Parmana

For fintech teams building agentic payments

Adopt agentic payments without losing control.

Let agents request payments. Your business decides what can happen.

AI agents can initiate payments, refunds, transfers, and other financial actions. Parmana checks each important action against your business rules before it happens.

If it is allowed, it proceeds. If it is not, it stops. The authorization is recorded so your team can verify what was approved before the action happened.

See how it works
Diagram: a payment request is sent to Parmana. Parmana checks the request against business rules and creates proof of authorization. The payment system can then verify the authorization before executing the payment.

Request → Check your rules → Prove authorization → Execute

The payment system verifies the authorization before the payment is allowed to happen.

Before money moves, three things happen.

No Direct Access

The agent can request a payment, but it never holds the credential that can move the money. The ability to execute stays separate from the agent.

One Action, One Decision

Every payment is checked against your rules. It is either allowed or stopped. There is no partial approval and no continuing after a refusal.

Proof Before Payment

Parmana signs every authorization with Ed25519 before the action executes. Your payment system, or any regulator, can verify the signature independently with the public key and decision record.

An agent can request an action. Your business rules decide whether it is allowed. Parmana makes sure the action matches that decision before it happens.

Request → Check your rules → Authorize → Execute

Your payment system knows what happened. Can you prove what was authorized?

AI agents and automated systems can initiate payments, refunds, transfers, and other financial actions. But giving a system access to make payments does not mean every payment it requests should be allowed.

An audit log can tell you what happened after the fact. It can show that an agent made a payment. It does not, by itself, prove that the payment was authorized under your rules before it happened.

The difference

After the payment

“The agent made this payment.”

Before the payment

“This payment was authorized under these rules.”

Parmana checks the requested action against your rules before it happens. Parmana uses Ed25519 digital signatures to prove every authorization independently, so the system that executes the action can verify it without trusting Parmana.

Allowed means it can proceed. Not allowed means it stops.

Start with the actions that move money.

Parmana helps fintech teams make automated financial actions checkable before they happen.

Agentic Payments

Let agents request payments while your rules decide which payments can actually happen.

Refunds & Transfers

Check refunds, transfers, and other money movements before they are executed.

Payouts

Make sure automated payouts match the amount, recipient, and rules that were approved.

Other Financial Actions

Apply the same control to any automated action where the wrong decision can have a financial impact.

The same approach can later protect other high-impact automated actions. The principle stays the same: check first, execute second.

Whatever makes the request, your rules still decide.

An AI agent, employee, application, or automated system can request a payment. Parmana checks the request before it can move money.

Who is making the request?What can go wrong?What Parmana does
AI AgentCan request a payment that falls outside its intended useCheck the payment before it happens
EmployeeCan act outside their allowed authorityCheck the payment against your rules
Third-Party AppCan request more than it shouldStop anything not authorized
Automated SystemCan act without someone reviewing every actionCheck every important action
Fraudulent RequestCan try to make an unauthorized paymentRefuse the request

The source of the request does not change the rule.

A simple example: an agent booking a flight

Imagine a customer allows an AI agent to book a flight to Mumbai for up to ₹8,000.

The agent finds a flight for ₹7,999 and requests the payment. Being under the limit does not automatically mean the payment should go through.

Parmana checks the actual payment request against the customer's rules before the payment happens.

What happens next

1. The agent asks to pay

Flight to Mumbai ₹7,999

2. Parmana checks

Does this payment match the customer's rules?

3. Parmana decides

Allowed → continue. Not allowed → stop.

4. The payment system verifies

The payment can proceed only when the authorization is valid.

The important question is not only “Did the agent have permission?”

It is “Was this payment authorized before it happened?”

Why this matters now

AI is moving from making recommendations to taking actions. In financial systems, that makes controlling each action just as important as making the decision.

AUGUST 2026

AI agents can take unexpected actions

AI security evaluations have shown agents can take actions beyond what people expect when pursuing a goal. An agent may find its own way to achieve an objective, including actions that were never explicitly approved.

What this means:

Giving an agent a goal does not mean every action it takes should be trusted.

2026

Financial services are moving toward agentic AI

Financial institutions are moving from AI that recommends and assists toward systems that can perform tasks and workflows. As financial actions become more automated, controlling the action becomes as important as making the decision.

What this means:

The question is no longer only what AI recommends, but what it is allowed to do.

2026

India is preparing for agentic payments

India's payments ecosystem is moving toward greater use of AI and automated payment experiences. NPCI is building AI capabilities for the payments ecosystem while UPI continues to expand delegated and automated payment capabilities.

What this means:

AI is moving closer to real payment workflows in India's financial system.

JULY 2026

RBI is putting stronger controls around AI

RBI's 2026 draft Model Risk Management guidance includes human oversight for AI models, including mechanisms to override, suspend, or deactivate models. This explicitly includes kill-switch arrangements.

What this means:

Financial institutions need the ability to intervene when an AI system behaves unexpectedly.

A kill switch stops a system. Parmana helps control the action before it happens.

The goal is not to stop agentic payments. It is to let fintech teams adopt them without losing control.

Parmana does one important thing

It makes sure an important action is authorized before the action happens.

