How to boost a conversion rate using a neural network


80 operators, 40,000 calls, 100% control, +3% of revenue and zero additional supervisors: how to boost a call center with a neural network

A case study on how to use neural networks to fully automate quality control to drive sales conversion.

If your company has operators, you know you can't control everyone: even 10% of the calls would take a lot of supervisor time. Add to that 80 people, 40,000 calls a month, and personal relationships between employees, and you get a poorly managed machine with jumping efficiency.

This case study is about how a network of medical centers learned to control the efficiency of this machine 100% with

Customer's Team

The customer is a Network of medical facilities. Details are still under NDA. The developer is Our service understands the context of the conversation, evaluates operators' dialogues and uploads analytics to managers.

What's important

This case is not about an enterprise client with a lot of data and a lot of money for experiments. It's about an average business with a data-driven approach that counts money and requires a 100% return on investment. Read below how we got an ROI of 328% already in the pilot phase.

Start Point

Personal relationships between operators and supervisors make it difficult to assess the quality of calls, only 0.5% of conversations are monitored, the spread of conversions - from 10 to 20%.

We started working with the medical center in July 2022. At that time, their team had 80 operators, who took 40,000 calls a month, and five supervisors to supervise.

The operator's script consisted of two parts: a regulated part and a framework. The regulated part prescribed specific questions: for example, when making an appointment for an MRI, it was obligatory to specify the weight. The framework consisted of stages: gather anamnesis, consult, handle objections, tell about the advantages of the equipment... Specific wording depended on the context of the conversation and the operator's manner of speaking. The client did not use rigid scripts because they reduced the conversion rate.


The more primary visits to the clinic, the more the medical center will sell high-margin services.

If we increase the number of visits and sales of marginal services, profits will begin to grow non-linearly.

Conversion from call to visit depends on the operator. If the operator passes the 10 stages of the framework, conversion will increase. It is important not to turn a person into a machine that reads out questions according to a rigid script: humanity, rapport, empathy are the medical center's competitive advantages.


  • Listening and training 80 operators is difficult and expensive. One supervisor costs about $100,000. The entire supervisory department costs $500,000 annually.
  • Supervisors have time to listen to and evaluate 0.5% of the recordings.
  • There is no dashboard with error statistics, which means that it is impossible to track where all managers or any one of them make mistakes.
  • The conversion from a call to an appointment jumps from month to month. Revenues, too, accordingly.

A system was required that automatically evaluates all calls and gives an objective assessment of all conversations.

What you wanted to get from

Understand how operators work. We wanted to analyze each call by 10, and then by 20 parameters. The call should be evaluated within and give metadata to DWH. If an operator says hello according to medical center standards, but doesn't collect anamnesis, then the first stage of the framework is passed by 1, and the second stage by zero. The point of data collection is to see and correct employee errors immediately after the conversation or at the end of the day.

With the help of control and training, reduce errors. If a patient is signed up in the wrong place, the doctor's time is wasted. For example, there is only one MRI machine that can take a patient weighing more than 120 kg. Every time the operator forgets to specify the weight, the schedule gets clogged with irrelevant requests and the doctor's time is wasted.

Save money on controls. Reduce the time spent listening to calls and do not inflate the staff in the quality control department.

Lay the foundation. Create a minimal product on which to build the necessary features, such as online prompts to operators during a conversation with a customer or autofill the card in the CRM at the end of the call.

The global goal of the project is to affect conversion rates, and not to waste money on ineffective advertising and additional controllers.


In the POC version, the folks at the medical center wanted to see if the operators went through the important 10 steps for conversion.

How we determined what the operator was saying

At each stage of the framework, operators say similar phrases. To make understand these phrases, we have assigned labels to them. For example, "Hello, you called the medical center, my name is Michele," is the "Greeting" label. "Signed you up for November 22," is the "Patient Record" label. "We're having a <some> promotion," is the "Promo" label. The task of the labels is to track which stages the operator forgets to go through, and how this affects conversion.

There is no labeled data for such a business task, so we did the markup ourselves. So we divided the operator's phrases into 10 labels:

  1. introduced himself and recognized the name
  2. collected anamnesis
  3. consulted before the visit
  4. Handled the objections
  5. made a presentation on medical equipment
  6. named contraindications
  7. offered additional services
  8. made an appointment
  9. said about the cost of the service
  10. Spelled out the details of the appointment

The operator asked the obligatory question at the stage of collecting contraindications to the appointment for a regular MRI machine, identified the label "Contraindications" with a probability of 0.939.

How integration with PSTN works

Simplified scheme of integration with PSTN

Simplified scheme of integration with PSTN


Briefly about how understands context

Let's say the operator said about contraindications: "Before the abdominal MRI, don't eat anything for six hours, i.e. from 9 a.m.". To understand this phrase, doesn't just search by keywords, but remembers and takes into account previous phrases in the conversation.

Only 100 calls were enough for us to learn

Normally, for a neural network to learn how to accurately identify labels, you need to mark up thousands of calls and feed them to the neural network. By using the new technology, we only needed 100 good calls to train the first 10 labels.

During the markup process, we had to assess the quality of the company's in-house supervisors. Our markup technicians drew the customer's attention to the discrepancy between the actual labels and the estimates on the checklist. That is, the label was not actually in the conversation, but it was on the supervisor's checklist. When the team of MFN promoted this story, it turned out that there was an informal relationship between some supervisors and operators, as a result, they were understandably overrated, which affects their motivation... They accidentally found another advantage of neural networks over humans)

People and Time

It took two months to develop, train and debug. Testing took another month. Five to ten people from our side and three from the customer's side worked on the project.


Medical center results after three months

  1. The medical center now has a system that understands the context of the conversation and evaluates all dialogues. Previously supervisors selectively monitored a maximum of 0.5% of calls, but now they monitor 100% of conversations. Data on each operator with different filters are displayed on the dashboard and enable the medical center to make management decisions quickly.
  2. Before the implementation, conversion to an appointment jumped from 10% to 20%; now it stays at 20%. For the client, each percentage means millions USD per month. Our next goal is to raise the conversion rate to 25% in three months.
  3. During the pilot, revenue increased by 3% and ROI was 328%.
  4. There was a fear that employees would start complaining about the innovation, but this did not happen. The first reason is the accuracy of phrase detection. The second is that the medical center uses analytics to educate the operator, not to punish it.
  5. Right now the medical center has more than 10 sites with end-to-end analytics. When we combine traffic sources, conversions and call analytics, we can see the advertising channels that are not making money and manage the user journey more effectively based on the data.
  6. When we implemented the labels, we saw that the supervisor ticked off the checklist just for fun. For example, the operator didn't talk about the promotion, but the supervisor checked that the "promo" stage was passed. When the team of MFN dug deeper, it turned out that there was an informal relationship between some supervisors and operators, which overstated the motivation. So we found another advantage of and neural networks in general.

What's next

In the next iterations we will add real-time prompts for operators and teach to determine a client's mood. In general, we will do everything we can to increase conversion rates and make decisions on the basis of the data.

Get the FREE Demo of


Similar posts

Get notified on new marketing insights

Be the first to know about new B2B SaaS Sales insights to increase the conversion rate with the tools and knowledge of today’s industry.