How to set up your first AI agent at work

To set up your first AI agent at work, pick one job that repeats every week and connect ChatGPT or Claude only to the tools that job needs. Let it read, sort, and draft on its own, and make anything that sends, pays, deletes, or publishes wait for your approval. Run it next to your old way for a couple of weeks, keep a record of each run, and only then let it do more.

Here is each step in more detail, with an example at the end. Product names and features change quickly. The details below are as of October 2026.

Pick one job

Start with one job. Not your whole inbox, and not a whole department. One task that comes back every week, takes at least half an hour, and follows rules you could explain to a new colleague.

Do it once more by hand and write down every step. Note each time you copy something from one tool and paste it into another. Those are the places where an agent can connect the tools directly.

I use three questions to check a first job. Could you explain the rules in ten minutes? Would you notice if the result were wrong? Does the work happen in tools an agent can reach? They are explained in Ask, copy, paste, repeat. The services page has more examples, sorted by role.

Choose the tool

For a first agent, use what you already have. ChatGPT and Claude can both connect to email, calendars, documents, and other apps, and both can work through a job in several steps. For a team, a business plan of either one is the better start, for the data reasons further down.

What turns a chat into an agent is that it can start without you. In ChatGPT, scheduled tasks run at a set time. On eligible plans, a task can also start on the web or mobile app when a new Gmail message arrives, when someone posts in a chosen Slack channel, or on activity in a GitHub pull request. If a task in the ChatGPT desktop app needs files on your computer, the computer has to be on and the app running. Claude’s Cowork can run scheduled tasks too. On Team plans its cloud sessions, still in beta and on by default, keep scheduled tasks running when your laptop is closed. On Enterprise, an owner has to turn them on first.

Sources: OpenAI, Scheduled tasks; OpenAI, Connected apps in ChatGPT; Anthropic, Claude Cowork on Team and Enterprise plans; Anthropic, Connectors.

Connect only what the job needs

An agent can only do what its connections allow, so this is where most of the safety comes from. Connect the apps the job needs and nothing else. If the job reads invoices from email, it needs email. It does not need your bank.

Also know whose access the agent uses. Anthropic’s documentation says a Claude connector inherits each person’s permissions in the connected service, and that a custom connector with a shared credential reaches whatever that credential can access. On ChatGPT Business and Enterprise, OpenAI says connected apps respect your organization’s existing permissions, and each person signs in to an app before using it. In both cases, an agent connected with your account can see what you can see.

Where you can, start with reading only. On Claude’s Team and Enterprise plans, an owner can let a connector read from a service while blocking it from writing changes back.

Decide what waits for you

My rule is simple. Reading, sorting, summarizing, and drafting can run on their own. Anything that sends, pays, deletes, or publishes waits for a person.

The easiest way to set this up is to let the agent prepare the work and stop before the last step. It writes a reply and saves it as a draft in your email. It lists the invoices to pay, and you pay them. It writes the post and leaves it unpublished. Your approval is the moment you open the draft and press send.

Some tools let you set this for each kind of action. On Claude’s Team and Enterprise plans, owners can set each connector action to Always allow, Needs approval, or Blocked.

Sources: Anthropic, Connectors; OpenAI, Enterprise privacy.

Check the data settings

Before you connect work data, check what happens to it. The main question is whether the provider may use your conversations to train its models.

• On personal ChatGPT plans, you can turn off “Improve the model for everyone” under Settings, then Data controls. Your new conversations are then not used for training. On personal plans the same setting also covers Codex tasks.
• On ChatGPT Business and Enterprise, OpenAI does not train on your workspace content by default. Deleted conversations are removed from its systems within 30 days, unless the law requires keeping them longer or OpenAI needs them to protect its services.
• On Claude Free, Pro, and Max, training is also a setting you control. If you allow it, Anthropic may keep de-identified data in its training pipelines for up to five years. Claude’s Team plan lists no model training on your content by default.

If the job touches customer or employee data, use a business plan, and tell whoever looks after data protection in your company what you are connecting.

Sources: OpenAI, Data controls in ChatGPT; OpenAI, Enterprise privacy; Anthropic, How long do you store my data; Anthropic, Claude plans.

Run it next to the old way

For the first couple of weeks, keep doing the job the old way as well, and compare the two. Check whether the agent found the same items, and whether it missed an exception you would have caught.

OpenAI’s guidance for scheduled tasks on the web is the same: test the prompt in a regular chat first, review the first few runs, then adjust the prompt, tools, or timing if the results are too broad or miss context.

Most fixes in this phase go into the instructions. The agent didn’t know that one customer always gets a different price, or that one supplier sends its invoices as two files. Add each of these to the instructions as you find them.

Keep a record of each run

Every run should leave a trace you can check later: when it ran, what it read, what it did, and what you changed before you approved it. ChatGPT’s Scheduled view shows each task’s status and recent runs. I also keep a simple log, one line per run, in a shared document or spreadsheet.

The record does two things. When something goes wrong, you can see where. And after a few weeks, it shows how often you had to correct the agent. That is how you decide whether to give it more.

Then widen it

Once the runs are boring, and you approve most of them without changes, widen the job one step at a time. Let it handle one more kind of case. Connect one more tool. Or remove the approval for one low-risk action, such as filing, while you keep it for sending.

Change one thing at a time, so you know what caused any change in the results. Then pick the next job and start again from the top.

A worked example

This is an example, not a client story. It shows how the steps fit together for a common job.

Imagine a small design studio that sends each client a short status email every Friday afternoon. Someone spends an hour going through the project board and the week’s emails to write them.

Nothing in this example needs custom software. It needs one clear job, the right connections, and a person who reads the drafts.

If you would like help choosing and setting up your first job, write to me. I do this as a personal setup for individuals and as a short sprint for teams.