ChatGPT, Claude, Copilot or Gemini · Practical walkthrough

Map an AI workflow before trying it yourself

The demo showed a chat, a spreadsheet and something posting to Slack. Find out which parts you can use before borrowing the whole setup.

Task by Task6 min read

What you'll make

a short map of what your technical team’s AI workflow produces, with the unknowns marked, and one small job of your own with a result you can check.

What you'll need

the AI tools your organisation has given you, a note or demo of the workflow you want to understand, one everyday task you know well enough to review, and about 20 minutes. You don’t need to connect accounts or ask for new permissions to follow this.

Example result.

What I saw	Kind	What I’d need before using it
Release notes summarised in chat	Chat	A work account allowed to handle those notes
CSV helper page	Small tool	The file it was tested on, and a known answer
Auto-post to #releases	Connected action	Who set it up and what it’s allowed to post

1. Ask how one workflow really works

Paste the workflow documentation or demo notes into your work chat.

Explain this workflow for someone who isn't technical. Map its input,
what starts it, what it produces, who reviews it and anything it reads
from or writes to another system. Say which parts are chat, file work,
a small tool or a connected action. Use only what's in these notes and
mark unknown tools, licences, permissions and steps as questions.
Then suggest one small job I could try with the access I already have.

Why this works: the prompt asks for unknowns as questions, which is where the useful conversation with the workflow owner starts.

What you should see: a plain map, a short list of questions (often about permissions) and one suggested job.

2. Sort the map into four kinds of output

Most workplace AI work falls into one of these. Each needs a different check, and only the first is something you can usually do straight away.

Kind	Examples	Check before trusting it
Chat	Answers and drafts in ChatGPT, Claude, Copilot or Gemini	Facts and conditions against the source
Files	A spreadsheet or document from an uploaded file; Claude’s file creation, ChatGPT with uploads	Open the file; check units and a known answer
Small tools	A web page or helper made in a Claude artifact or generated as an HTML file in ChatGPT	Does it work on real input? What is only a mock-up?
Connected actions	Something that reads or writes another system: a Copilot agent, Zapier, Power Automate	What it can access, who allowed it, and proof it ran

Why this works: “the team uses AI” covers four very different things. Naming the kind tells you what question to ask.

What you should see: each thing from the demo sorted into one row, with anything you couldn’t place marked “not sure”.

3. Try one job you can check

Choose something close to your work, with a source you can compare against. A good first job: paste last sprint’s release notes into your chat.

Using only these release notes, tell me: what shipped, what is still
pending or in testing, and what decision, if any, I need to make.
Keep every condition in the notes ("if", "pending", "in test").
List anything you couldn't tell from the notes.

Why this works: you know the release well enough to spot a wrong answer, so you learn what the tool is good and bad at on safe ground.

What you should see: three short lists you can tick off against the notes in a couple of minutes.

Check the result

Everything the explanation says the workflow does is in the notes or the demo.
Permissions and connected actions you haven’t been told about are questions, not facts.
Your first job’s answer keeps every condition from the source.
Nothing suggests you have access you haven’t been given.

The usual failure: the explanation assumes you can use everything you saw. It says “you can use the team’s automation to post this” when nobody has told you what that automation is allowed to do, or calls a mock-up page a working tool. Spot it by looking for “you can” next to anything in the connected-actions row.

You said I can [action], but these notes don't say who can use it or
what it's allowed to do. Rewrite that part as a question for the
workflow owner, and suggest what I can do with my own chat instead.

[action] is the capability the answer assumed.

Worked example: a release-notes demo

Fictional example. At a lunchtime demo, Joel from engineering showed three things:

1. Release notes for sprint 41 pasted into Claude, summarised for support staff.
2. A small web page, built in a Claude artifact, that checks a CSV of customer IDs.
3. A bot that posts the summary to the #releases Slack channel every Friday.
   (Who set it up and its permissions were not mentioned.)

Lena in customer support asked her own chat to explain it. The answer ended: “You can use the Friday bot to post your own support summaries.”

The failure. Nothing in the notes says who can use the bot or what it can post. That’s a connected action with unknown permissions, presented as Lena’s to use.
The repair. The repair prompt, with “use the Friday bot”, gave: “Ask Joel: who owns the Friday bot, what can it post, and can support use it? Meanwhile, you can summarise the release notes yourself in your own chat.”
Her first job. She pasted the sprint 41 notes into the step 3 prompt and got three lists: shipped, in test, and “no decision needed from support”. She compared each list with the notes before using it.

Go deeper

Keep "allowed to", "asked to" and "did" apart

For a connected action, these are three separate facts: what it has permission to do, what someone asked it to do, and evidence that it did it. A confident summary of an automation is not evidence that the change happened. Ask for the record (the posted message, the updated row) before you rely on it.

Learn the tool you already have

Before borrowing a colleague’s setup, learn what your own account does. The beginner’s guides to ChatGPT, Claude, Copilot and Gemini each start with a first job in that product. A familiar brand name doesn’t mean every feature is switched on in your account.

Turn a first success into a small routine

If the release-notes job works, save the question rather than the answer and paste fresh notes each sprint. Save useful spreadsheet questions shows how to keep last month’s facts out of a reused prompt.