ChatGPT is most useful in a small business for work where a good first draft saves time and a person can quickly check the result: drafting and rewriting, summarizing long material, translating, structuring messy notes, and thinking through a problem out loud. It is weakest where accuracy depends on facts it does not have, such as your prices, current regulations or what a customer actually said. The same is broadly true of Claude, Gemini and Copilot. Judge any of them by the work it helps you finish and the mistakes it leaves for you to catch.

Every new model release arrives with benchmark charts and bold claims. I have stopped reading those as reviews. A useful review of an AI tool needs context: what work it was used for, how often the output was usable, and what had to be fixed. This is mine, from using these assistants daily in my own work on websites, systems and operations for businesses in Japan.

Where does ChatGPT save real time?

Drafting and rewriting

This is the most reliable use. A proposal outline, a reply to an awkward email, a process document, a job description, an announcement to staff. The first draft is rarely the final one, but starting from a structured draft is faster than starting from a blank page. The gain is largest for writing I find tedious rather than writing I find hard.

Summarizing long material

Meeting transcripts, long email threads, vendor proposals and documentation all compress well. I ask for decisions, open questions and action items with owners, rather than a general summary. The output still needs a skim against the original, because summaries sometimes drop the one caveat that mattered.

Translation and bilingual work

For businesses that run in English and Japanese, this may be the single biggest benefit. Translating internal documents, drafting bilingual announcements and understanding a Japanese vendor contract before a proper review are all much faster. For anything customer-facing, a fluent speaker still reads the result. I cover how to brief the model for Japanese business writing in how to get reliable work answers from AI tools.

Turning mess into structure

Rough notes into a checklist. A description of how someone handles invoices into a step-by-step procedure. A pile of requirements into a comparison table. This is quiet, unglamorous work that keeps projects moving, and AI handles it well because the source material is all in front of it.

Thinking out loud

Explaining a problem to the model and asking it to push back is useful, particularly with a reasoning mode switched on. It surfaces options I had not listed and questions I should ask a vendor. I treat the answers as prompts for my own thinking, not conclusions.

Where does it still need checking?

Facts it was not given. Ask about a Japanese regulation, a vendor’s current pricing or a company’s history and the answer may be outdated or invented. Web search features help, but I still open the sources.

Numbers and analysis. Reasoning modes are much better at working through data than fast modes. They still make errors in calculations and in reading tables. I check any figure that will go into a decision.

Confident filler. When a request is vague, the answer is fluent and generic. It reads well and says little. This is the easiest failure to miss, because nothing is obviously wrong.

Your context. The model does not know your customers, your team or the history behind a decision unless you tell it. Projects and custom instructions help, but the burden of supplying context stays with you.

How should you judge a new model release?

New versions of ChatGPT, Claude and Gemini arrive every few months, and each is described as a major step. Rather than switching based on announcements, keep a short set of your own test tasks:

  • A real email you had to write in Japanese, with the context you would give a colleague
  • A messy document to summarize into decisions and actions
  • A small data analysis where you already know the right answer
  • A question about your industry where you know the common misconceptions

Run the same tasks through the new model and the one you use now. Note which output you would actually send, and how much you had to fix. Twenty minutes of this tells you more than any benchmark, and it stops you chasing releases that do not change your work. The same habit of checking results against a baseline applies to any automation, as I found in what 851 unattended bot runs taught me about automation.

Does it matter which assistant you use?

Less than the marketing suggests. ChatGPT, Claude and Gemini are all strong at the core tasks above, and each has periods where it leads on something. For most small businesses the practical questions are different:

Many people use more than one. That is fine for individuals. For a team, pick one main assistant so prompts, rules and training are shared.

The honest verdict

Used well, a general AI assistant takes a real share of the connective work in a small business: first drafts, summaries, translation and structure. It does not remove the need for someone who knows the business to decide, check and take responsibility. The time it saves is real, and so is the time it costs when its output is used unchecked.

The bigger gains come when the business around the tool is organized: a current FAQ, clean templates, documented processes and data in one place. That is where I spend most of my effort, and why I argue in why AI adoption stalls in Japanese SMEs that the setup matters more than the model.

If you want help choosing an assistant for your team and setting it up with sensible rules, I can do that as part of implementation work, or you can tell me what you are trying to do.

Further reading: how to get reliable work answers from AI tools · how to evaluate an AI vendor before you buy · which work should stay with people when you adopt AI