The AI panel (⌘J or Ctrl J) models as you talk, then checks and improves what you have. Here is what to ask, and when. Copy a prompt, change the words in [brackets], and send it.
Three ways to work
Interview: you (or a subject-matter expert) describe the work, and the AI draws it as you go and asks what it needs to know. Start here for a new process.
Feedback: click an element, type a comment, click the next one, and so on; then send them all and the AI applies them together. Best for a reviewer's list of changes.
Review: ask the AI to check the model or make a change. It works through edge cases one at a time, compares the model with the processes around it, and says what it changed and what is still open.
Every change the AI makes appears on the diagram straight away and can be undone turn by turn. Anything it guessed is marked with a ? until someone confirms it.
Getting a first draft
Let the AI lead
Interview
Let's start with a high-level overview.
The same as Interview me on an empty model. It asks about purpose, who is involved, what starts the process and what counts as done, then sketches the main stages.
Draft from what you know, refine after
Interview
Draft an initial version of [process name] from this description. Keep to the main path for now; we will do exceptions next.
[Paste a few paragraphs: who does what, in what order, using which systems.]
A rough draft in one go beats ten questions when you already know the process. Then ask for the exceptions.
Turn an existing procedure into a model
Interview
Here is our current procedure for [process name]. Model it as a flow: the roles become lanes, the systems go on the steps as applications, and any checks become controls. List anything that is unclear instead of guessing.
[Paste the procedure text.]
Paste text, not customer data. Long documents work best a section at a time.
Checking it
Compare with industry practice
Review
Compare this process with common practice for [process, e.g. SME loan origination] at a [bank / insurer / utility], for example the BIAN service domains and the APQC process classification framework. Which steps, decisions or controls are usually there that we don't have? List them with a reason each; don't change the model yet.
A checklist to discuss with the people who do the work, not a verdict: they decide what applies.
Find risk and control gaps
Review
Go through each risk on this process. Where is a risk not covered by a control at the same step, or by a control in our Controls register? For each gap, suggest a control and whether it is preventive or detective.
Twiggy already flags a risk with no control at its step; this goes further and proposes what to add. Controls the AI adds are defined in the model: add the keepers to the Controls register so every process can use them.
Check against an obligation
Review
Which steps in this process help us meet [obligation, e.g. customer identification under the AML/CTF Rules]? Link the matching requirements from our Requirements register, and tell me where the process shows no evidence that the obligation is met.
The AI knows public rules and standards in general terms. Treat its answer as a starting point for your compliance team, not as advice.
Work through what can go wrong
Review
Ask me what happens when things go wrong at each decision and risky step: information missing or wrong, an approval refused, a deadline missed, a system down, the customer withdrawing. One question at a time.
Each answer becomes a rule, an exception path, a timer, or a risk and control, whichever fits.
Check it fits with its neighbours
Review
Check this flow against the processes before and after it and its siblings: do the hand-off events match, and are roles, systems and controls named the same way?
Tidy the names
Review
Check the names: steps as verb and noun (Verify identity), events as outcomes (Application received), decisions as questions, and lanes matching our Roles register. Fix what's clearly wrong; list the rest.
Improving it
Find waste
Review
Where are the hand-offs, waiting, duplicated checks and rework loops in this process? Rank them by likely impact, and say what information you'd need to be sure.
Sketch a simpler future state
Review
Suggest a simpler to-be version of this process that keeps every control. Describe the changes as a list first; once I agree, apply them.
Keep the as-is: make the to-be a variant of it (a new model with variant_of=), so both can be compared and signed off separately.
Spot automation candidates
Review
Which steps could be automated or assisted by a system? For each, say what it would need (data, rules, an integration) and what could go wrong.
Using it
Explain it to a new starter
Review
Explain this process to a new starter in [team] in plain English: what it's for, their part in it step by step, what they must check at each step and why, and who they hand over to. Don't change the model.
Then walk them through it with Present, step by step.
Brief an approver
Review
Summarise this process for the person signing it off, in five bullet points: its purpose, who is involved, the key risks and controls, and anything still open.
Draft work instructions
Review
Write work instructions for the [role] lane: each of their steps as a numbered instruction, with the system they use and the checks they must make. Don't change the model.
Good habits
One thing at a time. Small, specific requests give better changes than “fix everything”.
Give it context. Who the process is for, the product or channel, and any rule that matters. The AI knows common practice, not your policies, unless they are in the model or your registers.
Check its guesses. Anything marked ? is an assumption: confirm it with the people who do the work, or correct it.
Undo freely. Every AI turn can be undone on its own, and every model keeps its version history.
Park what can wait. Notes for later go to the parked list instead of cluttering the model.
Keep personal data out. Describe the work, not real customers. The AI only sees what you can see, and you can switch it off for the workspace.
Regulation needs people. Use the AI to find gaps and draft, and your risk and compliance teams to decide.
