How can AI help with emails, summaries and translations?

Most people who try a chatbot end up using it for the same three chores: writing emails, shrinking long documents into something readable, and making sense of text in another language. These are good jobs for it. The tools behind ChatGPT, Claude, Gemini and Copilot are large language models, programs trained on huge amounts of text to predict what words should come next, so rewording and condensing text is close to what they were built for.

The catch is that the same skill makes their mistakes hard to spot. A wrong date in a fluent, polite email looks exactly like a right one. So the useful habit is simple: let the AI do the first draft, and keep the last look for yourself.

If you want the general rules for asking well, our guide on how to write a good prompt covers them. This piece sticks to the three tasks, with prompts you can copy.

Where can I use AI for email?

You don't have to open a separate chatbot. Gmail's Help me write can draft a new email or "refine existing text in your draft for tone and clarity", with buttons to make a draft more formal, friendlier or shorter, according to Google's help page. On personal Google accounts the page says it is "restricted to the US only"; elsewhere it needs a paid Workspace or Google AI plan.

Outlook has a similar Draft with Copilot button above a new message, and Microsoft's help page says you can then change the draft's length and tone from a menu. Whether you see it depends on your Microsoft 365 plan, and at work on your IT department.

Both companies tell you to review before sending. Google's page warns that Gemini "may suggest inaccurate or inappropriate information". Microsoft asks you to check and edit the draft before you press Send.

A general chatbot works just as well: paste in what you have, say what you want, copy the result back.

What should I ask for when writing an email?

Three kinds of request cover most of what people need. Replace the bits in capitals with your own details.

Drafting from notes. Start with the facts in any order and let the AI arrange them.

Turn these notes into a short, polite email to my child's school. Notes: off sick Monday and Tuesday, fever, back Wednesday, please send any homework. Sign it from NAME. Under 100 words.

Rewriting something you already wrote. This is often more useful than starting from nothing, because the content is yours.

Here is an email I wrote to my landlord. Keep every fact, but make it calmer and more direct, and cut it by a third. Don't add anything I didn't say. EMAIL TEXT

Replying to something tricky.

I received the email below from a supplier saying my order is delayed again. Draft a firm but courteous reply asking for a confirmed delivery date by Friday, and mentioning that I may cancel if it slips again. EMAIL TEXT

The phrase "don't add anything I didn't say" earns its place. Chatbots like to be helpful, and sometimes that means inventing an apology, a promise or a detail you never gave them.

What should I check before I press send?

Read the draft as if someone else wrote it, because something else did. A quick checklist:

• Names, dates, times, amounts and addresses. These are the details most worth checking against your own notes. An AI can swap a Tuesday for a Thursday without blinking.
• Promises. Make sure the draft doesn't commit you to something you didn't intend, such as a refund, a deadline or a meeting.
• Tone. Does it sound like you? Over-polished phrasing can feel cold to people who know how you normally write.
• Recipients and attachments. The AI can't see those. You still have to.

Tone matters more than it seems. In a survey of 1,100 professionals by Anthony Coman of the University of Florida and Peter Cardon of the University of Southern California, employees saw managers as sincere in 83% of cases when the managers' messages used little AI help, but only 40% to 52% when they used a lot, the university reported. The researchers found AI acceptable for grammar fixes and routine messages such as meeting reminders, and poorly suited to congratulations, praise and other personal messages.

How good is AI at summarising long documents?

It can read a 40-page report or a long email thread in seconds and give you the gist, which is genuinely handy for a tenancy agreement, a school policy or meeting minutes you missed. You can paste text in or upload a file. OpenAI's help page, for example, says ChatGPT accepts most common document types, up to 512MB per file.

Two limits are worth knowing. First, what the AI actually reads may not be everything on the page. The same OpenAI page says that outside its Enterprise plan, ChatGPT reads PDFs by extracting the digital text and discards images, so charts, scanned pages and photos of tables may simply be skipped.

