at par with larger players. 

 

What we are working on is called an agent marketplace for merchants where they can download certain agents and use them for specific use cases. It could be productivity agents for automating recon, automating integration with gateways, or help them power their commerce like recommendation agents. The definition of value-added services is shifting from legacy products to more AI-driven products. One of our large bets is the agent marketplace we are building. 

 

Would these agents come with specific use cases or does that mean you provide capability to just use different models? 

 

This includes agents trained for specific use cases solving specific business cases for the merchant. The agent would be working on the merchant’s data and wherever possible using other forms of information. The intelligence layer that it can use is what the power of the agents would be. 

 

Are you seeing interest from small and medium merchants (SMBs) for this? 

 

The kind of agents SMBs would use versus large merchants would be a little different. Many of the SMBs who are not as technically savvy as larger merchants would be areas of focus, where we want to empower them with plug-and-play technology. We are working with one of India’s largest direct-to-customer (D2C) aggregators to help with such agentic flows.