The Latest Payments Battle Point: AI Agents For Merchants

Almost every Indian payments giant delivered a similar pitch at the Global Fintech Fest 2026 organised last week: merchants no longer need to inspect settlements, chase failed transactions and untangle disputes. An AI agent will do it instead.

Five of the country’s largest B2B payments platforms introduced agents that can recover lost sales, reconcile chargebacks, pursue refunds and execute transactions, turning AI from an assistant into a protagonist in the payments chain. 

For starters, BharatPe, a merchant payments platform, launched an Agentic AI assistant built on Google Cloud’s Gemini Enterprise that works across more than 60 systems to resolve complaints and chargebacks, while also recommending products such as credit cards on UPI and loans. 

Similarly, digital payments provider PayU launched Agent HQ, a store where merchants deploy one agent per job. Cashfree also rolled out an agent for payment operations, Relay. Razorpay put an AI account manager named RAY on WhatsApp with IndusInd Bank. Meanwhile, Pine Labs went a step ahead, allowing an agent to complete the purchase itself.

Pine Labs is now powering an agentic marketplace for L&T Finance. Cashfree’s Relay is live for all its merchants. PayU’s Agent HQ and Razorpay’s Agent Studio stay in early access. 

Together, these launches suggest that AI is moving beyond assistance. The questions are: who needs them, what work they can take on, and how much control businesses are willing to give them.

What Are These Agents Up To?

Proving adept at recovering failed payments, confirming orders, filing disputes and reconciling chargebacks, payments giants are building agents to take work off a merchant’s plate.

Describing the shift, Reeju Datta, the cofounder of Cashfree Payments, said that most merchant AI worked like a copilot a year ago, preparing something for a person to review or send. “Now, the brief has changed. Agents are being built to complete the task themselves,” he added.

He explains a widening list of jobs an AI agent can now own:

• Recovering abandoned carts and retrying failed payments
• Confirming cash-on-delivery orders before dispatch
• Managing failed subscription renewals
• Filing disputes before the deadline
• Reconciling chargebacks, escalating refunds and following up on settlements

Some have upped the ante. At GFF, Pine Labs demonstrated an agentic purchase with Big Basket, where a customer could ask an agent to track the price of an iPhone and complete the purchase once it fell by 40%. While it may look like a consumer use case, it benefits the merchant directly.

Tanya Naik, who heads online and omnichannel business at Pine Labs, pointed out the commercial logic behind the push, noting that a human-run payment never quite reaches complete success, while an agent-triggered transaction is captured against an amount already authorised, so the success rate is structurally higher. 

The same mechanism opens a second use case: re-engaging dormant customers. A deal seeker who never returns to a storefront can set an agent to buy the moment a price condition is met, turning the agent into another channel alongside Instagram, Facebook and Google ads.

Who Needs These Agents? 

Not every merchant needs an AI agent. The first market is likely to be businesses where there is plenty of payment-related work, but very few people to handle it.

Cashfree, for instance, has built Relay with small and medium businesses (SMBs) in mind, particularly lean teams of around five people with no dedicated payments or finance function. 

PayU’s chief product officer, Manas Mishra, points to the same group, naming SMBs and mid-market merchants as the most relevant segment. Early interest is from D2C brands automating chargeback reconciliations, refund escalations, settlements, and transaction follow-ups, while also using agents to stay discoverable where their customers now search.

Cashfree estimates a single agent working across its merchant base could recover as much as ₹20,000 Cr in sales lost to failed payments every year.

For now, however, these agents are mostly recovering money and handling routine work rather than running the business. This may change as they are given more autonomy.

How Much Autonomy Should Agents Enjoy?

The more work merchants hand to agents, the more important the boundaries become. This is why payments platforms take different approaches towards setting a control layer.

• Cashfree’s Relay, for instance, only accesses a merchant’s own transaction data. It also has two checkpoints that an agent cannot cross without approval: before money moves and before a customer is contacted.
• PayU’s Agent HQ sits within the merchant’s existing dashboard, with the merchant setting the operational limits. Its agents cannot act beyond those boundaries.
• Pine Labs has taken a different approach with Grantex, an authority and trust layer built into its payments protocol. It checks each transaction against the mandate given to the agent and creates an audit trail that can be used to trace a transaction or resolve a dispute.
• Razorpay’s Agent Studio follows a review-first approach. Large transfers and deletions require a second confirmation and cannot be approved automatically.

Despite the differences, the principle is broadly the same: agents can prepare, decide and execute routine tasks, but actions that move money, contact customers or cannot be reversed remain subject to human control.

However, these boundaries are already starting to stretch.

Cashfree’s Relay can generate custom agents once a merchant defines the business logic and desired outcome, moving beyond a fixed set of pre-built tasks. Pine Labs, meanwhile, wants to extend this self-service model to banks, allowing merchants to ask about a payment problem and have an agent resolve it.

In other words, the agent is evolving from a feature into a layer that merchants can build on. However, if an agent makes a mistake, who bears the loss? Will it be the merchant, the payments platform or the bank? For now, there is no clear answer.

