I was tired of listening to calls and training sales people to sell better over the phone – and delegated this to ChatGPT. This is what came of it

In my company, sales managers call 30–40 times a day. It seems that I have tried all sorts of solutions on how to control calls and give feedback to sales people. I’ll tell you about these solutions: how they are implemented technically, what results they produce and how much they cost.

I am developing the finishing department in a construction holding in St. Petersburg. Somehow I came across an article – We taught amoCRM to determine the negative tone of a client via ChatGPT. I then thought, it’s cool, the guys understood how to use the neuron to benefit, and talked about it in the format of a detailed guide with screenshots. I wanted to implement something similar at home.

Often, businesses implement a neural network for the sake of implementation – simply because neurons are trendy. I have many acquaintances of company directors who use neurons only to create avatars on Instagram.

I am not a supporter of this approach; I prefer the jobs-to-be-done approach, when we hire a product to perform a specific job, to solve a business problem. You must always answer the question of how much money this brought to the company or evaluate the performance in some other way.

Some calls last 5 minutes, others 40 minutes. The company has a separate expert who listens to them all and makes recommendations to managers, and then meets with them and discusses them.

This wastes a lot of sales people’s time. Let’s say the manager spent 1.5 hours talking with the client about the property. Then the expert listens to the recording for the same 1.5 hours, then sits down to write recommendations, then discusses them with the manager for another 1–2 hours. But the manager could have spent these couple of hours on additional sales. In addition, the expert gives recommendations based on his own experience, which is different for each person and, of course, limited – unlike ChatGPT.

We decided that we needed to make this process automatic. Strictly speaking, it was necessary to automate not only checking how managers speak, but also making recommendations based on shortcomings in the conversation.

Let’s say the customer said, “I’ll think about it,” but the manager did not press him to make a deal. Or he made the classic mistake of not agreeing on the date of the next contact. Then ChatGPT creates suitable material on how to lead to a purchase in such situations.

A friend of mine solved a similar problem. Here’s what he did:

• Uploaded call records from CRM to Google documents using Zapier (an American service for integrating various applications).
• The document contained a transcript of the original recording, it was uploaded to ChatGPT
• ChatGPT compiled recommendations and placed them in the same document as a separate field.

I decided that this was too confusing and tried to find a simpler solution.

We had a powerful desktop computer in our office, on which I installed the Whisper model. I also connected a telegram bot, which (as it is written in the bot itself) works on the basis of ChatGPT.

What the process looks like

• The call recording is manually sent to Whisper.
• Whisper transcribes the call.
• I manually send the text of the call to the telegram bot along with the prompt “Be a professional seller, give recommendations to the seller. Rate your sales skills on a scale of 10. Tell me what went well and what the seller did poorly.”
• I receive an assessment of the quality of the conversation and recommendations.

By the way, this is how we transcribed not only calls, but also 1.5-hour online meetings. I have some employees go to meetings with voice recorders, then upload the file into Whisper, get a transcript and read the main points that were discussed at the meeting.

How good is the result? The process was effective, but at the same time it was a bit of a crutch… It was time-consuming and annoying. I wanted the same amount of functionality, but simpler and more accessible.

What about the money? 0 rubles. The office already had a computer with an installed model, and the telegram bot was free.

We used this solution for three months.

I decided that it would be easier with my own bot. Through the API, I added ChatGPT and Dalle-3 to generate pictures, as well as computer vision, voice acting through the same Whisper and several other tools.

What the process looked like

• I send the text of the call to a telegram bot with ChatGPT and Whisper with the same prompt.
• Whisper transcribes the call
• In the same bot I enter the same prompt and attach a recording of the call
• I receive a quality assessment and recommendations.

How good is the result? The process has become shorter, and we’ve managed to automate some of the manual work—now you don’t need to go into the terminal, transcribe and transfer the text of the call to the bot.

But I wanted to automate it completely, so that I could configure everything once and it would work in one system.

What about the money?. On average $10 (939 rubles) per month purely for the API. I made the Telegram bot myself, a corporate proxy was used for the browser, so there is 0 here.

This option lasted about 5 months.

I remembered about “Team F5” and that article about ChatGPT and the negativity and went to their website. I saw a description of Voice AI – some kind of clever widget that does the same job that I delegated to the telegram bot with ChatGPT. This solution seemed more convenient to me: everything happens right in amoCRM, no need to go anywhere.

I was offered an online meeting. We asked about the task and told about another widget – Triggers, which automates processes in amoCRM. They offered solutions and chose the most suitable one, and also helped to build its logic. And all this in 1 hour.

Initially, I wanted to connect only Triggers, but the process would have been more complicated: you need to write a special script that receives the recording, sends it via API to the “brains” of OpenAI, where the recording is decrypted via ChatGPT and sent as plain text, in which the speech of the client and manager is not divided – it is unclear who is saying what. Voice AI does the same thing and marks roles at the same time. Voice AI simplifies the entire process. We managed to connect them so that the problem was solved.

What the process looks like

• Voice AI decrypts the call and immediately sends it to ChatGPT with the old prompt.
• If the score is below 8, ChatGPT writes training material based on previous recommendations. The prompt was slightly changed: “Be a professional sales trainer and methodologist, based on the recommendation, write a short training material on how to improve your sales skills.”
• The Triggers widget sets the manager the task of studying the training material.
• The manager studies the material and marks the task in amoCRM as completed.

It’s convenient that everything happens right in amoCRM:

How good is the result? The number of steps has been reduced because you don’t have to do anything with your hands. Everything happens in one place in amoCRM: managers receive a task with training material, they watch it and close the tasks, I see the number of mistakes made and tasks completed and I can collect statistics on how successfully he completes these tasks.

Another useful thing: Voice AI allows you to download scripts and other materials used by the sales department.

What about the money? 2,910 rubles/month for setting tasks through Triggers and 20 rubles for processing 1 call in Voice AI. In total this is ≈ 3500 rubles/month.

We have been using this solution for a month now.

Saves sales time and at the same time teaches you to learn independently. They receive text recommendations, follow them themselves, apply them in their work and report back.

Saves my time. In the first two solutions I did a lot with my hands, now I don’t do anything. I just look at analytics on completed tasks for managers.

Recommendations based on the knowledge base. This makes the recommendations more accurate, because they are based on our regulations, and not just information from the Internet.

Lower costs. With ChatGPT, you had to pay $20 per month for requests, the number of which is also limited, and about another $10 per month for the API, i.e. total $30, 2817 rubles/month. An individual specialist received 100 thousand per month. And here it turns out to be 3,500–3,700 rubles per month – more expensive than the first option, but it frees up a lot of time for me and my employees.

Next, I will teach the neuron to give even more relevant recommendations: for example, so that it determines the client’s psychotype and gives the manager recommendations on how to communicate with this or that psychotype.

Tell us about your experience in implementing neural networks in business processes. What did they implement? What problem were you solving?

{“entityType”:”post”,”entityId”:1139443,”awards”:[]}