How Pine Labs is building for the agent economy: CTO Sanjeev Kumar on trust, control and execution

Pine Labs is building an AI-powered commerce operating system for businesses that want agents to handle real work. The system combines agent permissions, action tracking, enterprise applications, and developer tools. It matters because agents are moving beyond answering questions and beginning to process complaints, support teams, and run campaigns.

The platform already supports internal compliance, finance, HR, and marketing workflows. Its Cora agent handles cyber-fraud complaints, while an ads agent helps small merchants create campaigns. Grantex controls permissions and records actions, and AgenticOrg gives developers visibility into agents, confidence scores, and audit trails.

Pine Labs is extending the system beyond its own operations. Banks in India and overseas are implementing it, while the Agent Store is intended to support third-party developers. The broader goal is an ecosystem where merchants, banks, and enterprise teams can use agents without giving up accountability or human control.

An AI agent is software that can pursue a defined task by using connected tools, following instructions, and taking actions. Unlike a chatbot that only generates answers, an agent can move work through several steps and interact with business systems. This distinction matters when the task involves records, approvals, notifications, or transactions.

Cora shows the difference. It takes a cyber-fraud complaint, matches it to transactions, categorizes it, routes it to the right teams, and prepares a response. It can also use robotic process automation to access a government cyber police portal and fill in an action-taken form.

This does not mean agents operate without limits. Grantex assigns permissions, expiry dates, and authorizers, while Cora records a 15-step trace. Humans still review important actions, such as submitting the portal form. The article presents agents as supervised systems that execute work, rather than chat interfaces that only answer questions.

Pine Labs’ effort spans internal operations, financial institutions, merchants, and outside developers. Over 18 months, it built an agent-focused commerce operating system and first deployed the technology across compliance, finance, HR, and marketing. This gives the company experience using agents inside its own environment.

The effort now reaches banks in India and overseas, merchant partners, and developers. Cora supports cyber-fraud resolution, while a merchant-facing ads agent helps smaller businesses create campaigns. Developers can write skill files, connect tools, and define tasks in plain English through the AgenticOrg platform.

Pine Labs also organized a hackathon with The Ken that attracted more than 380 agent submissions from students, fresh graduates, and engineers across India. The Agent Store is planned to open the ecosystem to third-party developers. The article therefore describes both active deployments and an expanding platform, rather than a single experiment.

Grantex is Pine Labs’ control layer for AI agents. Kumar described it as “Open Authorization 2.0 for AI agents.” It defines what an agent may do, how long that authority lasts, and which person approved it. This makes an agent’s operating boundaries explicit before it begins work.

For example, an agent authorized to read and draft emails cannot send one automatically. Grantex can block that attempt and revoke the agent’s permissions, including the permissions of any child agents. A tamper-proof, hash-chain audit trail records the agent’s actions, creating a detailed history for review.

These controls are intended for banking and enterprise settings, where organizations need to understand and audit automated decisions. Pine Labs also uses traces, evidence, and confidence levels in its applications. Together, the controls allow agents to act within a defined scope while giving people ways to stop misuse and investigate what happened.

Cora is Pine Labs’ compliance and resolution agent for cyber-fraud complaints. It takes cases from intake through closure, matching each complaint to transactions, categorizing it, routing it to the relevant teams, and preparing responses for review. This connects several parts of the process in one workflow.

The speed gain is significant. Pine Labs said some complaints can now close on the same day, compared with a typical turnaround of three to five days. Cora records a 15-step trace covering validation, transaction lookups, settlement checks, team notifications, supporting evidence, and confidence levels.

Cora can also use robotic process automation when a system connector is unavailable. In a demonstration, it accessed the government cyber police portal and filled out an action-taken form. A human still controlled the password and one-time code and reviewed the form before submission. Faster handling therefore comes with documented steps and retained human oversight.

Human intervention remains at the points where an automated action could create significant consequences. Pine Labs’ approach lets agents prepare and execute much of the workflow, but people retain authority over credentials, final submissions, and campaign approval. This keeps automation from becoming an unchecked decision-maker.

In Cora’s demonstration, the agent used robotic process automation to access the government cyber police portal and fill in an action-taken form. A human operator kept the password and one-time code private. The operator also reviewed the completed form before it was submitted.

The ads agent follows a similar pattern. It researches local demand and competitors, then creates copy for search, display, and YouTube. Merchants can ask for changes and approve the campaign before launch. The article shows agents preparing work and handling routine steps, while humans retain control over sensitive access and public-facing decisions.

Banks and enterprises need more than an agent that can produce plausible answers. They need to know which systems it can use, what information supports its decisions, whether its behavior remains reliable, and when a person must step in. Permissions define authority, while audit trails make actions reviewable after the fact.

Pine Labs also highlights retrieval-augmented generation, tool and system connections, continuous evaluations, and decisions about human intervention. These pieces help an agent find relevant information, use business systems, test its performance repeatedly, and stop or escalate work when automation is not appropriate. Cora’s evidence, confidence levels, and action trace illustrate this approach.

The company first deployed these capabilities across its compliance, finance, HR, and marketing teams inside its own environment. It is also implementing them with banks in India and overseas. The article’s central point is that reliable enterprise agents require a coordinated control system, not just a language model.