Omise AI: Humanizing merchant onboarding with a conversational agentic AI on AWS Bedrock

Omise (Opn Payments) is a payment gateway serving merchants across Southeast Asia. Merchant onboarding depended on manual coordination across CX, Sales, KYC, and Operations, which was slow for merchants and costly for internal teams. This experiment asked whether a conversational agentic AI could make it self-serve.

This was the third AI product experiment at Omise, and the most human-centred one: an attempt to humanize AI in a fintech context. Working closely with the AWS Thailand team on Amazon Bedrock, we built a conversational agentic AI that guides merchants through payment setup and API configuration, and answers the basic product, pricing, and KYC questions that previously required arranging time across multiple internal teams.

I led the design end to end, from stakeholder research and service blueprinting to defining where AI should appear across the merchant journey and how the conversation should behave at each step. Because onboarding cut across CX, Sales, KYC, and Service Operations, much of the work was mapping the as-is process before proposing any AI touchpoint.

For this case study, I focus on the research foundation, the service blueprint with its AI-opportunity layer, and the phased design of the conversational agent.

Timeline

Q1 2025 - Q2 2025

Skills

UX Research Service Design Conversational AI Design UX Design

Client

Omise

My Role

Product Designer

Design Team

Product Designer: Watis Kotitrakool (Main owner) Engineering Manager: Sylvain Dormieu UX Researcher: Kun Wattanayothin In collaboration with AWS Thailand solutions team

Website

omise.co/en/omise-ai

What we wanted to achieve

Make AI the first touchpoint of onboarding
Let qualified merchants onboard end to end, from first product questions to payment setup, API integration, and KYC, without waiting for a team to be arranged.
Deflect repetitive inquiries
Research showed over 90% of support tickets were transaction and service enquiries, and a third of call volume was basic product and KYC questions whose answers already existed. The agent should answer these instantly.
Shorten merchant time-to-live
Merchants under integration deadlines went with the provider that replied fastest. Instant guidance and proactive KYC status updates directly protect conversion.
Earn trust with a humanized, transparent AI
In a financial product, the agent must communicate clearly, show what it did and why, and hand over to humans gracefully, so merchants can verify its work rather than take it on faith.

Constraints at the start

Onboarding knowledge lived in people, not systems: CS answers in Help Scout, KYC rules with analysts, and sales context spread across HubSpot notes and Google Sheets.
KYC communication was long, text-heavy, and English-only, making it hard to digest for Thai merchants.
Teams worked across disconnected tools (Help Scout, HubSpot, admin dashboard, Google Sheets, Jira), with no single source of truth to power an agent.
In a regulated payment product, a wrong answer about KYC or fees is worse than no answer, so the agent needed clear boundaries and graceful human handover.

Key solutions

Grounded the problem in field research
Ran stakeholder interviews and workshops with the CS ticket owner, the call-center team, Sales, Account Management, and Solution Consultants. Mapped the as-is flows (lead nurture, self-serve onboarding emails, the KYC review loop, and account activation) and pulled ticket and call statistics to size the repetitive-inquiry problem.
Synthesized personas and the merchant journey
Built an internal persona (a CX lead buried in repeated basic inquiries and manual processes) and an external one (a finance manager onboarding through KYC). Mapped the full Payment Gateway Onboarding Journey with pain points per stage, and defined which merchant profiles suit self-serve onboarding versus sales-led.
Designed the service blueprint with an AI layer
Mapped the journey against front-stage (CX / KYC / Sales), back-stage (Service Ops, Solution Consultants), and supporting systems, with an explicit AI Role & Opportunity layer marking where the agent takes over. Every pain point was classified as solvable by chatbot or not, keeping the scope honest.
Phased the agent by journey stage
Phase 1: support chatbot across pre-sale, onboarding, and post-sale, plus intelligent requirement collection. Phase 2: KYC processing, with self-serve document recheck, guideline feedback, and proactive status notifications. Phase 3: transaction monitoring, generative reports, and intelligent dashboards.
Built with AWS Thailand on Amazon Bedrock
Used retrieval over Omise's product and integration documentation for accurate answers, with guided conversational flows for payment setup and API configuration, replacing repetitive, team-arranged onboarding steps with a single conversational touchpoint.

OMISE AGENT

Meet your personal AI assistant, designed to empower and simplify your workflow

Final outcome

The experiment validated that the highest-volume, lowest-complexity work in onboarding (basic inquiries, KYC guidance, and status updates) can shift to a conversational agent, freeing CX, Sales, and KYC teams for complex cases and revenue-generating conversations. Just as importantly, it set a pattern for humanizing AI in a financial product: transparent about what it did, honest about what it can't, and designed to hand over to a human at the right moment.

Impact delivered

Sized the opportunity with data: over 90% of support tickets were repetitive (56% transaction, 35% service enquiry), and 61% of call volume was product, service, and KYC questions the agent can answer instantly.
Delivered a service blueprint with an AI-opportunity layer that aligned CX, Sales, KYC, and Operations on a three-phase agent roadmap.
Built a working conversational onboarding agent on Amazon Bedrock with the AWS Thailand team, covering payment setup, API guidance, and basic Q&A.
Shifted team effort toward revenue-generating work by targeting manual tasks like payout-reconcile support (around 45 minutes per case) for AI assistance.