When we said an AI assistant was coming, the obvious version was a chatbot that spits out Five9 XML. We didn’t build that, because the obvious version is dangerous: a model that confidently invents a <moduleId> or drops a branch can quietly break a live call flow, and you’d find out from a customer, not a diff.

So the assistant we shipped works the way a careful engineer does. It reads before it writes, it proposes instead of applies, and it checks its work against your tests.

It reads your flow with tools, not vibes

Ask “are there any dead branches?” and the assistant doesn’t guess from a blob of pasted XML — it calls find_dead_branches against your actual flow. The same goes for find_nodes, read_node, read_flow_summary, read_bundle_summary, and find_unused_prompts. These tools run in your browser, against the IR that’s already loaded. The backend is a thin proxy; your bundle stays local. You can watch each tool call appear in the transcript as the assistant works.

It proposes — you approve

When you ask for a change (“add a high-volume announcement before every Spanish skill transfer”), the assistant doesn’t touch your flow. It stages a set of edits. Each one is validated against the IR schema the moment it’s staged, so a malformed change is rejected on the spot rather than committed and discovered later.

Then you get a review screen: a node-level diff with a checkbox on every change. Approve all of them, approve some, or discard the lot. Only the changes you tick are applied, and they land as a single undoable step.

It re-runs your simulations before you commit

This is the part we’re proudest of. ivrloom’s call simulator is also its test framework: you can save scenarios — “press 2, after hours” — and replay them. When the assistant proposes a change, we replay every saved scenario for that script against the proposed flow and compare it to today’s behavior. If a call that used to reach the Spanish queue now hits a dead end, you see it in red, in the review, before you apply.

A diff tells you what changed. The regression check tells you whether it matters.

You’re in control of the spend, and the privacy

Every plan has a monthly token budget. There’s a usage meter in your account, a soft alert at 80%, and a hard stop at 100% — no surprise bills. And if you’d rather keep everything local, one toggle in settings turns the assistant off completely; with it off, nothing is ever sent to our AI provider.

Try it

Drop a .zip, open the AI tab, and start with one of the suggested prompts — “What does this flow do?”, “Are there any dead branches or unused prompts?”, or select a node and hit ⌘/Ctrl+E to ask the assistant to explain it.

If you’re a Five9 admin and you put this through its paces, we want to hear what broke and what surprised you: [email protected].