Building XwFin: What AI for GST and Indian Tax Compliance Actually Takes

Xwits has an accountant. She is excellent. She was also spending a significant portion of her working week on GST-related administration — generating invoices, preparing GSTR-1 filings, reconciling purchase records, chasing payment records from multiple sources. The work was accurate but mechanical, and it was eating time that should have gone to more complex financial analysis. XwFin started as our attempt to automate the mechanical parts. The first version was built in six weeks for internal use. It took another eight months to turn it into something we could offer to other businesses.

Why GST Compliance Needs AI, Not Just Software

Standard invoicing software handles the straightforward case: one GST rate, clean supplier data, standard invoice format. Indian GST compliance is not the straightforward case. The structure involves 5%, 12%, 18%, and 28% tax slabs, with exemptions, reverse charge mechanisms, and composite schemes layered on top. HSN codes that determine which slab applies number in the hundreds of thousands. E-way bills add another compliance layer for goods movement. GSTR-1, GSTR-3B, and annual returns have different formats, frequencies, and reconciliation requirements.

Rule-based software handles static mappings well but breaks on ambiguity. What HSN code applies to a product that could reasonably fall into two categories? What GST treatment applies to a mixed supply? AI handles these judgment calls better because it can reason across context rather than look up a fixed rule. Our AI layer does not replace the rules — it applies them to situations the rules did not explicitly anticipate.

The Data Challenges Nobody Warned Us About

Indian business data is messy in specific ways that are different from the messy data problems described in most ML literature. Supplier names are inconsistent — the same supplier might appear as "Mehta Traders", "S. Mehta Trading Co.", and "Mehta Trading" in different invoices. PAN and GSTIN numbers, which should be the canonical identifiers, are sometimes entered incorrectly or missing entirely. Bank statement formats from 30+ Indian banks are each subtly different. Our OCR and entity resolution layer had to be trained on this specific messiness, not generic document understanding.

The second data challenge was building training data for Indian tax edge cases. No public dataset covers the full range of GST compliance scenarios a real business encounters. We built our training corpus from anonymised transaction data, synthetic edge case generation, and expert annotation by CAs. This took longer than building the model itself.

Lessons from Building Financial AI in India

Three things I would tell anyone starting a fintech AI product in India today: First, accuracy requirements are non-negotiable. A 95% accurate model sounds impressive until it means one in twenty GST filings has an error. Tax accuracy needs to be 99.5% or higher, which means your AI needs human review workflows for edge cases, not just a threshold for escalation. Second, trust is earned through transparency. Our users want to see why the system classified something a certain way. A black-box recommendation gets ignored. An explanation that references the relevant GST circular gets accepted. Build explainability from the start. Third, the CA (Chartered Accountant) is your distribution partner, not your competitor. We almost made the mistake of positioning XwFin as a CA replacement. The CAs who use it now bring it to their clients and manage 10-15x more compliance work with the same team. The AI is a leverage tool for professionals, not a replacement for them.