In December I wrote about my first weeks with ChatGPT. That was first-date enthusiasm: I threw everything at it to see where it would break. Three months after its launch on 30 November 2022, ChatGPT has moved from “interesting toy” to “tool I open every day, like email”. In the meantime OpenAI launched the paid ChatGPT Plus subscription, and on 1 March it opened the API, which means developers can now build the model straight into their own apps.
I do not want to write again about how impressive it is. Plenty has been written about that. I want to be concrete: what I use it for in the companies I am involved in and in my advisory work with startups, where it has helped and where you need to be careful. No promises that it will double your productivity. Just what I see on my own desk.
Where I actually use it
After three months, my use has settled into a few clear categories. None of them are spectacular, and that is exactly the point: the gain comes from small tasks repeated dozens of times a week.
- First drafts of emails and proposals. A sales email, a diplomatic reply to an awkward request, a partnership proposal. I give it the key points and the tone I want and get a version to work from. I never send the first draft, but I no longer start from a blank page.
- Summarising long documents. A report, a long contract, a market study. I ask for the main points, the risks, what is missing. Then I read the document myself, but I already know where to look.
- Brainstorming. Product names, campaign angles, questions an investor might ask. Out of twenty ideas, eighteen are bland. But two make me think, and two good ideas are worth it.
- Structuring presentations. Before a pitch or a meeting, I ask for a skeleton: which slides, in what order, what message on each. I fill in the substance.
- Translating and adapting between Romanian and English. I work in both languages every day. I do not just ask it to translate but to adapt: “make this sound natural to an American reader”. It is very good at this.
- Simple code and spreadsheet formulas. I am not a programmer. But a tricky spreadsheet formula, a small script to clean up a list, an expression for a filter: I get them in seconds, with an explanation.
- Draft replies for customer support. Answers to frequent questions that a person reads, corrects and only then sends. A human stays in the loop, always.
The limits you cannot ignore
This is the part I repeat to every founder I work with. ChatGPT is extremely convincing. That is exactly the problem.
It makes things up, confidently. It will give you a number, a name, a study, a law, a source, all beautifully written and all possibly false. It is not lying on purpose; it simply generates the most plausible text. My rule: every fact, every figure, every quote gets checked. No exceptions. If you do not have time to verify, do not use it for facts.
Do not paste confidential data. Whatever you type goes to someone else’s servers. Client data, contracts with names and amounts, internal financials, passwords: no. If I want a contract summarised, I strip out the names and sensitive figures first. It is a new discipline teams have to learn, and it is better learned now than after an incident.
It does not replace judgment. It can write ten versions of a strategy. It cannot tell you which one fits your people, your market, your money. The decision stays with you, and so does the responsibility. An email sent with AI-generated mistakes is still your email.
Quality depends on how you ask. The same tool gives one person bland answers and another excellent ones. The difference is not luck, it is the question.
How I write a good prompt
A “prompt” is the instruction you give it. Through trial and error I have learned a few simple things that make the difference between a generic answer and a useful one:
- Give it context. Who you are, who the text is for, what you want to achieve. “Write an email” gets you any email. “Write an email to a partner who delivered late, firm but without damaging the relationship” gets you something usable.
- Give it a role. “Answer as a sceptical lawyer”, “as an investor looking for reasons to say no”, “as a customer in a hurry”. The role changes the perspective a lot.
- Give it examples. If you have a text whose style you like, paste it in. The model imitates very well.
- Specify the format. A list, a table, three paragraphs, 100 words maximum. Otherwise you get essays.
- Iterate. The first version is rarely the right one. “Shorter”, “more direct”, “no clichés”, “now in English”. It is a conversation, not a command.
Honestly, the most useful trick: I ask it to ask me questions before it answers. “What information are you missing to do this well?” Its questions often show me what I had not thought through.
What is moving around it
It is not just OpenAI. In early February Microsoft announced the new Bing with built-in chat, and Google announced Bard, for now available only to a small group of testers. Web search as we have known it for about twenty years is starting to change. For anyone doing online marketing, that is a signal to watch closely, just as I have long watched how algorithms decide what we see online.
The API released on 1 March is, I think, the most important news for entrepreneurs. Until now ChatGPT was a window you type into. From now on it can become a feature inside your product: an in-app assistant, an automatic reply, a summary generated on the spot. When I wrote in 2020 about the first machine-written text that sounds human, access was for the few. Now it is within reach of any small team.
What I tell teams and founders
I tell the startups I work with three things. One: try it on your real tasks, not on riddles. Two: set clear rules from day one, which data never goes in and what a human must check. Three: do not use it as an excuse to hire weaker people. Use it to let the good ones do more of what matters.
Because that, really, is my conclusion after three months. ChatGPT does not turn an inexperienced person into an expert. Someone who does not know what a good proposal looks like will not spot a bad one written by AI either. But for someone who knows their job, it gives back hours every week: the drafts, the summaries, the translations, the formulas. It is a multiplier. And a multiplier applied to zero still gives zero.
In short
It is an excellent tool for the first draft, not the final one. Check everything, keep sensitive data out, make the decisions yourself and learn to ask well. With those rules, it is the best assistant I have ever had for the price of a subscription.
How do you use it in your day-to-day work? And what is the biggest mistake it has made for you so far?