Second, summaries drift. Researchers Uwe Peters and Benjamin Chin-Yee compared 4,900 summaries of scientific papers written by 10 popular AI models with the originals, in a study published in Royal Society Open Science. Some models overstated the findings "in 26 to 73% of cases", according to the paper's abstract, and AI summaries were nearly five times more likely than human-written ones to make broad claims the original did not support. Telling the models to avoid inaccuracies did not help.

News is no easier. A study coordinated by the European Broadcasting Union and led by the BBC, with 22 public broadcasters in 18 countries, checked more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity about news stories. It found that 45% had at least one significant issue and 20% had major accuracy problems, such as made-up or outdated details, VRT reported in October 2025. These invented details are called hallucinations.

So use a summary to decide what to read, not instead of reading the part that matters.

Which summary prompts work well?

Ask for the summary to point back to the original, so you can check it.

Summarise the attached document in five bullet points for someone with no background in the subject. After each point, quote the sentence from the document it is based on. If the document doesn't say something, don't guess.

This is my rental contract. List every date, deadline, fee and penalty it mentions, with the clause number for each. Then tell me which clauses are unusual or unclear.

Here are the minutes of a residents' meeting I missed. What was decided, who agreed to do what, and by when? Leave out the discussion.

Then open the original at the quoted sentences and see if they say what the summary claims. It takes two minutes and catches most problems.

Can AI take notes in my meetings?

Video-call tools now offer this. Google Meet's Take notes for me, for instance, writes a summary with action items into a Google Doc for the organiser, and Google says Meet tells all participants that notes are being taken. The same help page warns that the summary "can be incomplete, inaccurate, or not generated", and that it works only in a handful of languages, one per meeting.

The courtesy rule is the same as for recording: tell people before you switch on any AI note-taker, and read the notes before you send them round. A wrong name next to an action item causes real friction.

Is AI good enough for translation?

For getting the gist of a letter, a menu or a message from a relative, yes, usually. For anything where one wrong word has consequences, not on its own.

The clearest evidence comes from medicine. Researchers at the University of California, San Francisco, writing in BMJ Quality & Safety, had ChatGPT and Google Translate translate 50 sets of real hospital discharge instructions from English into Spanish, Chinese and Russian. Sentence by sentence, ChatGPT was accurate 97% of the time in Spanish, 95% in Chinese and 89% in Russian. But because each set had several sentences, 56% of the Russian sets translated by ChatGPT contained at least one inaccurate sentence. The authors concluded that AI could be used cautiously for low-stakes English to Spanish messages when no professional is available, and advised against using machine translation without professional review for high-stakes content in any language.

The lesson carries over to everyday life: quality varies a lot by language, and small error rates add up over a long text.

Which translation prompts help?

Translate this message from my neighbour from Italian into English. Then list any words or phrases that could mean more than one thing, and explain the options.

I want to send this reply in German to a landlord. Translate it, keep it polite and formal, then translate your German version back into English so I can check the meaning.

The back-translation trick is not proof, but it catches meanings that went missing. If you know someone who speaks the language, a 30-second read from them beats any tool.

Is my information safe when I paste it in?

Treat anything you paste as shared with the company that runs the chatbot. Emails and documents often contain other people's names, phone numbers, account numbers or health details, and those people never agreed to share them. Replace them with placeholders such as "Client A" or "my bank" before pasting. For work documents, use only the tools your employer has approved.

Our guide to your data and AI chatbots explains what each company keeps and how to switch off training on your chats.

When should I not use AI at all?

• Legal, medical and official documents. Contracts, court letters, prescriptions, visa or benefit forms. Use AI to understand them if you like, but get a professional to translate or check anything you submit. Official bodies often require a certified translation anyway, so ask before you start.
• Personal and emotional messages. Condolences, apologies, congratulations, a difficult conversation with someone you love. The University of Florida findings suggest people notice, and the effort is part of the message.
• Anything you can't check. If you don't know the subject or the language well enough to spot a mistake, the AI's confidence will hide it from you.
• Other people's secrets. If it isn't yours to share, it isn't yours to paste.