A poorly tuned machine
If your company has its own call center, I don’t need to tell you how difficult, almost impossible, it is to monitor the efficiency of operators. After spending hours of his time, the supervisor will be able to listen to only a small part of the recorded conversations, and if, as was the case in the company of one of our clients, you have 70 operators who take up to 40,000 calls per month, you can forget about strict control. Essentially, it was a poorly tuned machine that stuck on corners and had questionable levels of efficiency.
We had good experience in healthcare call center outsourcing, but here the client did not plan to outsource it. Moreover, we were not talking about a giant company, which had no need to save on experiments for the sake of profit in the long term, but about an ordinary representative of a mid-level business, relying on the fact that every investment – in this case, in the help of our agency – will pay off as soon as possible. I proudly admit that his expectations were met: already at the pilot stage we received Return on Investment of 328%. Now I'll tell you how. Only without specific names, since no one canceled the non-disclosure agreement.
Features of the script
The script was built according to principles that were well known to us, as, indeed, to any healthcare call center outsourcing companies, and consisted of a clearly defined part for each specific case – for example, the operator was obliged to find out the patient’s weight when registering for an MRI examination – and a less regulated framework, when they collected anamnesis from the patient, answered questions he had, and found counterarguments to objections. Experience has shown that strict control in this part led to a decrease in conversion, so operators were allowed to formulate phrases as they saw fit.
Ideas
The medical center will be able to sell more services with high margins if it achieves an increase in the number of initial customer visits, which, in turn, will provide a significant increase in profits. Moreover, all this should be carried out by the company’s call center with outsourced technical support on our part.
The conversion from a call from a potential client to a visit to a medical center depends on the operator’s ability to competently build a conversation. At the same time, it is important that the operator scrupulously goes through all the points of the framework, but does not speak in memorized phrases of the script. Among his mandatory qualities should be empathy, the ability to build trusting relationships during a short conversation, and humanity.
Obstacles
5 supervisors are physically unable to listen to recordings of conversations of 70 operators. Consequently:
- There are no detailed statistics with common operator errors;
- There is no opportunity to correct and train them;
- Operators either spoke strictly according to scripts, turning into a kind of ecommerce answering service, or moved far from the framework.
As a result, call-to-record conversion remained unstable and showed different results in different months, and along with it, the medical center's revenue rose and fell.
At the same time, the company could not afford to scale the call center – say, hire a dozen more managers to manage operators. Such innovations would be unprofitable even if they led to positive results.
What task was set for Eleveo?
All actions should ultimately ensure an increase in conversion and income without inflating the staff and additional costs:
- Provide clear analytics for each operator. 20 parameters were compiled to evaluate the quality of work of the person who answers phone calls: how the greeting sounded, whether an anamnesis was collected, whether an appointment was made, whether its details were discussed, whether additional services were offered, etc. Each call was scored on a 20-point system depending on how well all the points were met - thus, at the end of each day, the operator who completed it could receive adjustments to correct errors in his actions.
- Take the burden off supervisors. Managers could avoid spending an insane amount of hours analyzing records, and the company could avoid doubling the number of call center supervisors and resorting to outsourced admin support.
- Create a basis for further development. So, this could be auto-filling cards for CRM or real-time prompts for operators that would help them while talking with customers.
Where did we start?
When creating the pilot version, we targeted the first 10 parameters/phrases that were significant for conversion to see if they were regularly spoken by operators. In a very simplified form, the Eleveo integration diagram looked like this:
How did we create labels and why are they needed?
Within the framework, all inbound call center operators use phrases with the same meaning, which may vary slightly depending on the situation, the rhythm of the conversation and the context. These phrases can be assigned labels:
– “Good afternoon, you have called such and such medical center” – “Hello”. – “I’m making an appointment for you to see the ophthalmologist for November 15” – “Registration”. – “There is a promotion for you” – “Promo”.
