Importing broker and CAS PDFs: why AI helps but still needs you

"Connect my account" in a product pitch is a trust and regulation story. "Import this PDF" is a data accuracy and time saving story. Serious fintech has to be good at the second before it promises a nationwide read-only API everywhere. This post is about uploads—contract notes, CAS PDFs, and broker exports—and about treating models as proposals, not oracles.

Two very different product promises

Account linking with a bank or broker, where you hand over credentials to a read-only pipe, is a large security and contract program: consent, revocation, incident response, and often regulatory questions that move on their own clock. A file you export yourself and upload into Artha has a smaller legal surface for most households. It is the same PDF a registrar or broker would email you; you are not inventing a new source of truth, only changing where you paste it for analysis.

Table extraction from a PDF is a hard problem when tables break across pages, or when a bonus or split line sits in a sub-row that a one-shot OCR pass treats as a smudge. That is why well-designed software separates suggested cells from locked rows with a review step, even when a model is strong enough to guess most of the grid the first time.

What a contract note and a CAS are doing for you

A contract note in Indian retail is a settlement list: your order, your fill, price, quantity, and the charges the broker was allowed to print on that day. You are not looking for a philosophy of markets; you want one clean row per fill in your stock portfolio tracker so average cost and P and L can stay consistent. Your own broker’s help site (for example the current pages at Zerodha or Groww) is the right vocabulary for what a field is called, not a third-party gloss.

A consolidated account statement from a depository (see NSDL and CDSL for what a CAS is for) is a list of ISINs and units. You want that list to line up with your mutual fund tracker and then to sit in the full household portfolio beside NPS, EPF, and listed names. The software can suggest table rows; you still have to be the person who says the row is yours for the year you are closing.

Failure modes to expect

• a photo of a glossy printout and the OCR drops a digit;
• a table that continues on the next page and the parser duplicates a line;
• a name on a note that does not match a joint holder;
• a dividend that lands in cash in the broker while you were still thinking in shares only on the line above.

Artha is built as a propose-then-confirm loop. You upload a PDF or CSV as the Artha page and features surface describe, you look at a grid, you fix a cell, and only then the portfolio and the stock and US sleeves are allowed to move together. That is a different feeling from pasting a password you cannot rotate into a new box.

Security in one line: a file you downloaded while you were already logged in to the bank or broker is a different trust line from handing a third party your live credentials. Read the Invesh privacy story with the same care you use for a bank; do not use any wealth product, ours or not, to avoid reading what you are signing up for. We are not your CA; we are a ledger and a set of suggested rows after you say an import looks good.

Importing with Artha is the boring part of a Sunday when it works and the loud part when a single cell is wrong. Fix the cell, then return to the market questions you actually care about. The Nifty will still be there on Monday; a false lot line should not live past the review step you just promised yourself you would take.

If the same long weekend is when you update mutual funds and stocks, run both through one reconciliation pass so the portfolio view is not half-migrated. A book that is two days behind on NAV in every line is more honest for tax prep than a book with one perfect sleeve and one you forgot to import.