AI in OnGravy

AI in OnGravy does the busywork, never the signing. Five surfaces: Ask OnGravy for questions, the AI agent for commands, voice vouchers for capture, the OCR stack for documents, and the WhatsApp assistant for doing all of it from chat. One principle runs through everything: AI drafts and suggests; consequential actions wait for a human.

Ask OnGravy — ⌘I

Press ⌘I anywhere in the app to open the chat sidebar (there is also a draggable launcher tab, and a full-page view at /dashboard/ask). It answers three kinds of questions:

  • How-tos — “how do I set up recurring invoices?”
  • Tax law — current as of July 2026, across the regimes OnGravy supports.
  • Your own books— “show expenses over ₹50,000 last quarter”, “sales vs last quarter” — real answers computed from your data, including comparisons.

Conversation history is synced, so a thread started on one device continues on another.

Ask also knows the specialised ledgers. It can explain, and route you straight to, the vertical accounting surfaces — company capital register, LLP and partnership accounts, trust/NGO, agriculture, real-estate, professional practice, import/export, e-commerce, manufacturing, and the tax record-action tools (advance-tax challans, GST TDS/TCS, insurance register, ITR-6, loan EMI). Ask “where do I book director sitting fees?” or “how do I record an APMC commission?” and it points you at the right screen. See Industry & specialised accounting.

The Virtual CA advisor feed

/dashboard/virtual-cais the proactive side of the same intelligence: a ranked “what needs you” feed that merges exception findings, compliance deadlines and a scrutiny risk band into one prioritised list, with an action link on every item and a books-closed indicator up top. It is advisory — it surfaces and links; you act. Details in Month-end & advisory.

The AI agent — commands with a visible trace

/dashboard/ai-agenttakes plain-English commands — “compile the VAT return”, “build the audit pack”, “run compliance checks” — and the agent picks the tools to do it. Every run shows a per-tool trace: you see each tool the agent invoked and what came back, not just a final answer.

Human approval — always, for high-stakes actions

Four categories of action always queue for human approval at /dashboard/agent — the agent cannot execute them directly, no matter how it was asked:

  • Filing a GST return
  • Posting a journal entry
  • Approving payroll
  • Creating a purchase

The queue shows what the agent wants to do and why; you approve or reject each item. This is the same review-first stance as OCR drafts and bank-match suggestions elsewhere in the product.

Voice vouchers

Speak a transaction and get a structured voucher draft: /dashboard/voice on the web, the mobile app, or a WhatsApp voice note. The recording is transcribed, the transaction is parsed into a draft voucher, and — as always — you review before it posts.

The OCR stack

Document extraction (the Smart inbox in Daily bookkeeping) runs on a multi-provider stack — AWS Textract, Azure, and Google Document AI — with a validation layer and mandatory human review. For India, extractions are grounded against your GSTR-2B: GST details on a scanned bill are cross-checked against what the supplier actually reported, which is where much of the accuracy comes from. Nothing OCR produces is ever auto-posted.

The WhatsApp assistant

Run the routine parts of your books without opening the app:

  • Record transactions — sales, purchases, and payments by text or voice note.
  • Snap a bill — send an invoice photo and OCR drafts it into the inbox.
  • Ask for reports — request figures and summaries in chat.
  • Manage staff join codes — onboard team members from the same thread.

The assistant is defended against prompt injection — instructions smuggled inside a forwarded message or document do not become commands. Message templates and the assistant policy are configured in Settings.

Bring your own AI provider

/dashboard/byo-llm lets each business choose the AI provider behind these features:

  • Anthropic — the default.
  • Sarvam — for India-resident AI processing.
  • OpenAI or Gemini.
  • A local endpoint — point OnGravy at AI you host yourself.

The setting is per business, so different businesses in one account can use different providers.

Gotchas & good to know

  • The agent never files, posts, approves payroll, or creates purchases on its own. If those items are not appearing in /dashboard/agent, nothing has happened yet — check the approval queue before assuming the command failed.
  • Ask answers about your books are computed from your books. If the books are behind (unposted drafts, unimported statements), the answers will faithfully reflect that.
  • Tax-law answers are current as of July 2026. For the authoritative record of regulatory changes shipped in the product, see the compliance changelog.
  • Voice and OCR produce drafts. Review is part of the flow, not an optional extra — accuracy comes from extraction plus your confirmation.
  • Choosing a BYO provider applies per business — switching providers on one business does not affect your others.
  • GSTR-2B grounding is India-only; in other jurisdictions OCR still runs multi-provider extraction and validation, without the 2B cross-check.
← Previous
🏦 Banking & reconciliation
Next →
🔒 Month-end & advisory