Parmana is not

  • An AI safety product
  • A policy documentation tool
  • An audit log
  • A risk score
  • An approval inbox

Parmana is

  • A check before money moves
  • A way to keep execution separate from the agent
  • A decision your payment system can verify
  • Proof of what was authorized

Control stays with your business.

Your customers, agents, and applications can request actions. Your business rules decide what is allowed, and only authorized actions are allowed to happen.

Questions fintech teams ask

Straight answers about keeping control, proving authorization, and adopting agentic payment flows.

What problem does Parmana solve?

It keeps you in control when AI agents, applications, or automated systems can initiate financial actions. Parmana checks every important action against your business rules before that action happens, not after.

What if the agent is compromised?

Parmana does not need to trust the agent. Say a compromised agent asks to send ₹500,000 to a new account. It can still only make a request, the same request any agent would make. That request is checked against your business rules before execution, the same as any other. The agent never holds the credential that could move the money itself.

What if our guardrails fail?

An agent's own guardrails are never the final authority. Say a duplicate-refund guardrail fails silently and the agent asks Parmana to issue the same refund twice. Parmana checks that second request against your rules regardless, the same way it checks the first one, and refuses it if it does not match. Your protection does not depend on the agent behaving correctly.

What if a smarter frontier model finds a way around the rules?

Parmana does not rely on the model being harmless or predictable. A more capable model might phrase a request differently or reach it by a different chain of reasoning, but Parmana only evaluates the request itself, the amount, the recipient, the timing, whatever your rules check, never how the model arrived at it. Capability does not become authority.

What if we change AI vendors, switch models, or turn AI off?

Your authorization does not depend on which AI model or vendor you use. Parmana sits outside the agent's decision-making and checks the action against your business rules. You can change models, replace an agent, change vendors, or turn AI off without changing the authority that controls what can happen.

Does Parmana stop agents from making payments?

No. If a payment matches the rules you have set, it proceeds. If it does not, it is refused. Parmana is built to help fintechs adopt agentic payment flows, not to block them.

What happens if an agent behaves unexpectedly?

Every payment or other important action is checked before execution, every time, not just the first one. If a request no longer meets your rules, Parmana stops it there rather than letting the agent continue.

Is Parmana only for AI agents?

No. A refund request typed into an internal admin tool by an employee goes through the same check as one an AI agent requests. The source of the request, agent, employee, application, third-party system, or automated workflow, does not change your business rules.

How is this different from an audit log?

An audit log tells you, after the fact, that an agent sent a payment. Parmana tells you, before it happened, that the payment matched your rules, and records that decision as evidence. One is a record of what occurred. The other is proof of what was allowed to occur.

Can Parmana help with auditability and compliance?

Yes. Parmana creates verifiable evidence for each authorization decision: what was requested, what rules were applied, whether it was allowed, and what was authorized before execution. It does not replace your compliance program, audit systems, or regulatory obligations. It gives them evidence to work with.

How is this different from a kill switch?

A kill switch stops a system, or all of it at once, when something goes wrong. Parmana checks one request at a time: it can let 999 out of 1,000 requests through and stop only the one that breaks a rule, without touching the other 999.

Can Parmana run in our own cloud or infrastructure?

Yes. Parmana runs as a standard container and can be deployed on your own cloud account or infrastructure you control, not only ours. Your business rules and signing keys stay in your deployment, not in ours.

What if today's cryptography changes tomorrow?

Parmana separates authorization from the technology used to prove it. As cryptographic standards evolve, the proof mechanism can evolve too, without changing the business rules that decide what is allowed.

How do I know a signed authorization is real?

Parmana signs every authorization with Ed25519. Anyone can verify it independently by checking the signature against the public key and the decision record, no special tools, and no need to take Parmana's word for it.

What if Parmana disappears tomorrow?

Your business should not lose control because a vendor goes away. Parmana is built around explicit authorization rules and verifiable evidence that live in your deployment, not around Parmana itself being the permanent source of authority. Your authority belongs to your business, not to Parmana.

Does Parmana replace our existing payment or authorization systems?

No. Your existing payment infrastructure and business rules stay exactly where they are. Parmana adds one control point in front of them, checking important actions before they execute.

Can Parmana be used outside payments?

Yes. Say an agent requests revoking a user's database access rather than issuing a refund. Parmana checks that request the same way it checks a payment, against your rules, before it happens. The same approach covers refunds, transfers, access changes, infrastructure operations, data movement, and other automated business workflows.

How do I know the claims on this page are accurate?

Every technical claim on this page describing how Parmana works was checked, one by one, against Parmana's own implementation before publishing, not written by a marketing team and left unverified. Read the validation write-up. Parmana's implementation itself is private, so the write-up documents what was checked and how, rather than something you can independently re-run against the source yourself.

More intelligence should not mean more authority.

Your AI can change. Your models can change. Your cryptography can change. Your business should still control what can happen.

Adopt agentic payment flows without losing control.

We help fintech teams design and test payment flows where AI agents can act within clear business rules, without giving up control over what actually gets executed.

Talk through your payment flow

Working with fintech teams building agentic payments.