diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
index a453dd1..52932ae 100644
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -19,6 +19,7 @@ reverse.
 
 ## 2026-08-25
 
+- An AI integration recorded as an idea in the plan — edit the tree in a dialogue, with your own API key and `llms.md` as the model's guide
 - A warning flags the bottleneck when one `@name` carries more than half of the open work on the cheapest path — their tag pills turn amber on the open nodes
 - The estimated size of an unsized node now prices the remaining work — done sub-packages drop out, and a subtree with everything named done counts as XS
 - The fold chip is solid white with an ink glyph now — the translucent box vanished on the pastel node colours
diff --git a/docs/examples/werkbaum.werkbaum b/docs/examples/werkbaum.werkbaum
index 1af3d89..b1828ca 100644
--- a/docs/examples/werkbaum.werkbaum
+++ b/docs/examples/werkbaum.werkbaum
@@ -196,6 +196,10 @@
       - [-] #trk.plugin: A plugin inside Taiga itself (L)  %% legacy frontend, self-hosted only
     = [?] #trk.tenzu: Tenzu, the successor (XXL) https://tenzu.net
     = [ ] #trk.none: No tracker, plain text is enough (XS)
+  - [?] #ai: AI integration (XL)
+    - [?] #ai.dialog: Edit the plan in a dialogue with the AI (L)
+    - [?] #ai.key: Bring your own API key, no server in between (M)
+    - [?] #ai.ground: llms.md grounds the model in the notation (S)
 
 ---
 #wb
@@ -1034,3 +1038,24 @@
 #trk.none
   The honest alternative: a plan in a text file may simply be enough. It is
   the cheapest node in this group, and the cheapest path duly picks it.
+
+#ai
+  Connect an LLM or an agentic AI to the editor, so the plan can be shaped in
+  conversation instead of by hand alone. An idea for now: the pieces exist
+  (the text is the whole state, and llms.md already teaches agents to read
+  and write it), but nothing here is designed yet.
+
+#ai.dialog
+  A chat beside the diagram: ask for a decomposition, a size estimate or a
+  restructuring, and the AI answers by editing the notation text — every
+  change lands as an undoable text edit, with the human deciding what stays.
+
+#ai.key
+  The user brings their own API key and the browser talks directly to the
+  chosen provider — no Werkbaum server in between, following the D20/D23
+  line: requests only at the user's request, only to the host they configure.
+
+#ai.ground
+  The system prompt is essentially llms.md (D43): the agent guide already
+  explains the notation for reading and writing, so the model needs no
+  training — just the document and the request.