ChatGPT usually fails to save a one-person business time because a chat window does not remove work, it moves it. The minutes saved on drafting go into giving the model context it does not have, checking what it wrote, and fitting the result back into your email, spreadsheet or client file, and when you are the whole business all three land on you. A study of 25,000 Danish workers in 11 occupations, including accountants and legal professionals, found chatbot users reported saving an average of 2.8% of their working hours, and 59% of the new tasks the tools created were about reviewing, integrating and governing AI output rather than using it. That does not mean AI cannot help a solo operator. It means the chat window on its own is the version where nobody has done the fitting-in work, and what to do next depends on which of three things actually went wrong.
The story tends to go the same way. You run the business on your own (a bookkeeping practice, a consultancy, a small agency) and the admin has drifted into the evenings. You pay for ChatGPT, use it hard for a fortnight, and it is impressive in the moment. A month later the evenings look the same, and you quietly stop opening it.
Most advice reads that as a skills problem: better prompts, a course, a list of fifty things to ask. Sometimes that is true. More often the tool did roughly what it was asked, and the hours went somewhere you were not measuring.
What the research on real use found
Most of the dramatic numbers about AI and productivity come from controlled trials: a defined task, a set time, a tool that fits it. The more useful evidence for you is what happens when people use chatbots in their ordinary working week. Anders Humlum (University of Chicago) and Emilie Vestergaard (University of Copenhagen) surveyed about 25,000 workers across 7,000 Danish workplaces, in late 2023 and again in 2024, and linked the answers to official records of earnings and hours. Their working paper was published by the Becker Friedman Institute in May 2025.
Most users, between 64% and 90% depending on the occupation, said the tools saved them time. But on the days they used a chatbot it saved about 25 minutes, the typical user only reached for it every third or fourth day, and the average saving came to 2.8% of total working hours. Even on the days they used it, the saving was 6 to 7%. Accountants and financial advisors, two of the occupations closest to a typical solo practice, reported savings about half as large as the occupations where the well-known trials had been run.
The limits are real and you should weigh them. These were employees, not owners. They were in Denmark, using the tools of 2023 and 2024, and the time savings are their own estimates rather than a stopwatch. Nothing in the study is about one-person businesses directly. What it does show is that a few percent, not a few hours a day, is a normal result for chatbot use at work, so a disappointing month is not evidence that you used it wrong.
Where the time went
The same study asked whether the tools had created any work. For 17% of users they had. When the authors sorted the new tasks people described, 59% were about enabling and managing the tools (checking the quality of AI output, fitting it into how work was done, working out the rules for using it) and only 41% were about actually using AI to draft or analyse something.
A separate study explains what that feels like from the inside. Researchers at Carnegie Mellon and Microsoft Research surveyed 319 knowledge workers who described 936 real tasks they had done with AI, and published the results at the CHI conference in 2025 (Lee, Sarkar and colleagues). Participants found gathering information easier and checking it harder. Solving the problem got easier, and fitting the answer into the actual work got harder. Doing the task got easier, and supervising it became a job of its own. One lawyer said the tool tended to “make up information to agree with whatever points you are trying to make, so it takes valuable time to manually verify.” A programmer said that checking generated code against what already existed, and altering it to fit, “was more effort than just doing it myself in the end.”
Map that onto an evening of admin. The model can draft the reply to a client in seconds. Before that, you paste in the email thread, explain who the client is, and remind it of what you charged last time, because none of that is in the chat unless you put it there. Afterwards you read the draft properly, because a fluent reply is not a correct one, fix the figure it guessed, and paste it back into your inbox. The drafting really was faster. The work around it was not, and all of it was yours.
The work a company does that you don’t have
The Danish study found one more thing that matters for anyone working alone. Every benefit people reported (time saved, better work, more creativity) was 10 to 40% greater in workplaces where the employer actively encouraged the tools, and that encouragement usually came with training (30% of employees had been on a course) and in-house versions of the chatbot (38% of firms had one). The most common new task of all was integration: somebody working out where the tool fitted into the way work was actually done. The authors suggest that the big trial results describe a carefully managed rollout that many ordinary users simply do not have.
That evidence is about employees, and it does not prove that encouragement causes the bigger gains. But the shape is hard to miss. In a company, someone gets paid to do the fitting-in work. When you are the whole business, you are the person doing the job and the person who would have to redesign it, and the redesign gets the same slot as everything else: after the clients are dealt with, in the evening, which is why it does not happen. The chat window is what AI looks like when nobody has done that work for you.
Three things that go wrong, and only one is fixed by building
“It didn’t work” covers several different failures, and they have different fixes. Go back to the task you tried it on and work out which one you had.
- The output was wrong in ways that took effort to catch. If checking each answer properly takes about as long as doing the work yourself, as with the lawyer’s legal research above, then stopping was the right call. A better setup will not change that much, and neither will we.
- The output was mostly fine, but you were the go-between. You copied the inbox into the chat, explained the client again, and copied the answer back out. That is an integration problem, not a model problem, and it is the one case where building something into the tools you already use genuinely changes the result: the context arrives without you, and you check one draft in the place you would have sent it from.
- The task never needed a model. Chasers on fixed dates, quotes worked out from a price list, the same welcome email to every new client: these are rules, and a template or the scheduling feature already in your software will do them more cheaply, the same way every time.
There is a limit to the second answer as well, and it should be said plainly. Building something does not get rid of the checking. Done well, it gathers the checking into one place and makes the mistakes easier to see, but someone still reads the output before it goes to a client, and if you are not willing to do that, no system makes it safe to hand over. If you cannot tell which of the three you had, measure the task for two weeks before spending anything: how long it takes, how often it goes wrong, and who catches it.
Common questions
- Why didn’t ChatGPT save me time?
- Usually because the time saved on drafting went into other work: giving the model context it does not have, checking its output, and moving the result back into your email or files. A 2025 study of 25,000 Danish workers found chatbot users reported average time savings of only 2.8% of their working hours, and most of the new tasks the tools created were about reviewing and integrating AI output.
- Is ChatGPT worth it for a one-person business?
- It is worth trying on drafting tasks where you would read the result anyway, but expect a saving of a few percent rather than hours a day unless the tool is set up around your own work. In a large Danish study, reported benefits were 10 to 40% greater in workplaces that actively backed the tools, where someone else had done the work of fitting them in. In a one-person business that someone is you.
- Is it that I was not prompting it properly?
- Sometimes, but it is rarely the main reason. In a 2025 Carnegie Mellon and Microsoft Research survey of 319 knowledge workers, people reported that AI moved their effort toward verifying information, integrating the output into their work and supervising the task. Better prompts reduce some of that. They do not remove the checking or the copying in and out of the chat.
- When should I stop trying to use AI for a task?
- When checking each output properly takes about as long as doing the task yourself, or when a mistake would be hard to spot. In those cases a better setup will not change the economics much. Tasks that follow fixed rules are also better handled by a template or a scheduled feature in software you already use.