AI

AI at work: what it really changes, from inbox to itinerary

In depth·International

How AI helps with email, meetings, spreadsheets, code, research and travel, and where it slips.

A woman works at a laptop at a long wooden kitchen table, writing notes under black pendant lamps
A woman works at a laptop at a long wooden kitchen table, writing notes under black pendant lamps
Photo CreditImage courtesy of © Shixart1985, CC BY 2.0, via Wikimedia Commons

The new colleague arrives without a lanyard, reads everything and never takes lunch, but occasionally invents a court case. Here is how AI assistants now help with real work, from the inbox and the spreadsheet to code and travel plans, what the best evidence says they change, and the habits that keep them honest.

A colleague with a very long reading list

The best evidence on AI at work is quieter than the headlines, and more interesting. A study published by America’s National Bureau of Economic Research followed 5,179 customer-support agents given an AI assistant. They resolved 14 per cent more issues an hour. The newest and least skilled improved by 34 per cent; the most experienced barely moved.

The tool, in other words, worked less like a robot replacement than like a patient mentor sitting beside the beginners.

That finding frames the question every office now faces. AI is excellent at the first draft, the summary and the search; it is weakest exactly where experience is strongest, in judgement and in knowing what has been left out. The useful way to think of it is as a capable junior: fast, tireless, widely read, and in need of supervision.

Email and documents

This is where most people start, and where the assistants now live inside the tools themselves. Google AI Pro puts Gemini in Gmail, Docs and Vids. ChatGPT offers extensions for Excel, Word, PowerPoint and Google Sheets; Claude’s paid plans connect to Microsoft 365. The trick is to give the assistant your own voice: paste in a few emails you were proud of and ask it to write like that. Left to itself, it writes like a very polite committee.

Meetings

Google says Gemini in Google Meet can take the notes and translate the conversation. A good summary turns an hour into a page of decisions and owners, which is no small mercy. Two habits make it work: tell everyone the meeting is being transcribed, as courtesy demands and many employers require, and read the action points before you send them, because a transcript cannot hear a raised eyebrow.

Spreadsheets and analysis

Assistants have become surprisingly able analysts. ChatGPT offers data analysis with interactive tables and charts; Claude’s paid plans add Claude Data and Sheets, still in beta. OpenAI says early customers used GPT‑6 Astra for spotting discrepancies in financial statements, and Google says Gemini 4 Carbon is strong on charts and dense documents. Ask for the formula behind every number: a figure you cannot trace is a figure you cannot defend.

Coding

Software is where AI has changed working life most visibly. Claude Code comes with Claude’s paid plans, OpenAI’s Codex with every ChatGPT plan in some measure, and Google offers its Antigravity platform. GitHub Copilot has individual plans at $10, $39 and $100 a month, with a coding agent that turns an assigned task into a pull request. Its higher plans can also hand work to Claude Code and Codex, so the rivals now turn up in the same toolbox.

Cursor, the AI-first code editor, has had a dramatic year: in June SpaceX, now home to xAI, agreed to buy its maker, Anysphere, in a deal reported at $60 billion. Its plans are now sold in one bundle with Grok, from $8 a month, with other makers’ models billed by use. For developers, the choice increasingly comes down to which editor you like to live in, since most offer the same leading models.

Do the tools make programmers faster? The most careful study is a cautionary tale. In early 2025 the research group METR found experienced open-source developers took 19 per cent longer with AI, while believing it had made them about 20 per cent faster. Its February 2026 follow-up suggests they are probably faster now, but calls its own data very weak evidence: many developers refused to take part because they no longer wanted to work without AI.

Let the assistant draft, sort and summarise; keep the checking, the judgement and the signature for yourself.

Research

Deep research tools in ChatGPT, Gemini and others read dozens of sources and return a report with footnotes in minutes. They are superb for getting your bearings and dangerous as a final word. In June 2025 the High Court in London warned lawyers after fictitious case citations, of the kind AI tools invent, reached the court, saying those who fail in their duties “risk severe sanction”. Open every source the assistant cites; if it will not open, it may not exist.

Two colleagues at a desk with laptops, one pointing a pencil at a printed page of diagrams and notes
Two colleagues at a desk with laptops, one pointing a pencil at a printed page of diagrams and notes
Photo CreditImage courtesy of © helloquence, CC0, via Wikimedia Commons

Travel planning

For the traveller, an assistant is a tireless concierge for the first draft: a three-day itinerary for Kyoto, trains that connect, restaurants near the ryokan. The details are where it slips. In 2024 a tribunal in British Columbia held Air Canada liable after its own website chatbot described a bereavement-fare refund the airline’s policy did not allow.

Confirm visas, opening hours and fares on official sites, and book through the provider itself.

Business travel is changing too. Among the enterprise plugins OpenAI launched alongside GPT‑6 Astra is one for Navan, the corporate travel platform, so that an agent can work with the tool where trips are actually booked and expensed: the paperwork end of travel, which few travellers will mourn.

When the assistant does the work

The step change of 2026 is the agent, an assistant that operates software rather than advising on it. OpenAI says GPT‑6 Astra can work everyday applications even without a programming interface; Claude runs scheduled tasks and works in Chrome; Google AI Ultra includes Agent Mode. OpenAI lets companies require approval before consequential actions.

Use such controls wherever they exist: delegation is a fine thing, until it emails the whole company.

Good practice: check, protect, disclose

Check: treat every fact, figure and citation as a lead to verify, never as a finding. Protect: personal plans may use your conversations for training unless you opt out, and Google says a sample of Gemini chats is reviewed by people, so client files and colleagues’ personal details belong only in a work account your employer has approved. Business plans carry their own data terms; ask your IT team what yours say.

Disclose: tell clients, editors and colleagues when AI drafted something that matters. It costs nothing, and it is far cheaper than explaining later. And keep a human name on every decision: the assistant can write the letter, but it should not be the one to sign it.

The verdict

AI is now genuinely useful at work, most of all for the first draft, the summary, the spreadsheet formula and the routine code. The evidence so far says it lifts beginners most, and that experts can overestimate how much it speeds them up. The open questions, how jobs and training change when the junior work is automated, are being argued by economists in real time. For now the best practice is simple: delegate the drudgery, keep the judgement.

Why it's on our list

  • Email, documents and meetings
  • Spreadsheets and analysis
  • Coding tools compared
  • Research and travel planning
  • What the studies actually found
  • Good practice: facts and privacy

Location

The map comes from Google Maps, which sets its own cookies once it loads.

Open in Google Maps

The ULTIMATE letter

Brighton in your inbox

An occasional email with new places, events, neighbourhood guides and our latest in-depth articles. No spam, ever.