What AI can do for you

01

What you’re talking to

Whether you use Claude, ChatGPT or Gemini, at its core is a large language model (LLM). It has read most of the public internet, up to a few months ago, but knows nothing recent and nothing private to you: your business, your mail, your calendar, your files. Four things built around it close that gap, and turn a well-read stranger into a colleague.

• Memory

  What it keeps about you from one conversation to the next. A small, curated set of notes you can read and edit, not a transcript of everything you said.

• Built-in tools

  Search the web for anything newer than its cutoff or to check a claim, read files you give it, and run code to work out an answer rather than estimate one.

• Connected tools

  Your mail, calendar, documents, and the other apps your work runs on. Built-in tools find what anyone could find; connected tools find yours.

• The agent loop

  You hand it a task, not a question. It plans the steps, uses its tools, checks its own work and keeps going until it’s done or needs you, within limits you set. That is an agent.

Give your assistant what it needs, and it can become a trusted coworker.

02

What it has access to

Eight ways, from what the model knows on its own to an agent with a machine of its own. They run in order of how much of your world each one touches, and how much it does on its own.

1. 01

   What it already knows

   By itself, the model is strong at explaining, rewriting, summarizing and reasoning through anything you describe to it.

   For example“Explain the termination clause in this contract in plain English.”

   Good to knowIdeas, language and structure are its strength. For specific names, dates and figures, hand it the material or let it look them up.

2. 02

   What it remembers about you

   A small profile it keeps across conversations: your role, your projects, the people you work with, how you like things written. You can read it and edit it in settings.

   For example“Write my Friday update.” It comes back as three bullets to your manager covering the three projects you own, because it kept all of that from earlier chats.

   Good to knowIt is a profile, not a transcript, so a long thread from last month is gone unless it saved a note. Tell it what to keep, and glance at the list now and then.

3. 03

   What it can look up

   Web search, reading the files you hand it, and running code to calculate rather than estimate. All of it comes with the assistant, with nothing to set up.

   For example“Compare three vendors’ current pricing in a table.”

   Good to knowWhen something has to be right, ask it to look it up or give it the file. On hard factual questions, that one change takes it from right about half the time to right nearly every time.

   As of writing, search, file reading and running code come free on Claude, ChatGPT and Gemini alike.

4. 04

   What it can see of yours

   Connected tools that sign in to your mail, calendar, documents, Slack, CRM, or whatever else your work runs on. It sees what you can see there, and no more.

   For example“Go through this month’s support mail and the order system, and list every customer still waiting on a shipment, with their order number and the last thing we told them.”

   Good to knowConnecting your tools is the setup step that changes the most, and it takes minutes. Not every app has a connector yet, and some cover only part of the app. New ones arrive constantly and existing ones grow, so a gap you find today may be closed in a month or two.

   As of writing, Claude calls them connectors and ChatGPT calls them plugins. Many are built on one shared standard, MCP, so a connection often works in more than one assistant.

5. 05

   What it has been taught to do well

   A skill is a set of instructions for one kind of work, which it picks up whenever a task calls for it. Some come built in, you can add your own, and the format now travels between assistants.

   For example“Make the quarterly review deck in our company’s format.”

   Good to knowA skill makes it good at that one job, so the place to write one is work you do the same way every time. Add skills from people you trust, as you would any software.

   As of writing, Claude ships skills for Word, Excel, PowerPoint and PDF and takes your own as an upload. Google’s Gemini Skills read the same format, so a skill written once can travel.

6. 06

   What it can do in a browser

   When an app has no connector, or one that only goes so far, your assistant can use the app the way you do: through its website. It opens the page, reads it, clicks and types, so it can pull a statement from a supplier portal, fill a cart or submit a form.

   For example“Pull last month’s statement from the supplier portal.”

   Good to knowIt stops and hands over to you at logins, payments and the “prove you’re human” checks some sites put up. Plan to be around while it works: most browser tasks come back to you more than once, and the final click is usually yours.

   As of writing, this mostly happens on your own computer with the desktop app open. Both Claude and ChatGPT have a browser built into the app for the assistant to drive, and both can also work in your own Chrome, signed in as you. ChatGPT offers a browser in the cloud as well, which keeps working after you leave, though you sign in to each site there yourself and more sites turn it away.

7. 07

   What it can do on your computer

   Driving the desktop apps installed on your computer, not websites, by looking at the screen and clicking the way you would.

   For example“Pull the aged receivables report out of our desktop accounting software and save it as a spreadsheet.”

   Good to knowFor anything none of the above can reach. It is the slowest and newest of the eight and needs the machine awake, so give it short, well-defined jobs and expect to stay nearby.

   As of writing, Claude’s computer use and ChatGPT’s desktop app can both operate the apps on your Mac or PC, with the desktop app open.

8. 08

   An agent with its own machine

   The same assistant, running on a computer of its own instead of yours, so your laptop can be closed and in your bag. You hand it a job and leave. It works through it and comes back when it is done or needs you.

   For example“Go through every contract in the shared drive, pull the renewal dates into a sheet and flag anything renewing in the next ninety days. I’ll check in this evening.”

   Good to knowIt does its best work through your connected tools and on documents. Websites take more from you there: it starts with none of your sign-ins, so each new site is a request to you, and a site that turns away cloud browsers or asks for a human stalls it until you pick up its message. Save heavier browser work for when you are around.

   As of writing, Claude’s cloud tasks work this way: start one and it carries on after you close the laptop, though anything on your own computer or in a browser waits for the desktop app. ChatGPT’s dots, Grok Bot and Gemini Spark go further, each a standing agent with its own machine that you message like a colleague, and OpenClaw is the open-source version you host yourself.

