Introducing Arnab: the AI Operator by OnePay, for OnePay

9:14 AM.

The Zoom call ends. Twenty-nine minutes on Q3 priorities. Three open questions nobody wrote down. Two action items assigned verbally to people who have already jumped to their next meeting. A decision that needs to reach four people who weren't on the call.

You make a mental note to send out the Zoom AI transcript later. Slack has nine new messages. One of them is a Datadog alert — latency spike on the payments service. Probably nothing, but you should check. Your calendar flashes a reminder that you have one minute until your next meeting, and so you leave the other eight unread.

This is the part of the workday that doesn't show up in job descriptions: connecting the dots. Following up, routing information, tracking down context, making sure the right things reach the right people. It isn't hard. It's just constant. It fragments everything around it, and it doesn't stop.

This is where Arnab steps in.

The Cold-Start Problem

Frontier AI models are remarkable. They can reason across long contexts, write production code, and synthesize information faster than any human. But drop one into a real company on day one and a quieter problem shows up: the model is brilliant in the abstract and a stranger in your environment.

It doesn’t know your business, what products you build, or the type of culture you have. It doesn't know your mission, your values, how your org is structured. It doesn't know what "Tokki" is, who runs the Risk team, or which Slack channel is the right one to ping when payments latency spikes. And even if it did, most of the systems your work actually lives in — Slack, Gmail, Drive, GitLab, Datadog,  the ERP — speak their own languages. The model has no way in.

That gap is the cold-start problem, and it's the single biggest reason AI adoption stalls inside organizations. The technology is ready. The context isn't. Employees try the tool, hit one too many road blocks, and quietly go back to doing things the old way. You end up with a company where a small percentage of employees use AI well and the rest watch from the sideline.

We built Arnab to close that gap.

What Is Arnab?

Arnab is OnePay's internal AI Agent. It's every employee’s coworker. It shows up as a web app and natively inside Slack. It pairs the intelligence of frontier models with five things those models lack. Institutional knowledge, integrations into our tools, model choice, a place to share what we build, and memory.

We think of those as Arnab's superpowers. Each one builds the moat that separates Arnab apart.

Arnab (أرنب) means rabbit in Arabic. Tokki (토끼), our developer agent, means rabbit in Korean. The naming is deliberate. 

There's a story we tell at OnePay. A fox is out hunting and catches sight of a rabbit. He chases hard — and the rabbit gets away. When the fox returns home empty-handed, his family is astonished. "You're a fox. How did you let a little rabbit get away?" The fox shrugs: "I was chasing my dinner. The rabbit was running for its life."

That's the spirit we embody at OnePay, and it was only fitting that our AI coworkers embody that spirit too.

What Sets Arnab Apart

1. Knowledge: It Knows OnePay

The first question any new hire has to answer is "where am I, and how does this place work?" Arnab arrives with that answer already loaded.

On the organizational side, Arnab knows our mission and vision, our core values, our org structure, our HR policies, and how we work as a company. On the product side, it knows our products and features, our value props, and the FAQs customers ask. When someone asks "what's the eligibility for Banking+?" or "who owns the Pay Later roadmap?", Arnab doesn't have to guess. It answers from the same source of truth our teams use.

That sounds simple, but it's the difference between a chatbot and a coworker. A coworker has context. An assistant that knows the company can give you the right answer the first time, in the right tone, with the right caveats. That trust, answering “who owns OnePay Later?” the same way a teammate would, is what gets people to come back.

2. Integrations: Connected to The Tools We Use

Knowledge gets you started. Integrations are what let an AI agent actually do work.

Arnab is connected to the systems where OnePay's employees live: Slack, email, documents, spreadsheets, presentations, expense management, finance tools, legal systems, data analytics, production support tools, code, and a growing list beyond that. Today, Arnab supports hundreds of tools across dozens of integrations.

Most of those tools didn't have a way for an AI agent to access them when we started. We built them. And the hard part wasn't churning out Model Context Protocol (MCP) servers — we had coding AI agents on our side. The hard part was capturing the actual functionality of the tools behind each integration. A Slack integration isn't useful if it can only post a message. We needed Arnab to be able to send messages, with or without attachments, search across channels, read and summarize threads, schedule follow-ups, operate on canvases and more. A document integration isn't useful if it can only read. Arnab needed to read, edit, leave comments, summarize and resolve feedback. The depth in each integration is the product. Which is where our morning resumes:

9:22 AM. The Datadog alert is still sitting in your Slack. You ask Arnab:

"What's happening with the payments service latency in the last hour? Is this related to a recent deployment?" 

Arnab pulls the latency timeline from Datadog, cross-references the deploy in ArgoCD, looks at the code merged in GitLab, surfaces a similar incident from six weeks ago in our incident history, asks Tokki to debug and tells you it looks like a connection pool misconfig. You didn't open Datadog. You didn't open GitLab. You didn't search a wiki. The integrations did.

