Agentic payments — in action

“Send $50 to Sarah”.

That’s the entire payment flow

Our client replaced payment forms with a text message. Together, we built a fully conversational AI agentic payments while the industry was still setting standards.

AGENTIC PAYMENTS — IN PRODUCTION

Built with our
fintech customer

Production-ready AI banking Agent with payment guardrails built in. No forms, no friction, no failed transactions.

5

months from kick-off → production

40+

countries covered

0

errors and no hallucinated transactions

WHAT WE SOLVED

The most important in Agentic Payments? Anti-hallucination guardrails

Zero dollars moved by hallucination.
That is not a claim, it is architecture.

01

Zero dollars moved by mistake

Hallucination in payments means money going where it should not. Every transfer is validated against the identifiers the user confirmed, so the agent can never move money to the wrong place.

02

Guardrails at the system level

The AI cannot catch a wrong IBAN or corrupted identifier on its own. Our guardrails work independently of the model, blocking mismatches at the system level.

03

Uninterrupted chat

KYC, OTP flows, and real-time FX across 40+ countries, with zero tolerance for error. A provider-agnostic architecture scales without interrupting the conversation.

5 months to prove
agentic payments
work. We hit it.

The deadline was a live demo in front of a strategic partner. The goal to show that onboarding, KYC, funding, and transfers all happen in a single conversation.

Everything happens in one chat. No jumping between dashboards, no filling out forms, no approval steps.


Marek Gajda
The Software House

Marek Gajda

SYSTEM DESIGN

From “send $50” to
settled transaction

How a conversational message becomes a real cross-border transfer with guardrails at every layer.

01

Payment flow

A user message enters the AI orchestrator, which detects intent and routes the request to the right specialized agent: Sign-up, KYC, Transfers, Funding or Support.

Every agent output passes through the guardrail layer before reaching a payment provider. Payloads are hashed on both sides, if they don't match, the transfer is blocked

No LLM output ever calls a payment API directly.

WHAT WE SOLVED

Four make-or-break decisions in agentic payments

01

Hallucination-proof transactions

LLMs can be confidently wrong. User-confirmed values are hashed. Mismatches never reach the payment API.

02

KYC in conversation

Traditional KYC breaks the conversation. We moved the entire flow in chat. No forms, no redirects, no drop-offs.

03

Provider-agnostic architecture

We standardized all provider communication through a single MCP gateway. New market, new provider — minimal changes needed.

04

Async transactions, uninterrupted UX

Event-driven workflows handle each step independently. The agent always picks up where it left off.

TECHNOLOGY

The stack behind a
production

Stack tool
n8n
Model Context Protocol
Redis
PostgreSQL
Notion

OUTCOMES

Shipped in 5 months.
Live across 40 countries.
Zero hallucinated transactions.

5

months from first line of code to live demonstration to a strategic financial partner

40+

countries cross-border transfers operational at launch

0

model-generated errors resulting in incorrect transactions

4

engineers: the team size that helped deliver it

We build agentic
payment systems.
Let's talk about yours

Your personal data will be processed in order to handle your question, and their administrator will be The Software House sp. z o.o. with its registered office in Gliwice. Other information regarding the processing of personal data, including information on your rights, can be found in our Privacy Policy.