Agentic workflows, step by step

What makes a workflow agentic

A chat answers one request at a time. An agentic workflow takes a job with several steps, works out what to do at each one, and keeps going until the job is finished or it reaches a point where you've told it to stop. In ChatGPT and Claude, that's now something you can set up yourself, without writing code.

The first example is one I run on my own inbox. Every couple of hours it checks for training enquiries, researches the organisation, drafts a proposal and a reply in my voice, and leaves both for me to check. The other two show the same idea in a finance team and a customer service team.

The parts it's built from

Each example is put together from the same handful of parts. None of them is new on its own. The workflow is in how they're joined.

• A schedule starts the work at set times, so nobody has to press go.
• Connectors read from and act in the tools you already use: email, calendar, Trello, a CRM or an accounting system. Each one can be limited to reading.
• Skills hold a written method the AI follows the same way every time, such as how you price, how you write a proposal or how finance checks an invoice.
• A design kit holds your house style, separate from the skill, so every document comes out looking the same.
• Folder access and browser control cover the jobs with no connector. It can work in one folder on a computer, or use a website the way a person would.
• AI judgment is the step where it weighs up what it found, such as whether an invoice adds up or how serious a complaint is.
• A person checks the work before anything leaves. Where that happens is written into the instructions, not left to chance.

Setting the limits

The prompt behind each workflow is short, and much of it is about what the AI mustn't do: never send an email, never pay or approve anything, never offer a customer more than the policy allows. The detail of how to do each step sits in the skills, so the instructions stay readable and the method stays the same from one run to the next. Choose See the prompt behind it at the first stop of any example to read the whole thing.

Where the person comes in matters as much. In the enquiry and invoice examples, someone checks at the end. In the complaints example, the team lead approves the remedy partway through, before a letter is written.

Expect it to slip, then fix the instructions

Setting a workflow up is the start. Once it's running, it meets cases nobody thought of: a reminder that looks like a new invoice, an urgent complaint put politely, a website that changes its layout. At the last stop of each example, choose What can go wrong to see the ones I'd expect and what I'd change.

Every slip points to a gap in the instructions. Fix the instructions, not just the one case, and the workflow gets better each time.

Learning to build these

I demonstrate workflows like these on Advanced ChatGPT and Claude at work, along with the projects, skills, connectors and scheduled tasks they're built from. For a team, I run in-house workshops built around your own documents, so the workflows you see are ones your people could run. Everyone gets the full presentation afterwards, and a set of interactive notes built to keep working through with ChatGPT or Claude.