The important part: none of that required setup on your end. Tools appear in Arnab as we ship them, available automatically.

3. Model-Agnostic: The Best AI, Everytime

The AI landscape moves fast. A model that's the clear frontrunner today may not be in a few weeks, and different tasks prefer different models. Locking a company's AI experience to a single model is a strategic bet we don't want to make.

Arnab is built on the Model Context Protocol (MCP), which decouples the intelligence from the tooling. That gives us the flexibility to swap models while maintaining the same integrations and knowledge base. Today, Arnab exposes six models that a user can pick from depending on the task — reasoning-heavy work, fast retrieval, creative drafting, code-heavy sessions — and we can expand that list as new models ship. Employees aren't waiting on us to "upgrade Arnab" when a new model lands; they're picking it from a dropdown.

4. A Sharing Ground: Skills, Apps, and Some Friendly Competition

When we removed token limits for our employees, we saw it quickly accelerated individual employees. But for those who hadn’t yet overcome the cold start, uncapping token limits did not help.

Most internal AI tools are built for consumption. Arnab also was designed for contribution and participation.Two ideas drive that concept.  

The first is Skills: reusable, multi-step workflows that any employee can build in plain language for tasks they execute over and over again. The Zoom follow-up workflow you could have run at 9:14 AM is a Skill — built by someone in Operations, published internally, available to anyone who wants it. You didn't build it. You didn't set it up. You just ran it, and three minutes later the action items were routed, the missed attendees had a summary, and the open questions were threaded with the right people tagged in.

The second is Apps: Arnab can build interactive apps on-demand, save, and host them for internal use. When a member of the Finance team needed to view vendor spend and highlight anomalies, previously we would’ve had to tie a few reports manually into a spreadsheet and contact someone from engineering to investigate the anomaly. Now they can ask Arnab. Arnab generates an app, wires in the data vendor sources, highlights anomalies and ties in operational data, logs and metrics, to explain the root cause of the anomaly. The app inherits the permissions of the user viewing the app, so authorization just works.

Both Skills and Apps live in a marketplace inside Arnab, and both track downloads, likes and uses. The leaderboard is real. So are bragging rights. It is not based on token usage, it is based on who made the best apps and skills. A culture of building for each other isn't something you can simply make happen — you have to consciously design it into the product, and then watch it take off.

5. Memory: It Learns You

The fifth is the most personal. Arnab arrives knowing the basics about you — your name, your role, your team, your team functions. From there, the more you use it, the more it picks up your preferences and adapts to your style. Which projects you're driving, who you collaborate with most, the recurring questions you tend to ask, the formats you like your answers in.

We built personal memory into Arnab, and we gave every user direct visibility into theirs and the ability to manage it. Memory shouldn't be a black box. The point isn't for Arnab to feel omniscient — it's for it to feel like it's actually paying attention.

The Platform Layer

Arnab is the control tower. It's where employees interact with AI capabilities across the company. But Arnab isn't OnePay's only AI agent. We've deployed a growing fleet of specialized agents across the business — engineering, operations, compliance, fraud, customer support, and more on the way. Each one is purpose-built for its domain. Arnab lets us reach them consistently through a single front door.

The architecture that holds this together is built on two open protocols. We made an early decision not to invent our own.

The first was MCP — the Model Context Protocol — for how agents access tools and data. Before MCP, every AI integration was a bespoke wiring job: proprietary APIs, custom auth, glue code that broke whenever something upstream changed. MCP standardizes that surface.

Then came ACP — the Agent Communication Protocol — for how agents talk to each other. ACP plays the same role for agent-to-agent collaboration that MCP plays for agent-to-tool access. When Arnab needs to access information that is available in our vast codebase, it communicates with Tokki using the ACP protocol to get the answer.

Why It Compounds

A meeting follow-up, an incident investigation, a custom dashboard — each is useful in isolation. When AI stops being something you configure and starts being something you just reach for, the learning compounds. Someone on Finance builds a Skill that someone on Support adopts. An App that started as a one-off dashboard becomes the way a whole team monitors a process. The floor lifts for everyone, and the ceiling keeps rising because everyone's contributing to it.

That's what we mean when we say AI-native at OnePay. Not that every employee is a power user — that the cold start is gone, the tools are connected, and the company gets smarter and faster at the edges every day.

The Point

The latency spike got investigated faster. The action items got routed. The dashboard got built. None of it required opening four different tools, remembering where the context lived, or finding someone on a specialized team to do it for you. None of them dropped, and what once took multiple context switches now happens in the same context: Arnab.

That isn't one dramatic transformation. It's a lot of small frictions removed, compounding across a lot of people, every day. That's what AI-native actually looks like in practice — it’s not a single moment where everything changes, but a steady reduction in the overhead that keeps work from moving as fast as it should.

Arnab is available to all OnePay employees today. More integrations, Skills, and Apps are landing each day.

If you want to help build the next generation of how organizations work with AI, we're hiring. Visit https://onepay.com/careers