Arbiton

01 · The intelligence layer for next-generation banking

The intelligence layer for next-generation banking

Every customer, answered.Accurately. Privately. For pennies.

Try it yourself Talk to us
Demo Bank Ltd.Asha · 4521

Hi Asha. Ask me anything about your spending, income or balance.

How much did I spend on coffee in September?

₹1,555.00 across 4 purchasesStarbucks₹755.00Blue Tokai₹800.00Approved plan · 0 model tokens
Recorded demo · Asha’s account

Behind the screen

  1. IUnderstandSystem 1Coffee · September · Starbucks, Blue Tokai
  2. IIPlanSystem 2An approved plan already exists
  3. IIICalculateClosed environmentClosed environment · Asha’s accounts only
  4. IVAnswerCheckedChecked · 0 model tokens

Arbiton sits between your customers and your bank. It starts with the questions customers ask most, about their own spending, and answers them correctly, inside your walls, on models you control. The same engine grows into investment guidance and any banking question.

Not an off-the-shelf bot. Every installation is built around your bank’s own data.

Your customers already ask.
Nobody answers well.

Where did my money go this month? Can I afford to start investing? How do I raise my UPI limit? Customers want answers about their money and their bank, in their own words. The tools they get today leave them to work it out, or to wait for a call centre.

A guessed number is worse than no number at all.

Carried over from today, set against what changes
Dr. What customers get todayCr. What Arbiton gives them
Statement PDFs and filter screens. The customer does the arithmetic.A plain question and a plain answer, with the exact figures.
General chatbots that guess numbers and sound sure about them.Every figure is computed from the ledger and checked before it is shown.
Account data sent to outside AI services.Runs inside the bank, on models the bank chooses.
Product and policy questions wait in a call-centre queue.Answers from the bank’s own products and policies, as the bank approves them.
A model bill on every message, rising with every customer.Repeat questions are answered without calling a model at all.

A transformer layer between your customers and your bank.

Arbiton turns a customer's question into a correct, checked answer, built from that customer's own data and your bank's own rules. It is built to the standard a bank needs: secure, compliant, cheap to run, and yours to operate.

Runs inside your walls. Your data does not leave.

One engine, many use cases

We start with one use case and do it properly. The same engine, the same security and the same way of checking answers carry over to the next.

  1. In the demo

    Transaction analysis

    Spending, income, merchants, trends and balances, computed exactly from the customer’s own ledger.

    “How much did I spend on coffee in September?”

  2. Same engine

    Investment guidance

    Guidance built on the customer’s real income and spending, inside the products and suitability rules your bank sets.

    “How much could I put into a monthly SIP?”

  3. Same engine

    Banking questions

    Products, rates, limits, charges and how-to questions, answered from your approved documents.

    “How do I raise my UPI limit?”

  4. Built with you

    Your next use case

    Whatever your customers ask most, added with the same data preparation, testing and approval.

    “Tell us what your customers ask.”

Four steps.
Only one of them thinks hard.

Fast, specialised models handle the routine. A deliberate model is called only when something is genuinely new. Everything that touches money is plain, checked computation.

Most questions skip step II entirely and cost no model tokens.

  1. I

    Understand

    System 1

    A small, fast model running inside your environment reads the question: what is asked, for which period, which merchants or categories, and whether it should be answered at all. A vague question gets a clarifying question back.

    Model useSmall local model
  2. II

    Plan

    System 2

    Only for a kind of question the system has not seen before. A larger model drafts a plan for the answer. Once a person approves it, the plan is reused for every customer who asks that kind of question.

    Model useNew questions only
  3. III

    Calculate

    Closed environment

    The plan runs in a sealed environment holding this customer’s data and nothing else: no network, no other accounts.

    Model useNone
  4. IV

    Answer

    Checked

    The result is checked, then written in plain language with the exact figures, ready for your channel.

