Most AI tools ask you to make an uncomfortable trade: useful answers in exchange for sending your data somewhere you cannot see. For a CA practice — holding PAN, Aadhaar, bank details and entire sets of books — that trade is hard to justify. Lipik is the assistant we have built into FirmsNest to avoid it. A munshi in the old sense: it reads the papers, finds the figures and prepares drafts, while you review, decide and sign.
It lives inside your firm’s admin workspace, works only on the documents your firm already holds, and is available to your team — never to your clients. It does not file anything, and it is not a tax, audit or legal adviser. It prepares; you remain the professional behind the work.
What Lipik does
- Answers questions from your own documents. Ask about a client’s file in plain language and Lipik answers from what your firm has uploaded — and cites the documents it used, so you can check it in seconds rather than take it on faith.
- Pulls the figures out for you. Point it at a Form 16, a 26AS or a set of accounts and it extracts the values into a structured form, so you are reviewing numbers instead of re-typing them.
- Drafts a working paper. It can assemble a draft tax-estimate worksheet in Excel — clearly labelled a draft for your review, not a filing. The arithmetic is done in plain code, not guessed by the model, so the sums add up the same way every time.
- Checks the latest rules. For a current due date or a rate change, Lipik can search the public web — using only a generic question, never a client’s name, PAN or any figure from a document — and shows you the sources.
What stays your job
Lipik does not compute or file returns. You continue to do that wherever you do today — Winman, Computax, KDK, the income-tax or GST portal. Every figure it surfaces is a draft for you to verify; every worksheet carries that label. The point is not to replace your judgement but to remove the hours of reading, extracting and re-keying that sit in front of it.
How the data boundary works
The reason we could build this at all is the boundary underneath it. The rule we held ourselves to is simple: your clients’ data never leaves the system. In practice that means four things.
The model runs in India, on Sarvam
The AI model that reads your documents runs on Sarvam — sovereign AI built and hosted in India — under terms that keep the data in-country. Your clients’ documents are not shipped to a server abroad to be processed.
Your data is never used for training
The model is not improved, tuned or trained on your firm’s documents. What you put in is used to answer your question and nothing else — no retention for model-building, no quiet repurposing later.
Documents and their index stay in India
To search your documents, Lipik builds a private index of them. That index, and the documents themselves, sit in encrypted storage in an India region, sealed to your firm alone — one firm’s data is never visible to another.
Web search carries no client data
The one time anything leaves the country is a web search for a public rule — and a guard strips out anything that looks like client data (a PAN, an Aadhaar, an email, a phone number) before the query goes out. Only a generic, public question is ever sent.
A scribe, not an oracle
We named it Lipik — the scribe — on purpose. A scribe prepares the record; the principal reads it and signs. That is exactly the relationship we want you to have with it. It drafts, extracts and summarises with its sources attached; you bring the professional judgement that no model has. When a number matters, the computation is done in code you can rely on, not produced by a language model improvising arithmetic.
Where it fits, and what’s next
Lipik appears in your firm’s admin workspace once it is switched on for your account. We are rolling it out carefully, firm by firm, rather than flipping it on for everyone at once — the kind of data it touches is not something to be casual about. If you would like your practice in an early group, tell us.
The promise of AI for a practice was always obvious; the catch was the data. Lipik is our attempt to keep the first without ever conceding the second.