Running your reselling books from ChatGPT
The reason reselling records go stale is never that the arithmetic was hard. It is that writing them down happens at the worst possible moment: in a car park, holding a box, with one hand free. A chat window is a genuinely good answer to that, provided the numbers it gives back are yours and not its own invention.
Status, so nobody is misled. The Flippo app for ChatGPT has been submitted to OpenAI's app directory and is in review. It is not live yet. The Flippo iPhone app is likewise submitted to the App Store and in review. Everything described below works today in the Flippo web app, which runs in any phone browser.
The part that actually fails
Ask any reseller where their bookkeeping breaks and you will hear the same shape of answer. Not "I could not work out the margin". It is that four purchases from a car boot sale in March were never entered, so the March figures are wrong, so nobody trusts the April ones either, so the sheet quietly stops being maintained.
Entry is the bottleneck, and entry is a typing problem. Every form asks for a date, a source, a total, a currency, a list of items, a condition for each one. All of it is reasonable and none of it is what you want to be doing while someone waits for you to move your car.
"Bought a box of twelve PS2 games at a flea market for 400 kroner" contains every one of those fields. Something just has to take it apart and put it in the right columns.
Why a chat alone cannot do this
It is worth being precise about the limit, because the failure mode is subtle and expensive.
A language model with no connection to your records will answer "what did I make last month?" anyway. It will produce a number with a plausible shape, in the right currency, with a sensible-looking margin. The number will be fiction. It has no rows to add up, so it writes something that reads like the answer to that question.
For a reseller that is worse than no answer at all, because it looks exactly like a real one. Nothing in the reply tells you it was invented.
So the only version of this worth building is one where the chat cannot answer from memory. Each question becomes a lookup against the ledger, and each figure returned is a figure that exists in a row, already converted to the currency you report in. If the data is not there, the honest answer is that the data is not there.
What the chat can and cannot be allowed to do
Reading is the easy half. Which items have been sitting longest, what is unlisted anywhere, what a purchase actually returned once the last unit went, which platform pays best for a given category, what the month looks like after fees and postage. Those are queries, and a query that returns nothing is merely unhelpful.
Writing is the half that needs care. Recording a purchase, logging a sale, marking one unit of five as sold, deleting something. A model that misreads an instruction here does not produce an unhelpful sentence, it changes your books.
The rule we settled on is that a write is described before it happens, never after. The assistant states exactly what it is about to record, you confirm, and the confirmed instruction is executed as written. Not re-derived, not re-asked, not "interpreted a second time" - the same staged call, which is the difference between one purchase and two identical ones.
The other half of that rule is scope. An assistant that can read a database needs a hard boundary at the edge of your own account: every lookup and every write filtered to the signed-in user, enforced by the server rather than by the wording of a prompt. A prompt is a request. Access control has to be code.
Where a sentence beats a form
Three moments, specifically.
At the point of buying. One sentence while the thing is still in your hands, rather than a form that evening, by which time you have forgotten what the seller knocked off.
At the point of selling. A sale notification arrives on your phone. One sentence records it, and the unit is off your inventory before you have put the phone away.
The questions you never bother to look up. "What have I got that has been in stock more than ninety days?" is a real question with a real answer, and almost nobody goes and builds that filter. Asked in a sentence, it costs nothing, so it actually gets asked.
What it does not change
A chat interface does not improve the numbers. Profit is still sale price minus what you paid minus every fee and every postage label, ROI is still measured against the purchase price, and a cost that was never recorded is still missing whether you failed to type it into a form or failed to say it out loud.
What changes is how much friction stands between a thing happening and the thing being written down. That is the whole of it, and for most small resellers it is the only part that was ever the problem.
The ledger underneath is live today
Purchases, individual units, listings, sales and expenses, with profit and ROI worked out per item. The ChatGPT and iPhone apps are in review; the web app runs in any browser now.