    Model useNone
Day book · demo accountevery figure computed, none guessed
  1. 001How much did I spend on coffee in September?₹1,555.00Starbucks ₹755.00 · Blue Tokai ₹800.00Approved plan0
  2. 002How many Uber trips did I take in September?2 trips₹581.25 in totalApproved plan0
  3. 003Break down my food delivery spending in September by merchant₹1,636.00Swiggy ₹1,098.00 · Zomato ₹538.00New plan → approvedfirst ask only
  4. 004What did all customers spend in September?DeclinedI can only look at your own accounts.Declined0
Model tokens spent on repeat questions0

Not a product you switch on.
A system built around your data.

Every bank’s data is different: its core systems, how it prints a purchase, the merchants and languages of its customers, its products and its rules. Good answers depend on understanding that data, so every Arbiton implementation is shaped to the bank that runs it.

Same engine. Your data. Your answers.

One purchase, as three banks print it
  • UPI/SWIGGY/swiggy@icici/Food order
  • POS 4512XXXX SWIGGY BLR IN
  • SWIGGY*ORDER 8812 BANGALORE
SwiggyMerchant · Food deliveryLearned from each bank’s own records, not guessed
What is standard in every installation, and what is built for each bank
Standard in every installationBuilt for your bank
  • The four-step engine: understand, plan, calculate, answer
  • The security model: scoped access, closed environment, audit trail
  • The operator dashboard
  • Evaluation, release and rollback tooling
  • Support for open-source and approved models
  • Which use case you start with, and which come next
  • Adapters to your core banking, data warehouse and identity systems
  • Your merchant, category and counterparty catalogue, built from your own transactions
  • System 1 trained on your customers’ language, products and terms
  • Your products, rates and policies, and the rules for investment guidance
  • The questions you choose to answer, and the ones you decline
  • Evaluation sets and an accuracy bar agreed with your team
  • Deployment in your environment, under your compliance rules

System 1, System 2, and the work in between.

Arbiton is not a model wrapped in an app. It pairs fast, specialised models for the routine with a deliberate model for the new, and adds the groundwork that makes both reliable. That groundwork is why answers stay with the facts.

01

Prepare

Your transaction data is cleaned and enriched before any question arrives: merchants named, categories assigned, people and transfers recognised. Good answers start with a well-kept ledger.

02

Train

The System 1 models are trained on the way your customers actually ask, in their language and with your products, merchants and terms.

03

Evaluate

Every change is tested against agreed question sets and replays before a customer sees it. Nothing ships on a hunch.

Less invention, by design: the model plans, the ledger supplies the numbers.

Notes to the product.

Your team runs Arbiton from a dashboard. Each part of it is explained here the way an annual report explains its figures: briefly, and in plain words.

Every behaviour can be changed from the dashboard, without a deployment.

  1. Note 1

    Question understanding

    System 1 decides what each question is about, which period and merchants it means, and whether it should be answered.

  2. Note 2

    Entities

    The merchants, people, categories and places customers mention, matched to the right records, including nicknames and misspellings.

  3. Note 3

    Intents

    The kinds of question you support, with examples, and the kinds you decline, such as payments or another person’s data. New use cases arrive as new intents.

  4. Note 4

    Features

    The facts an analysis may use, such as amounts, dates, categories and balances, defined once and shared by every answer.

  5. Note 5

    Skills

    Reusable calculations: totals, comparisons, trends, top merchants, recurring payments.

  6. Note 6

    Approved answers

    Plans learned from new questions, reviewed by a person, then reused for every customer at no model cost. A cache builds itself as customers ask.

  7. Note 7

    Pinned caches

    Common figures, such as monthly spend by category, prepared ahead of time so popular questions answer instantly.

  8. Note 8

    Evaluations and testing

    Question sets, replays and adversarial tests that every change must pass before it reaches customers.

  9. Note 9

    Releases and rollback

    Configuration is published in versions with checks before each release. Roll back in one step.

  10. Note 10

    Traces and audit

    Every step of every answer can be followed, with personal details redacted. Every change records who made it.

  11. Note 11

    Model choice

    Choose which model does each job, in-house or from an approved provider, and set daily spend caps.

  12. Note 12

    Statement upload

    Customers can add a PDF statement from another bank and ask about it in the same conversation.

