Think of workflows
as systems
Any workflow can be automated at some point. The question is how much the AI should do on its own today — step by step, with the guard each step needs.
- Use case
- Workflow
- Data
- Constraints
A use case has a workflow, the workflow has data, and the data has constraints. AI automation and inference are integrated based on the specifics of each use case, workflow and data.
Which use cases with which workflow
Most use cases are data-management use cases. Finding data, moving it, checking it, reporting on it. So every step starts from the data.
Done when you can say in one sentence what the AI is hired to do, today's cost is measured, and the result that would make you stop is written down.
How will the AI handle each step — and where must a person stay?
Each step's conditions decide the AI's role. Decide the shape of the system, who approves, the systems it touches and what it may change, the words it is given, and how you will know it works.
Done when every step has a role and a guard, every system an access path, and a test set says what “right” means.
Build it, prove it on real cases, then go live
Build against non-production data, run the test set on every change, test the protections rather than argue them, and make sure the people who run it know what to do when it fails.
Done when it passes its test set, nothing irreversible happens without a person, and someone is on call with a way to switch it off.
What workflow do you want to automate?
Map your workflow · 3 steps
- Start with what happens first — usually reading or gathering something.
- Write each step as an action — “Check the amount”, not “amount checking”.
- Say who does it today — a person, a team or a system.
- Say where — the system or tool it happens in.
- Name the data — what the step uses or produces.
- Mark the data — confidential, critical, or can't be undone.
- Add the next step — until you reach what it ends with.
Talking to Claude
You have the use case and the workflow. Now bring them to your Claude — one conversation per workflow, one step at a time.
- Give it the whole picture first — the job, what starts and ends it, and every step: what happens, who does it, where, and the data.
- Ask it to say it back — before any work, have it repeat the workflow in its own words. Fix anything it got wrong.
- One step at a time, with a real case — start with a step the AI can run today. Paste one real example with anything private taken out.
- It suggests, you decide — it shows what it would do; it does not change, send, delete or save anything.
- Say what was wrong, not only that it was — add one instruction, try the same example again, compare. Keep what works in the project instructions.
Try AI on your own work
Your first time using AI on your own work? Start here.
Seven small steps, done with an AI you already have — no technology or coding needed. The FDS workflow agent writes exactly what to paste, and you tell it what happened.
- Talk about your work — tell the AI what you do, the way you would explain it to a new colleague.
- See what it understood — ask it to repeat your work back. Fix anything it got wrong.
- Try one small thing — the easiest part of your work, with one real example.
- Look at the result — is it right? Is anything missing? Would you use it as it is?
- Make it better — add or remove one instruction, try the same example again, compare.
- Try something new — a harder case, or another part of your work.
- Put it together — line up what worked into your workflow. That is where Prototyping begins.
Steps 3 to 5 repeat as often as you like: try it, check it, improve it.
Nothing you type here leaves your browser. Your real work happens in your own Claude.
How the steps and conditions become a decision
Every workflow moves data — in from somewhere, changed at each step, out to someone. Every step is something done with that data: gathering, consulting, checking, analysing, deciding, presenting. Three conditions on the data set how far the AI goes on each one.
| Step's conditions | AI's role today | Guard that applies |
|---|
Confidential decides where the AI runs and who sees the data — not whether it can help. And “today” is not forever: a step moves down a level when the evidence says so.
The FDS workflow agent walks you through it
One step at a time, on your workflow, with every guide in reach and nothing skipped quietly.