I recently conducted an experiment as part of writing an article about AI-assisted process mapping and thought I’d post the results here.
I’ve been investigating using AI to turn narrative text into process maps and many of the roads I’ve followed have led to Lucidchart!
I decided to try narrative-to-map three different ways:
- Claude to Lucidchart via the MCP connection
- Lucidchart AI without the Process Agent
- Lucidchart with the Process Agent
Here’s the narrative description of an HR recruiting process.
This is the prompt I used in each case:
You're a business analyst with particular skill at capturing knowledge about and then mapping business processes. Please create a flowchart in Lucidchart for the process described in the attached PDF.
Note that I only included “in Lucidchart” when I was working in Claude; clearly that wasn’t necessary when I was in the Lucidchart app!
Case 1: Claude using Lucidchart MCP connection
Claude used the MCP to create this diagram. It’s an OK start but has several issues including 1) there are no arrowheads on the connector lines to show flow direction, and 2) the “No” path is missing from the decision diamond. (Actually, it’s not 100% missing: the word “No” is buried in the End shape for some reason, but there isn’t a connector line.)
When I pointed out those shortcomings, Claude created this version of the diagram. There are still no arrowheads, but the No path is present and correct.
NOTE: I manually increased the font size on each shape to improve readability for a screenshot; everything else is as Claude created it.
Case 2: Lucidchart AI without the Process Agent
The results using Lucidchart’s built-in AI capabilities were much more impressive. There’s an additional decision point, which is valid, and there are arrowheads to show direction.
Minor quibble: I asked for a flowchart and Lucidchart produced a swimlane diagram. This isn’t wrong – the latter is a superset of the former – but each format has a place and I would have liked to see a flowchart version.
Less minor quibble: The step just before the end of the flow is clearly labelled “subprocess” yet Lucidchart used the Process shape and not the Subprocess shape. I suspect that’s because the appropriate shape in the library carries the official designation of “Predefined Process.” In the process mapping world, “Subprocess” is such a common synonym for predefined process that I would suggest adding that knowledge to the AI drawing agent.
Note: I made minor changes to the layout, principally to reduce the overall width for screenshots.
Final point: As a process analyst, I would rearrange a few areas of the map and would add a separate swimlane for “Candidate” to contain the “Candidate accepts offer?” decision. However, this is a solid B+/A- result.
Case 3: Lucidchart AI with the Process Agent
The Process Agent also created a swimlane diagram. Interestingly, it’s different from the one created without using the agent. I suspect part of that is the non-deterministic nature of LLMs and AI agents but I’m sure it also relates to differences in the models used for each.
Observations: the lane assignments for some tasks are different from Case 2; there is only one decision, not two; The Yes/No outcomes don’t use the Lucidchart color and line type conventions.
Other than that, I really like the recently added feature in the Process Agent that creates a Context frame (upper left), and both short (upper left) and long (right) process summary panels. Bonus points for adding a shape key!
Summary
Based on my experience so far, I’ll stick to working inside Lucidchart rather than starting with Claude or Copilot when I want to create process maps. If you your experience has been different, I’d love to hear about it -- please comment below.
I’ll be working with longer and more detailed narratives in the future and would love to see your results if you create maps from your own process narratives.
Your privacy choices