You own the code.
We sit with your team.

Arbiton is not software as a service. You license the source code and run it in your own environment. Our engineers work alongside yours, forward deployed, until it fits your systems, your data and your customers.

No black box. The code is yours to read.

Schedule of works

  1. I

    Discovery

    Your data sources, identity setup, compliance requirements, the questions your customers actually ask, and which use case to start with.

  2. II

    Integration

    Adapters to your systems, access rules, and the models you choose, installed in your environment.

  3. III

    Preparation and training

    Your transaction data enriched; System 1 trained on your customers’ language and products.

  4. IV

    Evaluation

    We agree the question sets and the accuracy bar together. Nothing goes live until it passes.

  5. V

    Launch and handover

    Go live in your channels, train your team to run it, then add the next use case on the same engine.

Ask Asha’s account anything.

Asha is a fictional customer with a real-looking September statement. The demo covers the first use case, transaction analysis: pick a suggested question or type your own, and watch each step of the answer alongside.

Three visitors at a time. If it is busy, ask anyway: your question joins the queue.

Demo Bank Ltd.Asha · 4521

Tap a question or type your own to start.

Hi Asha. Ask me anything about your spending, income or transfers in September.

What is happening now

Ask a question to see each step of its answer.

  1. IUnderstand
  2. IIPlan
  3. IIICalculate
  4. IVAnswer

Asha’s statement · 1 to 22 September 2026

XXXXXXXX4521
DateNarrationDebitCreditBalanceAmount
01 SepOpening balance₹48,250.00₹48,250.00
01 SepNEFT-ACME TECHNOLOGIES PVT LTD-SALARY SEP₹85,000.00₹1,33,250.00+₹85,000.00
02 SepUPI/SWIGGY/Food order₹486.00₹1,32,764.00−₹486.00
03 SepPOS DECATHLON SPORTS BANGALORE₹3,499.00₹1,29,265.00−₹3,499.00
04 SepUPI/STARBUCKS COFFEE₹345.00₹1,28,920.00−₹345.00
05 SepUPI/UBER INDIA/Trip₹268.50₹1,28,651.50−₹268.50
06 SepUPI/BLUE TOKAI COFFEE ROASTERS₹420.00₹1,28,231.50−₹420.00
07 SepIMPS-RAHUL VERMA-Dinner share₹1,200.00₹1,27,031.50−₹1,200.00
08 SepPOS PVR CINEMAS FORUM MALL₹980.00₹1,26,051.50−₹980.00
10 SepUPI/AMAZON PAY/Order 403-112₹2,349.00₹1,23,702.50−₹2,349.00
11 SepUPI/SWIGGY/Food order₹612.00₹1,23,090.50−₹612.00
12 SepPOS DECATHLON SPORTS BANGALORE₹1,299.00₹1,21,791.50−₹1,299.00
14 SepUPI/BLUE TOKAI COFFEE ROASTERS₹380.00₹1,21,411.50−₹380.00
15 SepUPI/UBER INDIA/Trip₹312.75₹1,21,098.75−₹312.75
16 SepIMPS-PRIYA MENON-Rent share₹6,500.00₹1,27,598.75+₹6,500.00
18 SepPOS PVR CINEMAS FORUM MALL₹760.00₹1,26,838.75−₹760.00
19 SepUPI/STARBUCKS COFFEE₹410.00₹1,26,428.75−₹410.00
20 SepUPI/ZOMATO/Food order₹538.00₹1,25,890.75−₹538.00
21 SepPOS CROMA ELECTRONICS INDIRANAGAR₹4,999.00₹1,20,891.75−₹4,999.00
22 SepINT.PD:01-06-2026 TO 31-08-2026₹412.00₹1,21,303.75+₹412.00
Closing balance₹1,21,303.75

Open an account with us.

Tell us about your bank and what your customers ask. An engineer replies, and we can walk through the demo using your own questions.

Talk to engineers, not a sales script.

Enquiry slipNo. 000001

We use these details only to reply to you.