Our AI call center platform offers a conversational AI voice bot for automated phone calling and candidate screening. This TTS-enabled bot supports AI cold calling, efficiently handling conversations, rescheduling calls, and saving results to ATS. Leveraging AWS Lambda, the scalable voice agent solution reduces overhead and increases performance. By integrating generative and rule-based approaches, this AI cold caller software minimizes errors, lowers costs, and enhances call center efficiency—enabling fast, accurate, and automated call management for staffing companies.
A large American staffing company approached us. Their business is based on quickly finding candidates for companies. One of the most important metrics for their business is speed. Within a maximum of 3 days from receiving a request, they need to provide the client with several candidates who are currently available for hiring and meet the job requirements.
The most labor-intensive part of their work is calling candidates, screening them, and determining their availability for work. For this, the company employs thousands of recruiters who perform this routine work manually.
A major issue in this process is the overhead costs, including the salaries of recruiters who have to call potential candidates. Additionally, after each call, the recruiter must enter information about the conversation into the ATS, including data such as the call outcome and whether the candidate is suitable for the vacancy. In particular, it is necessary to record how the candidate answered the screening questions.
Moreover, when performing long routine processes, people inevitably make mistakes, both during candidate calls and when entering call information into the database.
Existing solutions for automating the calling process proved ineffective, as they were difficult to adapt to the company's needs. Therefore, the client turned to BroutonLab developers to create a custom ai call center solution to address these issues.
Our task was to develop an AI-powered system for automating candidate calls and screening. The system must be scalable and capable of making a thousand calls in parallel.
A specialized voice bot (or voice agent) should conduct the conversation according to the script, including asking screening questions. It should interact like a human, be able to answer questions, and maintain a conversation. For example, if the wrong person answers the call, it should politely ask for the candidate. It should also wait if the candidate needs time to find a pen to write down the details of the conversation. The number of such cases can be large, but all are important, and the voice bot must handle them properly.
After each call, the voice bot should save all call data in the ATS, including a screening score for each question as well as the candidate's final screening score.
The voice bot should call candidates according to a schedule. If a candidate is unable to speak at the given time, the voice bot should reschedule the call and call at the appointed time.
This tool should include a UI as well as an API for automating company processes.
Before starting the work, our team analyzed existing voice agent technologies. Since commercial solutions did not satisfy the client due to limited customization, we considered open-source technologies for creating chatbots (such as RASA and Deep Pavlov). Unfortunately, none of them met our architectural requirements. Therefore, we decided to develop a completely custom solution.
Our solution consisted of several modules:
To create the chatbot, we used a hybrid approach, combining modern generative methods based on LLM (ChatGPT, LLama) and a rule-based approach to control dialogue and reduce model hallucinations. We used Prompt Engineering and LLM fine-tuning to optimize key metrics for our AI. Additionally, an important part of the work was creating an automated validation system for our solution on a large set of cases, which helped make the bot training more stable.
For the voice service, we used ready-made technologies from ElevanLabs (voice generation) and DeepGram (analyzing the interlocutor's voice and converting it to text). These services performed well. As part of this module, we did extensive work to optimize its speed. In particular, we managed to reduce the voice bot's response time to 500 ms. Additionally, we developed unique algorithms for echo suppression and voicemail detection.
Our solution uses AWS Lambda, allowing it to be easily scalable regardless of the number of parallel calls. Additionally, this approach is serverless, eliminating the need for dedicated 24/7 servers. As a result, the client only needs to pay for the actual use of hardware resources during the call.
For UI development, we used Quasar/Vue.js. We created a user-friendly back-office interface that includes everything needed for configuring our solution and analyzing screening results.
As a result, our solution was successfully integrated into the client's company processes. Furthermore, they are now selling it to companies with similar needs.
The cost of calls decreased to 2 cents per minute, whereas before implementing our solution, the client had to pay their employees an hourly wage for the same work.
The need for hiring and managing thousands of call center specialists was completely eliminated, as this process is now fully automated.
The speed of finding candidates increased threefold (now it takes only one day to find a candidate instead of three as before). The error rate was reduced by 70%, as artificial intelligence continuously learns.
The client is completely satisfied with the results of our collaboration, and we are currently working on new technologies for further automating and increasing the efficiency of their operations.
