Ask your business anything
A conversational copilot grounded in your real numbers — answers, not hallucinations.
You have ₹2.15L collectible now across 3 overdue accounts:
- Sharma Traders · 18d₹1,24,500
- Verma & Co · 9d₹62,300
- Patel Retail · 4d₹28,200
Want me to draft WhatsApp reminders for all three?
Live product UI — sample data.
Plain-language answers
“How much GST do I owe?”, “Who owes me money?”, “Why did profit fall?” — answered from your books.
Grounded, never invented
It reasons over a compact snapshot of your real data and says so when the answer isn’t there.
Enterprise, key-safe
Powered by Claude/OpenRouter/OpenAI/Gemini via a metered gateway; keys never touch your machine, and Tier-A insights still run offline.
What this actually means in practice
Ask a question and the copilot never computes the answer. A router picks the deterministic engines the question needs, those engines read your posted books and return a structured snapshot — this month's sales, gross and net profit, GST liability, bank balance, the health score and its recommendations, who owes what — and only then is the language model given that snapshot and asked to write a reply. Every rupee in the sentence was already a number before any model saw it.
Switch the language model off entirely and the figures do not change, because an on-device narrator writes the same reply from the same snapshot. That is the test worth applying to any AI accounting claim: if turning the model off changes the number, the number was never yours. Here it cannot — the prompt builder hands the model only the structured grounding and instructs it to invent nothing, and where a figure is absent the reply says so instead of filling the gap.
The copilot can also propose an action — create a follow-up task, schedule a renewal reminder, raise an invoice against a contract milestone. Each is a named tool bound to an existing audited mutation, argument-validated before anything runs, and every tool that writes carries an approval requirement; only read-like refreshes go through unattended. Nothing bypasses the ordinary permissions your user account already has. It drafts, you approve, and the write is logged like any other.
| You ask | Which engine answers | What the model contributes |
|---|---|---|
| How much GST do I owe? | Output less input tax over the posted period | Wording only — the figure is the ledger's |
| Who owes me money? | Receivables ageing by party and due date | Ordering it into a readable list |
| Why did profit fall? | Period-on-period P&L comparison from the same books | A plain-language account of the movement |
| Which customers are drifting? | Churn arithmetic on each party's own ordering interval | Naming them in a sentence |
| What should I do this week? | Rule-derived action list with the triggering amounts | Prioritised phrasing, not new advice |
| What is my stock worth? | Closing valuation at your chosen method | Restating the total in context |
| Anything outside the snapshot | Nothing — that data was never fetched | It says so rather than guessing |
What’s included
- Natural-language Q&A
- GST, collections & cash answers
- Grounded business snapshot
- Multi-provider gateway
- Token metering & quotas
- Deterministic offline fallback
What it does not do
- · The conversational copilot is Enterprise at ₹2,499 monthly. The deterministic engines it quotes ship in every plan from ₹249 and are readable on their own screens with no AI at all.
- · It advises and drafts; it does not post. Every write-capable action requires your approval and runs through the same permissions and audit trail as manual work.
- · No external knowledge. It cannot see your industry, your competitors, market rates or the season ahead — only what is in your own books.
- · It is only as good as what you have entered. Expenses booked outside LekhaPro or stock never recorded are invisible to it, and it will not warn you they are missing.
Frequently asked
Does it make things up?
It answers only from the business data provided and flags when something isn’t available — no invented figures.
Is my data sent to the cloud?
Only a compact snapshot, only when you ask, and only on Enterprise. Keys live on our gateway, never on your device.
What does the copilot send to the cloud?
A compact structured snapshot, only at the moment you ask, and only on Enterprise. Provider keys live on our metered gateway rather than on your machine, and usage counts against your quota. The on-device path avoids the round trip entirely and answers from the same snapshot, which is what runs when the gateway is unreachable.
Will two people asking the same question get the same answer?
The figures will be identical, since both replies are built from the same snapshot of the same posted books. The sentences may be worded differently, which is what a language model does. If exact reproducibility matters — for a review pack or a board paper — read the underlying engine screen, where the output is byte-for-byte the same each time.
What exactly is included in AI · Enterprise?
natural-language Q&A, GST, collections & cash answers, grounded business snapshot, multi-provider gateway, token metering & quotas and deterministic offline fallback. Every item on that list is a working screen you can open in the trial rather than a capability described in the abstract.
How is ai · enterprise handled differently here?
It rests on plain-language answers, grounded, never invented and enterprise, key-safe. The distinction that matters is that these are structural rather than cosmetic — the behaviour is built into how records are posted, not layered on as a report you have to remember to run.
Does AI · Enterprise work offline?
On the Desktop edition, yes — completely. natural-language Q&A and GST, collections & cash answers run against a local database on your own machine, so the screens behave the same with the network unplugged as with it connected. AI and billing are the only two things that reach out. On the Cloud edition ai · enterprise runs in the browser and needs a connection.
Does AI · Enterprise post to the real books?
Yes, and that is the whole reason it lives in this product rather than beside it. What plain-language answers records lands in the same posted double-entry ledger the Trial Balance, P&L and GST returns are built from — so there is no second set of ai · enterprise numbers to reconcile against the first.
How does AI · Enterprise affect my GST?
It feeds it directly rather than sitting alongside it. CGST/SGST and IGST are split from HSN or SAC and place of supply as each document is raised, so whatever ai · enterprise produces is already correct when GSTR-1, 3B and 2B reconciliation are drawn off the same books. The return matches the accounts because it was never a separate exercise.
Does AI · Enterprise cost extra?
No. AI · Enterprise is part of the accounting core rather than a paid add-on, so grounded, never invented is there on the entry tier exactly as it is on the highest one. What the tiers change is reach and depth — devices, multi-branch and multi-warehouse reporting, and whether the conversational AI copilots are switched on.
Answers from your books — not guesses
Every answer is grounded in a snapshot of your own data, through a gateway where your keys never live. It can read your documents, too.
Ask anything, grounded in your numbers
- You askin plain language
- Your data snapshotassembled from your books
- Key-safe gatewaykeys never leave the server
- Grounded answercites your real figures
Illustrative. The copilot answers only from the data provided — it never invents figures.
Scan a bill, book the entry
- Snap the billphoto or PDF
- Fields extractedparty · GSTIN · items · tax
- Validatedstrict, checked
- Posted to ledgerbalanced Dr/Cr voucher
Illustrative. OCR extraction is an Enterprise feature; the deterministic core still runs offline.
Ask your business anything
One bar for your whole business. Ask in plain language and get answers grounded in your real numbers — never invented. The deterministic core still runs offline.
Conversational copilot is an Enterprise add-on (Claude / OpenRouter / OpenAI / Gemini, metered). Sample prompts.