Top Stories From India & Around the World

• Google Picks Indian Startups For AI For The Planet: Google has selected four Indian startups – Terrastack, Varaha Climate, Farmers for Forests and Climitra Carbon – for the maiden cohort of Google DeepMind Accelerator: AI for the Planet (APAC) accelerator initiative. Over three months, they will get access to Google’s AI stack, technical support and mentorship to scale sustainability-focused AI solutions.
• Lenskart Deepens AjnaLens Bet: The omnichannel eyewear brand has acquired an additional 1.8% stake in XR startup AjnaLens’ parent Dimension NXG for ₹8 Cr, taking its holding to 9.01%. It has invested ₹18.5 Cr across three purchases this year, as it pushes its AI-powered smart eyewear play alongside its Phonic audio glasses and a Qualcomm partnership.
• DeepSeek Releases V4.1 Flash: The AI lab has launched the smallest model in its new architecture family with native multimodal visual understanding. The 552B-parameter MoE uses an asymmetric causal encoder-decoder design with just 8B active parameters on input and 16B on output, and DeepSeek says it beats V4-Pro on performance, cost, speed and runtime, with lower API prices.
• OpenAI Pauses Pro Signups: OpenAI has put new Pro subscriptions on hold as demand for GPT-6 Astra, its first cyber-critical model, outstrips capacity, with enterprise access prioritised before wider ChatGPT availability.

The Weekly Buzz: We Must Pace the Frontier

Anthropic CEO Dario Amodei published a detailed public statement titled “We Must Pace the Frontier,” arguing that frontier labs must deliberately slow the rate at which they improve model capabilities so that safety work can keep pace. He was careful to distinguish this from a full halt: training and technical progress would continue, but companies would take adequate time to align and safeguard systems, with independent confirmation required before further leaps. Two developments drove the shift. 

First, recursive self-improvement has accelerated sharply since summer, with AI systems increasingly helping build the next generation of models. Second, recent agent incidents, including OpenAI’s swarm that breached external systems — showed that misaligned collective behaviour is already possible. Amodei warned that a more capable version of such a swarm could, within 6-12 months, assemble a persistent botnet across large parts of the internet.

His three-step plan begins with a unilateral commitment Anthropic is implementing immediately: giving third-party evaluators permanent, employee-level access (desks, badges, laptops) and the right to publish findings without editorial control. The second step calls for coordination among labs in democratic countries on common safety standards and limits on unchecked progress. The third, hardest step involves eventual global coordination, including with China, while democracies maintain their technological lead through export controls and security measures.

Sam Altman, Elon Musk, and later Satya Nadella publicly endorsed the call. Altman confirmed OpenAI would also adopt embedded evaluators. The unusual alignment among rival CEOs, coming amid researcher resignations and a string of agent incidents, marks the first time the leading US labs have so openly agreed that capability gains must be deliberately paced rather than maximised at all costs.

Startup in the Spotlight: Flash AI

Founded in April 2022 by Flipkart VP Ranjith Boyanapalli, Flash AI is building an AI-powered shopping assistant that delivers instant, AI-generated insights to help shoppers buy with confidence. 

Online shopping offers endless choices, but too many options can overwhelm consumers. From reading product reviews on platforms to following YouTube influencers, shoppers often spend hours trying to decide whether a product is worth buying.

The process is confusing, time-consuming and prone to information overload. Instead of juggling multiple marketplaces and manual research, Flash AI brings detailed reviews, pros and cons, price comparisons and purchase recommendations into one place.

By integrating data from multiple ecommerce channels, the assistant lets shoppers cut through the noise and make faster, better-informed decisions. Flash AI claims to cater to users across 100+ countries, tapping into a growing global ecommerce market hungry for tools that simplify the buying experience.

Prompt Of The Week

What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work? 

Here’s the prompt Niraj Nagrani, the chief data and AI officer at Altimetrik, uses to walk into client conversations. The company offers an integrated AI-powered enterprise software suite called AIOS. 

“Before this week’s client conversations, pull together a full picture of where AIOS stands.

Start with the deployment question: how a Gemini/GCP setup compares to OpenAI/AWS and a native Anthropic stack, and where the AIOS Substrate actually saves a buyer from vendor lock-in. 

Look at it across:

• Latency
• Governance surface
• Unit economics

From there, turn the AIOS Governance story into something a security architect would actually sign off on, not marketing language, hitting audit, guardrails, eval, human-in-loop, and model risk as concrete checklist items.

Then scan the Data & AI tracker from the last 24 to 48 hours and surface only what genuinely shifts our positioning or what a client CxO is likely to raise this week.

Finally, work through the YTD pipeline and identify:

• How many Data and Agentic deals carry real stated pain tied to the AIOS layer
• What single slide would land best in a 20-minute exec conversation”

Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.

[Edited by Shishir Parasher]
[Creatives by Varshita Srivastava]

The post The Latest Payments Battle Point: AI Agents For Merchants appeared first on Inc42 Media.