Somebody shows you a demo. They photograph an invoice on a table — folded once, slightly rotated, printed by a dot-matrix machine that last had a fresh ribbon in 2019 — and four seconds later the supplier name, invoice number, date, taxable value and tax are sitting in neat fields.
It works. It is genuinely good, and it is worth saying so plainly rather than being sniffy about it. Reading text off a page is a problem that has had a very large amount of effort thrown at it and it now mostly works.
Then you take that capability back to a practice and discover it has solved a step, not a job.
Five things the scan does not tell you
Look at what you actually have after those four seconds: a supplier name, a number, a date, an amount. Now try to file it.
Whose is it? The invoice names a supplier and a buyer. The buyer is a GSTIN belonging to a group that has four entities, two of which you handle and one of which the client's cousin handles from another city. The scan read the GSTIN correctly. It has no idea which of your files this belongs in.
Which period? The invoice is dated 28 March. It reached your office on 6 April, in a batch the client sent while clearing their desk. The date on the page is not the answer to the question you actually have.
Is it already recorded? The same invoice arrives twice as a matter of routine: once as a photograph from the client at the time of purchase, once in the month-end bundle from their accountant, often a third time when you ask for a legible copy. Three files, two of them slightly different photographs of the same paper. Only one of them should become an entry.
What was it sent for? This is the one that costs the most later. The invoice came with a sentence — "iska credit nahi mila tha" — and that sentence is the entire reason the document was sent. Extraction pulls the numbers off the page and leaves the sentence behind in the chat.
Is this the final copy? Revised invoices, cancelled ones, a proforma that looks identical apart from one word near the top. The scan reads all of them equally well.
None of these are reading problems. They are record problems, and no amount of improvement in reading the page will touch them.
The question that matters more than accuracy
Set the filing aside. There is a harder issue, and it is the one that should decide whether you let any of this near client files.
It is the second week of September. A notice has come. You need to show where a figure came from — not approximately, not "it was in the purchases", but this number, from this document, received on this date.
If the figure was typed by your article assistant from a paper on the desk, you know how to trace it. If the figure was produced by a tool that read a photograph, you need the tool to hand you the photograph back.
This is the whole game, and it is where most demos go quiet. A number displayed without a path back to the page it came from is not a time-saving. It is a liability you have accepted on the client's behalf without telling them.
So the rule, and it applies to every tool of this kind:
Any figure a machine shows you should arrive with the document it came from — the actual page, on the same screen, not a filename and not a reference to a folder.
A tool that cannot do this is not reading your documents. It is producing plausible text about them, which is a different activity that happens to look identical when it is right.
What to demand before you upload a single client file
Five questions. They are short, they are answerable, and a vendor who cannot answer them plainly has told you something.
- Show me the source. Click any number you have produced and show me the
- What do you do when you are unsure? The correct behaviour on a smudged
- Do you keep the message the document came with? If the sentence that
- Is my clients' data used to train your model? In Kognora the answer is
- Can this send anything to a client by itself? If yes, walk away. A
page it came from. If that takes more than one click, or produces a folder instead of a document, the answer is no.
figure is to say so and ask. A tool that always returns a confident value has chosen to guess on your behalf, and you will not know which values were guesses.
explained the file is discarded at the moment of upload, it is gone, and it was the most valuable part.
no. Ask anyone else the same question and get it in writing.
professional practice cannot have a system that communicates with clients without a human approving the words.
The scan is a convenience. The record is the asset.
Here is the part worth sitting with. If every scanning tool in the market disappeared tomorrow, a practice with a proper client record would lose some typing time. A practice with excellent extraction and no record would lose nothing it could name, because it never had anything except faster typing in the first place.
Extraction compounds nothing. It saves the same few minutes today that it saved last year, on each document, forever. A record compounds: the second year of it is worth more than the first, because the questions you can answer from it are questions about history — what did we tell this client in March, which documents came in that quarter, have we asked about this invoice before.
That is the asset. The scan is a feature attached to it, and any conversation that starts with the scan has started at the least interesting end.
Kognora gives a professional practice one record per client — messages, documents and meetings together — and answers questions from that record with the source document attached, so every figure has a page behind it that you can open. Client data is not used to train anyone's model, each client's record is isolated from every other, and nothing is sent to a client without a human pressing send. The plain version of what it does and refuses to do is written out here.