What ChatGPT can actually do for a founder, and where it stops
ChatGPT is reliable at four jobs inside a small company. It drafts from a brief you already hold. It reads your own documents. It does structured research against criteria you set, and it turns a repeating task into a reusable format. It fails quietly at three things: specifics it was never given, contradictions between sources, and anything dated. Telling the two apart is what you have to learn.
What can ChatGPT actually do for me?
Four jobs hold up under repetition. The first is drafting from a brief you already carry in your head. A founder selling a wellness range into a new market can name the buyer, the objection and the format, and get a proposal or an outreach note back in that shape. The draft is raw material. You still write the sentence that commits you.
The second is reading your own material. Upload the monthly sales report, the three research PDFs, the distributor contract, and ask for a table, the open questions and the next actions. The common file types a small company already holds are supported, spreadsheets and CSVs, Word files and PDFs. This is the cheapest work of the four, because the input is yours and the facts in it are already checked.
The third is structured research against criteria you set. Scoring three candidate export markets on acceptability, price point, customer acquisition cost and regulatory friction is the shape of it. You name the criteria, you name the candidates, you ask for a source against each cell. The criteria are the work. Without them you get an essay with a confident last paragraph.
The fourth is turning a repeating task into a reusable format. A daily report that always needs the same five questions answered becomes a standing brief, and then a five-slide deck in the same layout every morning. The first build costs an hour of your attention. The tenth run costs two minutes.
Where it stops, and how the failure looks
The output will often carry its own line on what it cannot reliably confirm. Read that line and take it at face value. The three failures below are the ones where it says nothing, and that silence is what makes them expensive. They do not arrive as errors. They arrive as clean paragraphs in good English.
The first is a specific that was never supplied. Ask for landed cost into a new market and you can get a figure that reads like a sourced number and is not one. The check is mechanical. Every figure you would act on gets a source you can open, or it comes out of a document you uploaded.
The second is a contradiction smoothed flat. Combine three research reports on the same market and the output tends to be one tidy answer where the three disagreed. Ask for the disagreements by name before you ask for the summary. A table with a conflicts column is worth more than a paragraph of consensus.
The third is anything dated. Duty rates, a regulation, a competitor's public claims, last month's closing prices. Treat a dated fact as a question to take to the primary page yourself, not as an answer.
None of the three gets fixed with a better prompt. They are the reason sign-off stays with the person whose name is on the pitch.
What AI tools should I use for my business?
One, and then the checks. For a company of two to twenty people the tool list is the wrong first question, because the constraint is which task you are willing to hand over and then inspect. Start with ChatGPT and the checks. Put your own documents in it and run the four jobs above on live work for a month.
A founder doing marketing, research and finance alone has no spare hour to evaluate software. There is a pitch due Thursday. So the sequence that works is task, then check, then tool. Take the weekly export enquiry log: every enquiry summarized, sorted by size and fit, with a reply drafted for each. Name it, then write down what a wrong answer looks like, a buyer's volume read wrong or a country named that you do not ship to. Then run ChatGPT on last week's enquiries and see whether it does that job.
Adding a second tool before the checks are habit leaves you with two outputs you cannot audit. I have watched that cost more time than the tool saved.
What do founders ask next about these limits?
What do I do when the same task fails the check twice?
Stop handing that task over and change what you give it. Two failures in a row on the same task usually mean the input was thin, not that the output was careless. Supply the document, the figures or the constraint that was missing. If the third run still fails, that task stays with you.
Should I upload my own files or paste the text in?
Upload the file when the structure matters, paste when only the words do. A spreadsheet, a contract or a long PDF keeps its rows and headings as a file and loses them as pasted text. Before uploading anything, decide what you would be comfortable seeing outside your company.
Does a longer prompt give a better answer?
Length helps only where it adds constraint. Role, context, task, format and constraints earn their words because each one removes a wrong answer. Adjectives do not. A brief that names your buyer, your market and the output format beats a paragraph of encouragement every time.