AI Process Mapping With Mermaid: As-Is to To-Be

To map a process with AI, give ChatGPT or Claude your interview notes and ask for a Mermaid flowchart of the as-is process, with who does each step and every decision point. Read the draft against what people told you, then paste it into CardBoard, where it becomes a process map you can drag, annotate, and share with the people who do the work. Duplicate that board to design the to-be, and keep the as-is as your baseline.

The draft takes a minute. What it can’t do is know your organization. The rest of this guide is about getting a draft worth correcting, and the part of the work that’s still yours.

Why business analysts are doing this now

Mermaid is a text format for diagrams, and it’s what language models write when you ask them for one. BA trainers have picked it up: Adaptive US teaches custom GPTs that generate Mermaid for business analysis diagrams, and the IIBA has run sessions on AI process mapping. The appeal is plain. An as-is map used to take an afternoon of redrawing notes into boxes. Now the boxes are free, and the time goes into checking them.

Step 1: Give the model your raw material

A model maps what you give it. Asked for “an invoice approval process,” it invents a tidy one that belongs to no company. Given the notes from your interviews, it maps yours, gaps included. A prompt that works:

Here are my notes from interviews with accounts payable, purchasing, and a manager about how invoices get approved today. Write a Mermaid flowchart of the as-is process, left to right. Start each step with who does it, like “AP clerk: key invoice into ERP”. Make every decision a diamond with labeled Yes and No branches. Include the rework loops people described. Mark the steps people complained about with a class named pain. Don’t add steps that aren’t in the notes, and list anything unclear as a question after the diagram.

What comes back looks like this:

flowchart LR
  A[AP clerk: receive invoice by email] --> B[AP clerk: key invoice into ERP]
  B --> C{PO number on invoice?}
  C -->|No| D[AP clerk: email requester for PO]
  D --> E[Requester: reply with PO]
  E --> B
  C -->|Yes| F[AP clerk: match to PO and receipt]
  F --> G{Amounts match?}
  G -->|No| H[AP clerk: email buyer about variance]
  H --> I[Buyer: approve variance]
  I --> J
  G -->|Yes| J[Manager: approve in ERP]
  J --> K[Treasury: schedule payment]
  classDef pain fill:#ff9384
  class B,D,H pain

Two choices in that prompt matter later. Putting the role at the start of each step keeps ownership visible, because CardBoard doesn’t draw Mermaid subgraph swimlanes. And the last line, asking for questions instead of guesses, is where the model tells you what your interviews missed.

Step 2: Read it for what the model can’t know

The draft is plausible, and that’s the risk. A model fills gaps with the common version of a process, so check every step against a named person:

• Is each step something someone actually does, or something the model assumed?
• Does every decision have both branches? Models forget the unhappy path.
• Are the rework loops there? In the example, the missing-PO loop sends the invoice back to be keyed again. That loop is usually where the time goes.
• Did it merge two roles into one? “Approver” is often a manager and a buyer.

Fix what you can in the text, or ask the model for a corrected version with your list of changes. How to make diagrams with AI covers the editing loop and the syntax errors you’ll hit.

Step 3: Paste it into a board

In CardBoard, create a board and choose Paste Mermaid. A flowchart becomes a process map:

Now it’s a working surface instead of a picture. Put the detail that doesn’t fit in a box, like volumes, systems, and handoff delays, in each card’s description. Invite the people you interviewed as guests, free on every plan, and let them comment on the step they own. The corrections they make are the ones the model could never have made.

Step 4: Design the to-be

Keep the as-is board as it is. It’s your baseline, and the argument for every change. From the Create board page, choose Duplicate a board, pick the as-is, and rename the copy to-be.

Then work the red cards first. In the example, keying invoices by hand and chasing PO numbers by email are the two loops that generate rework. A to-be might capture invoices from a supplier portal that rejects any invoice without a PO, which removes a step, a decision, and a loop.

An AI can help here too, differently. Connect Claude to CardBoard over MCP and ask it to propose to-be changes on the copy: steps to remove, a decision to automate, a new check. Each change arrives as a proposal on the board, and you accept or dismiss it. The model suggests. The analyst decides, with the stakeholders in the room.

From to-be to delivery

A to-be process map says how the work should flow. It doesn’t say what to build first. When the to-be needs software, the next map is a user story map: the to-be steps become the backbone, the changes become stories under them, and a release slice says which changes ship first. If your AI wrote the steps as a Mermaid journey, that imports as a story map directly.

What this doesn’t do

• No swimlanes. Mermaid subgraph blocks are ignored on import, so keep roles in the step names.
• No BPMN. CardBoard imports Mermaid flowcharts, not BPMN files, and doesn’t export either.
• One path per sequence diagram. A Mermaid sequenceDiagram imports as a single path of messages. Diagrams with alt or loop blocks are refused.
• No export back to Mermaid. Once the process is a board, the board is the copy you keep.

Frequently asked questions

Can ChatGPT or Claude create a process map?

Yes. Ask for a Mermaid flowchart of the process and either one returns a diagram as text, which renders as a flowchart. Give it your interview notes rather than a topic, or it will map a generic process instead of yours.

How do I make an as-is process map with AI?

Paste your interview or workshop notes into the model and ask for a Mermaid flowchart of the current process, with the role at the start of each step, every decision as a labeled branch, and the rework loops people described. Then check each step against the person who does it.

What’s the difference between an as-is and a to-be process map?

An as-is map shows how the work happens today, including the workarounds. A to-be map shows how it should happen after the change. Keep both: the as-is is the baseline that explains why each change in the to-be is worth making.

Can I edit an AI-generated process map without regenerating it?

Yes, if you import it somewhere built around the structure. Pasted into CardBoard, each step becomes a card you drag and rename, and an AI connected over MCP proposes changes to that same board instead of drawing a new diagram.

Does CardBoard support swimlanes from Mermaid?

Not from Mermaid. subgraph blocks are ignored on import, so put the role at the start of each step’s name, which keeps ownership readable on the map.

Written by Josh Colter, who owns and builds CardBoard, the story mapping tool featured in Jeff Patton’s User Story Mapping.