When it runs

What it has access to is one question. When it works is another, and it decides which jobs you can stop thinking about.

• When you ask

  The default. It works while you are in the conversation, and on the newer assistants carries on in the cloud after you close it.

  “Draft the reply now.”

• On a schedule

  Every morning, every Monday, or every hour at the finest. Each run starts fresh from the instructions you wrote.

  “Every weekday at 7, sum up the overnight tickets.”

• When something happens

  A message arrives, a ticket is filed, an order comes in. The newest of the three, and not yet on every assistant.

  “When an order email arrives, add it to the tracker.”

If you only ever chat, you are using the first three. The work worth handing over starts at the fourth.

03

What you end up with

Five kinds of result, in order of how much is already done by the time it reaches you.

1. 01

   An answer in the chat

   Text, a table, a plan, an explanation, right there in the reply.

   Good to knowFor something you want to read now and don’t need anywhere else. If you’d be copying it out, ask for it in a file or a document instead.

2. 02

   A document it made

   With its built-in tools it can write a real file for you to download: a Word document, a spreadsheet with working formulas, a slide deck, a PDF. It can also make docs, slides, dashboards and small apps that live inside the assistant and export to those same formats.

   Good to knowOpen what it made before you pass it on. The structure and the formulas are usually right; the exact figures are what deserve a look. And once you export from inside the assistant, later edits there do not flow back, so pick one place to finish.

3. 03

   A document in your own tools

   Created or edited straight in Google Docs, Sheets or Slides, in Microsoft 365 or in Notion, through a connector. It lands where the rest of your work lives, with the sharing you already have, and you or your assistant can pick it up again next month.

   Good to knowName the tool when you ask, as in “a doc in Google Docs”. Ask for “a doc” on its own and you will usually get a file instead.

4. 04

   An action through a connector

   Done directly inside the app: a Slack message posted, a ticket filed, a customer record updated, a meeting booked, an email draft prepared.

   Good to knowEverything it does here lands under your name, so set it to show you anything headed for another person before it goes.

5. 05

   An action through a browser

   Done the way you would do it by hand, on a website or in a desktop app: a cart filled, a form submitted, a statement pulled from a portal, a report pulled out of accounting software.

   Good to knowIt fills everything in and stops before paying, deleting or submitting anything it cannot take back, so the last click is yours. This is the kind of result to be around for.

The lower you go on this list, the less is left for you. What remains is the call that should be yours: to pay, to send, to submit.

04

How to work with it

Two things decide how well this goes: what you give your assistant to work with, and what you let it do on its own.

What to give it

• 1The outcome, not the steps. What should be true when it is done? It can work out the route.
• 2The names of the people and things involved. The client, the invoice, the file: whatever it should go looking for.
• 3What must be right. If one detail matters more than the rest, say which.
• 4Access to where the answer lives. A connector beats a description.
• 5What done looks like. “A draft in my drafts folder, addressed, under 150 words” gets you that. “Handle the invoice” gets you its idea of handle.

What you allow

As of writing, every major assistant has the same few permissions, under different names.

Tighten the rules in proportion to how little you will be watching. Start by letting it prepare and hand back rather than send, and widen from there as it earns your trust.

Where you come in

• While it works

  On a longer job it will stop and ask: to sign in, to choose between two ways forward, for something only you know. Quick answers keep it moving, so stay reachable for browser work and for anything that leans on what’s in your head.

• Facts from memory

  It states what it recalls as confidently as what it knows. Check the names, numbers and citations it pulled from memory; anything it looked up or read from your file is on firmer ground.

• Did it do all of it

  On a long task it can narrow the job without saying so, or report done before the result is there. Name the pieces when you set it, and ask for something you can open rather than a summary.

• The final polish

  What it hands back is a strong first pass. The tone, the detail only you know, the call on what to cut: that part is yours, and it’s far faster than starting from a blank page.

Why to start small

An assistant that holds your logins and acts on its own can go wrong in ways a chat cannot. Here are a few cases worth knowing, and each is a reason to give it a little at a time.

An unattended agent can keep going past where you would have stopped it, in what it does or in what it costs. Give a scheduled run a small, clear job.

An always-on agent with stored sign-ins is the richest target you own, and the first real leak from one has already been reported. Give it the few sign-ins it needs and no more.

A page or an email can carry hidden instructions. An assistant reading it while it holds your logins can be steered by them, and every browser agent tested so far has been caught this way at least once. Keep sensitive accounts out of browser work.

Treat your assistant like a capable new colleague: brief it well, check its work, and give it the keys one room at a time.

05

In practice

What this looks like on real work: four tasks, and how much of each came off your plate.

06

Where Underway fits

That is a lot to hold in your head every time you hand something over: what it can get to, what to give it, when to let it run, what to check. Underway holds it for you, so a task is a line you write and the rest is already worked out.

• It starts for you

  Capture a task in a line. Underway writes the brief, the whole of “What to give it” above, written for you, and your AI gets going before you have thought about where to begin.

• It keeps the work, not the chat

  Every task holds its brief, its history and where it stands. Close the chat, lose the thread, switch assistants: the task is still there, and the next conversation picks up where the last one left off.

• It spends your attention sparingly

  You are drawn in only for what is yours: a question, an action, a site to sign in to, each ready to act on. A scheduled run works whatever is ready, so you come back to drafts and findings, not a cold list.

• It learns how you work

  How each task went carries into the next brief: your tone, your rules, what worked with that client last time. The tenth task costs you less than the first.

One queue holds all of it, including the things only you can do, so there is never a second list. The guide has the rest: how a task flows, when to use Underway and when to just ask, and what you need to start.