Related Apps

Advanced Call Transcripts & SummariesAdvanced Call Transcripts for WorkflowAdvanced Sub-Account Analytics

Description

Learn how to set up a workflow using ChatGPT that analyzes phone calls to figure out if a lead booked or not.

Status

Completed

Overview

This workflow requires that you use Advanced Call Transcripts & Summaries in order to get the finished, formatted call transcript.

Once you have this transcript in your workflows, you can basically do anything with it, as this guide will explain.

Here’s a video walking you through how this works:

Install Snapshot

You can get this workflow from a snapshot. If you already have the app, you can get it from the app documentation page.

Otherwise, once you install the app you’ll be given instructions on how to install the snapshot.

Chain of Thought

This workflow uses my trademarked Double Prompting Algorithm™.

To enable branching and taking different actions based on what ChatGPT responds with, we need it to output a certain word or phrase that we can filter based on.

But each time we need ChatGPT to spit out a specific word or phrase, we prompt it twice.

The first time has it argue for or against a certain determination, and the second time reads that first arguments and produces our single-word ouput.

This is similar to the way that ChatGPT o1 and other similar reasoning models work.

In order to get a more accurate response, we allow ChatGPT to generate more tokens at first to explain its reasoning.

If we just asked it for a single-word response right off the bat, it would be highly innacurate.

This is how you and I think, since if we have a complex problem to solve, we actually work through each step–taking notes, writing things down, and talking to ourselves until we arrive at a solution.

This system isn’t recursive in the same way that ChatGPT o1 or more advanced reasoning models might be, but it does allow for better evaluations.

Debugging Notes

After each ChatGPT prompt in this workflow we have a dubugging note like so:

This is useful while you’re getting the workflow dialed in and testing it on a single sub-account.

For each completed phone call, you can look to that contact’s notes section to exactly what ChatGPT was thinking about, and what decision it made for each step of the process.

If you see the same note multiple times, that would be a case where the prompt is being retried since the determination was incorrect.

Once you get the workflow dialed in so it’s reliable, you can remove these steps so you’re not filling up the notes section all the time.

Prompt Retries

One of the (perhaps most confusing) features of this workflow is the automated prompt retries.

This can be necessary however, since sometimes even when you tell ChatGPT “Do not output any text besides BOOKED or OTHER”, it’ll give you “BOOKED”, “The response should be BOOKED”, etc.

These responses aren’t reliably filtered for, so it can make sense to retry the prompt if the response isn’t suitable.

To retry the prompt, all we do is wait a few minutes, then send contact back up to the start of the workflow to try again.

You could reduce this wait time to just a few seconds or remove it entirely, but once the workflow runs a few times it’ll automatically remove the contact once it detects a loop. In this case, that would be a good thing since if there’s an issue you don’t want this to just run indefinitely.

Associated Costs

The heaviest cost incurred from this operation (besides native call transcripts) will be the ChatGPT native action.

This action is billed at $0.03/action, so with 4 actions (or more) running, you’re up to $0.12 per run. Although, with the snapshot we’re using ChatGPT 4o-mini which the rebilling page doesn’t mention, but it could be charged at the cost of ChatGPT 3.5 since it’s very inexpensive.

This may be worth it to you, but in reality ChatGPT costs about 1/10 of that (or $0.003/response) so I may consider building a 💬Custom AI Workflow Action so you can use your own API key and cut down on this cost.

Frequent Questions