Compounding Intelligence for Indian SMBs
Most business software delivers the same value on day 300 as it does on day 1. A spreadsheet doesn't learn your vendors. Tally doesn't notice that a customer's orders have been declining for three months. These tools store what you put in and give it back when you ask. The value proposition is organization, not understanding.
The best businesses don't run this way. They run on accumulated understanding.
The long-tenured employee
Every business has that one person who just knows things. They know that Gupta Trading always delays payments in Q3 but makes up for it with large orders in Q4. They know that a ₹40,000 order from Verma is unusual and worth double-checking. They know that copper prices spike before Diwali, and that this affects margins on a specific product line. They know that when the owner says "Sharma ji" without further context, he means Sharma Trading in Jaipur, not Sharma & Sons in Delhi.
This knowledge wasn't taught. It was accumulated — through hundreds of transactions, thousands of conversations, years of daily exposure. It's not in any manual or database. It lives in someone's head: a rich, continuously updated model of the business and its relationships.
And the raw material of that model isn't structured data in an ERP. It's WhatsApp messages — "sharma ji se 25000 ka payment aaya" — forwarded invoice photos, voice notes in a mix of Hindi and English sent between a warehouse and a delivery truck. That's the actual information flow of an Indian small business.
The question we found ourselves asking: what if the system itself could develop this kind of understanding?
What compounding intelligence actually means
A system whose understanding of a business deepens with every interaction creates value that grows non-linearly over time.
Week 1: The system records transactions. "Sale to Sharma Trading, ₹50,000, 18% GST." It's a ledger — useful, but not much more than a spreadsheet.
Week 4: The system catches errors. "You logged ₹25,000 for Sharma but their orders are usually around ₹2,50,000 — did you miss a zero?" It has enough history to know what's normal for each party.
Week 12: The system predicts. "Based on your receivables and payment patterns, you'll be short ₹2 lakh in three weeks. But if you collect the ₹80,000 overdue from Verma, you're covered." It can model cash flow because it knows not just what's owed, but how each party actually pays.
Week 30: The system surfaces what you wouldn't think to look for. "Sharma's order frequency has dropped from weekly to monthly since October. Their payment times have also lengthened. Something may have changed in their business." It's connecting signals across multiple dimensions — volume, frequency, payment behavior — to surface a relationship risk the owner hasn't noticed yet.
Week 50: The system acts with earned autonomy. Routine invoices are generated. Payment reminders are sent at the right time, in the right tone, for each party. GST returns are prepared and filed on schedule. The owner reviews exceptions, not operations.
Each week builds on everything before it. And critically, this accumulated understanding becomes a switching cost — not through lock-in mechanics, but because the contextual intelligence that took weeks to build is genuinely irreplaceable. You could export your transactions. You can't export everything the system has learned about your business.
Two things changed
This kind of system wasn't buildable two years ago. Two prerequisites converged.
First, AI agents crossed the capability threshold for messy, real-world business input. The typical Indian SMB owner doesn't sit at a desk entering data into forms. They send WhatsApp messages, forward invoice photos, speak voice notes in Hinglish while walking through a warehouse. Until recently, no AI could reliably parse this into structured data. Now it can — consistently, with context, across languages.
Second, India's digital infrastructure created a rich substrate that an intelligent system can orchestrate. UPI processed 228 billion transactions in 2025. The Account Aggregator framework has 269 million customer consents and 780+ financial institutions. GSTN has 13.8 million taxpayers filing digital returns. E-invoicing thresholds are dropping. DigiLocker and EntityLocker offer verified business document APIs.
These rails are extraordinary. But they were built for systems, not for people. A small business owner doesn't interact with "the Account Aggregator framework." They say "mera bank statement check karo." The rails are waiting for an interface layer — something that translates between how businesses actually operate and what these systems require. An intelligent agent that accumulates understanding of the business is uniquely suited to be that layer, because it can orchestrate these systems on behalf of the business with full context about what matters and why.
Why this isn't just "add AI to accounting software"
Most of what's happening in this space right now is AI-augmented, not AI-native. The distinction matters.
An AI-augmented system takes an existing interface — forms, menus, dashboards — and adds intelligence to it. AI suggests which ledger to use. AI auto-fills a tax return. This is useful. It makes existing workflows faster. But the AI operates within the constraints of the original design. It's intelligence layered on top of a static architecture.
An AI-native system is designed around intelligence from the ground up. There are no forms to fill because the system extracts structure from natural conversation. There's no dashboard to check because the system surfaces what matters proactively. There's no learning curve because the interface is the same one the owner already uses — language.
More importantly, an AI-native system can compound. It maintains a persistent, structured model of the business that every interaction enriches. Party behavioral profiles. Product margin trends. Cash flow predictions weighted by actual payment reliability. Anomaly detection calibrated to what's normal for this specific business. None of this is possible to retrofit onto a system designed around structured input forms, because the data model, the interaction model, and the intelligence model are all intertwined.
Progressive autonomy
In practice, autonomy needs to be earned, not assumed. We think about this as a ladder:
L0 — Record. "Log ₹50K sale to Sharma." → Done.
L1 — Suggest. "Sharma is 15 days overdue. Send a reminder?"
L2 — Act with approval. "Your GSTR-1 is ready. Shall I file it?" → Yes. → Filed.
L3 — Act with notification. "I sent Sharma a payment reminder. He's now 30 days overdue."
L4 — Act autonomously. Recurring invoices generated, payments reconciled overnight, GST filed on schedule. Human oversight for exceptions only.
Each level requires two things: accumulated intelligence (the system needs to know enough to act correctly) and demonstrated trust (the system needs to have proven itself at the previous level). An agent that earns autonomy through accuracy, over time, with the owner always in control — "always auto-send payment reminders, but always ask before filing GST" — is fundamentally different from one that claims autonomy by default.
The 63 million
India has 63 million MSMEs. Most manage their finances in notebooks, WhatsApp chats, or basic spreadsheets. They aren't going to adopt traditional accounting software — that behavioral shift has been attempted for twenty years, by well-funded companies, and it hasn't happened. The gap isn't awareness or access. It's that the software demands something the owner can't give: structured input, consistent usage, and time to learn a new tool.
But the same owners who won't open an accounting application will happily send a WhatsApp message — because they already do, dozens of times a day, for business. The same owners who won't read a dashboard will act on a notification. The same owners who won't generate a report will ask a question — "Sharma ji ka kitna baaki hai?"
The opportunity isn't to simplify existing software. It's to build something that doesn't feel like software at all — an agent that lives where the business already operates, understands the owner's language and context, and builds a deepening model of the business from every interaction. Not a tool the owner operates, but a system that operates on behalf of the owner, getting better at it every day.
That's what we're building at Kaarta. We'll write more about the specific ideas, architecture, and challenges as we go.
Kaarta is building AI agents that run back-office operations for Indian SMBs. If this resonates, reach out at hello@kaarta.in.