Customer support · Production

Lindy AI replaces open prompts with structured on-rails workflows to make AI agents reliable

The problem

Lindy 1.0 used a giant prompt field and a collection of tools, leaving it to the LLM to decide when to invoke each tool—making workflows unpredictable and unreliable. Text-based configuration was also opaque to non-technical users, with 60 to 70% of user-typed prompts being unintelligible.

First attempt

The prompt-driven Lindy 1.0 approach could not guarantee that required workflow steps—such as consulting a knowledge base—would always execute; the LLM could skip them entirely.

Workflow diagram · grounded in source
1
Meeting begins: Lindy records
trigger
“when a meeting begins, you record the meeting”
2
AI generates coaching notes
ai_action
“my Lindy is constantly coaching me. And so you can see here in the prompt of the coaching notes, I've told it, hey, you know, was I unnecessarily confrontational at any point? I'm French, so I have to watch out for that. Or not confronta…”
3
Summary and coaching notes emailed
output
“after the meeting, you send me a summary and you send me coaching notes”
4
Notes disseminated to Slack
integration
“it goes on Slack. It disseminates the notes on Slack”
5
User reply restores meeting context
feedback_loop
“it's actually able to backtrack and resume the automation at the coaching notes email if I responded to that email”
Reported outcome

Lindy 2.0's on-rails visual workflow builder made agents reliably execute mandatory steps and enabled a wave of new use cases; the founder now skips most meetings and conducts a five-minute Q&A with Lindy rather than attending a 60-minute meeting.

Reported metrics
meeting time replaced by async Q&Ainstead of going to like a 60-minute meeting, I have like a five-minute chat
User prompts that are unintelligible60 or 70%
agent reliability after Lindy 2.0way more reliable, way easier to set up
new use cases from Lindy 2.0a ton of new use cases pop up
Reported stack
ZendeskSlackGPT-4 TurboGoogle DocYouTubeRAGGoogle Workspace
Source
https://www.latent.space/p/lindy
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Lindy 2.0's on-rails visual workflow builder made agents reliably execute mandatory steps and enabled a wave of new use cases; the founder now skips most meetings and conducts a five-minute Q&A with Lindy rather than…

What tools did this team use?

Zendesk, Slack, GPT-4 Turbo, Google Doc, YouTube, RAG, Google Workspace.

What results were reported?

meeting time replaced by async Q&A: instead of going to like a 60-minute meeting, I have like a five-minute chat; User prompts that are unintelligible: 60 or 70%; agent reliability after Lindy 2.0: way more reliable, way easier to set up; new use cases from Lindy 2.0: a ton of new use cases pop up (source-reported, not independently verified).

What failed first in this deployment?

The prompt-driven Lindy 1.0 approach could not guarantee that required workflow steps—such as consulting a knowledge base—would always execute; the LLM could skip them entirely.

How is this customer support AI workflow structured?

Meeting begins: Lindy records → AI generates coaching notes → Summary and coaching notes emailed → Notes disseminated to Slack → User reply restores meeting context.