Labels allow you to track whether an operator is saying key phrases. This is of great importance: for example, if an employee does not know that the patient’s weight exceeds 120 kg, he will sign him up for examination on a device not designed for such loads, and if he cannot present new cool equipment, his return on investment will decrease. Any digital marketing specialist, and in principle a novice marketer, knows about the importance of such phrases.
The “keys” that we have got:
- Greeting and asking the patient's name;
- Collection of primary medical history;
- Consultation on visiting issues;
- Processing of objections;
- Voicing of contraindications to the procedure;
- Offer of additional services;
- Make an appointment;
- Announcement of cost;
- Voicing details of recording;
- Presentation of equipment.
This is how Eleveo identified the “Hello” label with a probability of 0.773.
And here Eleveo noted the “Contraindications” label with a probability of 0.939 when collecting data from a patient who wants to get an MRI.
Difficulties processing live speech
In email marketing, unlike calls, the message is always built according to fixed rules. True, in spoken English it is also customary to use certain constructions, which greatly simplifies the task of the program – you can, for example, give it a command to find phrases in a call recording that begin with a certain word, and get exactly what you are looking for. But this does not mean that live speech does not make problems.
There are three possible options for “sabotage”:
- Slang, unintelligible exclamations, and words borrowed from other languages clog up speech, interfering with processing.
- Words that are similar in meaning or sound cause confusion.
- The text is checked character by character, which is quite slow.
After weighing all this, we decided to use LSTM recurrent neural networks with long and short-term memory, capable of applying natural language processing techniques.
As an example: to understand the operator's phrase that the client should fast for 6 hours before an abdominal MRI, Eleveo covers not only keywords, but also context phrases.
If the search were based solely on keywords, Eleveo would mistakenly label the phrase as “Recording”. All first-generation Speech Analytics technologies focus on keywords, and are therefore complex and expensive. Second-generation speech analytics analyzes context and performs its assigned responsibilities more effectively.
Hundreds of calls
Some time before this order, I had the opportunity “from the front row” to observe how a neural network was trained to operate labels for Canada outsourcing call centers. This required thousands of calls! But we didn’t have time and resources to follow the usual path, so we set the Data Scientist the task of finding a faster and cheaper option and, using the experimental technology presented, were able to train Eleveo with the first 10 labels in just 100 calls, after which we expanded to 20.
At the same time, we assessed the quality of work of current supervisors and found significant discrepancies between their assessments and reality. After conducting an investigation, we found out that supervisors overestimated operators with whom they communicated well personally, and this clearly demonstrated another disadvantage of the human factor and the advantage of a virtual assistant, if this name is applicable to a neural network – it does not have favorites.
Deadlines and staff
Development, training and debugging of the process as a whole took us two months. At different stages, from 5 to 10 people were involved in the project on our side, and three on the client’s side.
Most of the time was consumed by the initial markup, final debugging of bugs and back-end improvements for correct integration with customer services and IP telephony, but after this experience, creating pilot versions was no longer difficult for our employees. In particular, this experience came in handy a couple of months later, when we were hired to organize a telemarketing call center, but that’s another story.
What did all this lead to?
The medical center has acquired a system capable of evaluating operator dialogues with clients in context and monitoring 100% of conversation records versus the previous 0.5%. All assessments are provided to management in a timely manner, allowing the right decisions to be made to improve the quality of the call center’s activities.
The conversion stopped jumping and took a confident position at the level of 25%. In the next three months we plan to raise it to 30%. Considering that each percentage translated into financial language means thousands of dollars, simply fixing it at 25% for the duration of the pilot launch, we have already given the center an ROI of 328% without introducing a 24 hour answering service, inflating the staff and other expensive measures.
We were afraid of opposition from operators and supervisors, but here a smart system came to the rescue, showing high accuracy in identifying phrases, and the center’s management, which used analytics data to train employees, and not for fines and punishments. Now we're going to combine data from the company of sites with end-to-end analytics to see effective ways to manage user attention and figure out the best advertising channels. We will soon introduce online tips for operators and add Eleveo’s ability to calculate the client’s mood to further increase conversion.