Looking For The Best Way To Approach Agent Autonomy? Match The Scope To The Stakes

Enterprise AI agents are starting to transact with agents from other companies. A customer’s agent places an order, a merchant’s agent responds, and payment moves between them without a person signing off on each step. For CIOs, agent governance now has to extend past their own systems to every agent in the chain.

Payments is where this is happening first. In India, protocols already let agents complete payments after a single upfront approval, giving technology leaders an early look at the guardrails agent-to-agent commerce will need.

Vijeth Pandit is the Chief Product Officer for Payments and Fintech Infrastructure at Pine Labs, a payments technology company founded in 1998. It began with in-store card terminals and now runs gift card issuing, online payments, and real-time payments infrastructure across India, Southeast Asia, and the Middle East. In 2026, it launched one of those agent payment protocols on UPI, India’s instant payments network. Before Pine Labs, he held product roles at Indian fintech Razorpay and spent five years at Amazon, after earlier stints at healthtech startup Practo and telecom company Reliance Jio. He sees governing agents as a management task companies already know how to do with people.

"When you hire people, you regularly audit whether they're working within the guardrails you trained them on. It's almost the same with agents," Pandit said. The same habits apply when the agent on the other side of a transaction belongs to another company, and every party needs a clear view of whose agent is doing what.

• Personal service, rebuilt: Online shopping traded the attentive salesperson for convenience. Beyond suggesting a preferred payment method or routing a transaction to the right processor, checkouts had little room to tailor the experience to each buyer. "The commerce experience itself had become one size fits all in some way. Now that’s going to change," Pandit observed. Merchant agents can respond to each customer’s agent with answers specific to that request, bringing back the conversational selling that faded online. That upside is pulling businesses in quickly, and it raises the stakes for getting permissions right at every handoff.

• Rules set in advance: Transactions between agents depend on delegated authorization. A consumer or business approves once, up front, and the agent acts within the spending limits and permissions attached to that approval. Recording it creates an audit trail that can be checked after the fact. "You define the guardrails, and you also do the accountability check. That’s the number one thing that will be needed," Pandit noted. Buyer and seller agents can work from that same record, so neither side has to take the other’s word about what was authorized.

• Who owns the outcome: Companies also need to decide who is accountable when a transaction goes sideways. A customer’s intent and an agent’s reading of it won’t always match, so liability needs to be assigned ahead of time. “Earlier, a business only had to worry that a fraudulent customer could play around, and vice versa. Now there are four players in the mix,” Pandit added.

Those cross-company rules rest on an earlier decision about which work agents should handle at all. AI can help with a long list of tasks, from developer productivity to customer-facing processes like onboarding, refunds, chargebacks, and support. Not every one of them needs an agent, and Pandit sorts them with one consistent test.

• Judgment calls only: Some tasks follow fixed rules and should give the same result every time. Others need someone to weigh the situation and make a call. Automation fits the first kind, and agents fit the second. "We should not be putting agents in deterministic things. What can be solved through automation doesn’t need AI," Pandit said. Reconciliation shows how this works. Checking whether 100 orders match 100 payments is a simple rule, so automation can handle it. Figuring out why a payment is missing and deciding whether that’s okay is a judgment call, usually made by a finance supervisor, and that’s where an agent helps. Small businesses that can’t afford an ad agency can use an agent the same way, to decide where their ad budget should go. In markets like India, the main benefit is fewer costly mistakes, more than lower staffing costs.

• Autonomy on a probation period: Even within that second group, some tasks need more human oversight than others. An agent can plan a small business’s ad spend, while writing the ad itself takes a sense of taste that develops over time. The safer path is to start agents on tasks with more predictable outcomes and keep a person responsible for the final call. Open-ended work is also harder to evaluate, so an agent’s scope should widen in stages. "When someone joins a team, you spend the first 30 days training them. Once they’re comfortable with two-way door decisions, you take a step back, and once they’re seasoned enough for one-way door decisions, you step back more," Pandit explained. Agents can follow the same path, only much faster.

• Fixed goals, flexible tools: New models and agent tools arrive every few weeks, and chasing each release makes it easy to lose sight of what customers need. The steadier approach is to anchor on outcomes that haven’t changed through any wave of payments technology. Customers want transactions to go through and refunds to arrive on time, with no fraudulent chargebacks. They don’t want to be charged while the merchant says the payment never came. "In a world where there’s going to be change every day, the only way to stay on top is to focus on the constants. Those north star metrics are not changing. The means to the end will keep changing," Pandit noted. With those goals fixed, teams can test proprietary and open-weight models side by side and switch between them without disrupting the customer experience.

The groundwork companies lay for agent commerce now will shape what agents can do next, and personal finance may be where it matters most. Someone managing a credit card, a home loan, and a car loan can already get AI-generated insight into how they’re doing. The next step is an agent that acts on that insight within limits the person sets, which could do the most for financial inclusion in emerging markets.

Reaching that point depends on the accountability companies build into agent transactions first. "The goal is for AI to go beyond educating people and telling them what they could have done, to the point where it says, 'I’ve done this for you.' The question is how we do that in a safe way," Pandit said.