AI-powered call center solution with Azure and OpenAI GPT.
Sign in to like and favorite skills
AI-powered call center solution with Azure and OpenAI GPT.
Send a phone call from AI agent, in an API call. Or, directly call the bot from the configured phone number!
Insurance, IT support, customer service, and more. The bot can be customized in few hours (really) to fit your needs.
# Ask the bot to call a phone number data='{ "bot_company": "Contoso", "bot_name": "Amélie", "phone_number": "+11234567890", "task": "Help the customer with their digital workplace. Assistant is working for the IT support department. The objective is to help the customer with their issue and gather information in the claim.", "agent_phone_number": "+33612345678", "claim": [ { "name": "hardware_info", "type": "text" }, { "name": "first_seen", "type": "datetime" }, { "name": "building_location", "type": "text" } ] }' curl \ --header 'Content-Type: application/json' \ --request POST \ --url https://xxx/call \ --data $data
Enhanced communication and user experience: Integrates inbound and outbound calls with a dedicated phone number, supports multiple languages and voice tones, and allows users to provide or receive information via SMS. Conversations are streamed in real-time to avoid delays, can be resumed after disconnections, and are stored for future reference. This ensures an improved customer experience, enabling 24/7 communication and handling of low to medium complexity calls, all in a more accessible and user-friendly manner.
Advanced intelligence and data management: Leverages gpt-4.1 and gpt-4.1-nano (known for higher performance and a 10–15x cost premium) to achieve nuanced comprehension. It can discuss private and sensitive data, including customer-specific information, while following retrieval-augmented generation (RAG) best practices to ensure secure and compliant handling of internal documents. The system understands domain-specific terms, follows a structured claim schema, generates automated to-do lists, filters inappropriate content, and detects jailbreak attempts. Historical conversations and past interactions can also be used to fine-tune the LLM, improving accuracy and personalization over time. Redis caching further enhances efficiency.
Customization, oversight, and scalability: Offers customizable prompts, feature flags for controlled experimentation, human agent fallback, and call recording for quality assurance. Integrates Application Insights for monitoring and tracing, provides publicly accessible claim data, and plans future enhancements such as automated callbacks and IVR-like workflows. It also enables the creation of a brand-specific custom voice, allowing the assistant’s voice to reflect the company’s identity and improve brand consistency.
Cloud-native deployment and resource management: Deployed on Azure with a containerized, serverless architecture for low maintenance and elastic scaling. This approach optimizes costs based on usage, ensuring flexibility and affordability over time. Seamless integration with Azure Communication Services, Cognitive Services, and OpenAI resources provides a secure environment suitable for rapid iteration, continuous improvement, and accommodating variable workloads in the call center.
A French demo is avaialble on YouTube. Do not hesitate to watch the demo in x1.5 speed to get a quick overview of the project. Voice is hesitant on purpose to show the bot can handle it. All the infrastructure is deployed on Azure, mostly in serverless mode. Provisionning of the LLM resources can be done to reduce the latency.
Main interactions shown in the demo:
Extract of the data stored during the call:
{ "claim": { "incident_description": "Collision avec un autre véhicule, voiture dans le fossé, pas de blessés", "incident_location": "Nationale 17", "involved_parties": "Dujardin, Madame Lesné", "policy_number": "DEC1748" }, "messages": [ { "created_at": "2024-12-10T15:51:04.566727Z", "action": "talk", "content": "Non, je pense que c'est pas mal. Vous avez répondu à mes questions et là j'attends la dépaneuse. Merci beaucoup.", "persona": "human", "style": "none", "tool_calls": [] }, { "created_at": "2024-12-10T15:51:06.040451Z", "action": "talk", "content": "Je suis ravi d'avoir pu vous aider! Si vous avez besoin de quoi que ce soit d'autre, n'hésitez pas à nous contacter. Je vous souhaite une bonne journée et j'espère que tout se passera bien avec la dépanneuse. Au revoir!", "persona": "assistant", "style": "none", "tool_calls": [] } ], "next": { "action": "case_closed", "justification": "The customer has provided all necessary information for the insurance claim, and a reminder has been set for a follow-up call. The customer is satisfied with the assistance provided and is waiting for the tow truck. The case can be closed for now." }, "reminders": [ { "created_at": "2024-12-10T15:50:09.507903Z", "description": "Rappeler le client pour faire le point sur l'accident et l'avancement du dossier.", "due_date_time": "2024-12-11T14:30:00", "owner": "assistant", "title": "Rappel client sur l'accident" } ], "synthesis": { "long": "During our call, you reported an accident involving your vehicle on the Nationale 17. You mentioned that there were no injuries, but both your car and the other vehicle ended up in a ditch. The other party involved is named Dujardin, and your vehicle is a 4x4 Ford. I have updated your claim with these details, including the license plates: yours is U837GE and the other vehicle's is GA837IA. A reminder has been set for a follow-up call tomorrow at 14:30 to discuss the progress of your claim. If you need further assistance, please feel free to reach out.", "satisfaction": "high", "short": "the accident on Nationale 17", "improvement_suggestions": "To improve the customer experience, it would be beneficial to ensure that the call connection is stable to avoid interruptions. Additionally, providing a clear step-by-step guide on what information is needed for the claim could help streamline the process and reduce any confusion for the customer." } ... }
A report is available at
https://[your_domain]/report/[phone_number]http://localhost:8080/report/%2B133658471534
--- title: System diagram (C4 model) --- graph user(["User"]) agent(["Agent"]) app["Call Center AI"] app -- Transfer to --> agent app -. Send voice .-> user user -- Call --> app
--- title: Claim AI component diagram (C4 model) --- graph LR agent(["Agent"]) user(["User"]) subgraph "Claim AI" ada["Embedding<br>(ADA)"] app["App<br>(Container App)"] communication_services["Call & SMS gateway<br>(Communication Services)"] db[("Conversations and claims<br>(Cosmos DB)")] eventgrid["Broker<br>(Event Grid)"] gpt["LLM<br>(gpt-4.1, gpt-4.1-nano)"] queues[("Queues<br>(Azure Storage)")] redis[("Cache<br>(Redis)")] search[("RAG<br>(AI Search)")] sounds[("Sounds<br>(Azure Storage)")] sst["Speech-to-text<br>(Cognitive Services)"] translation["Translation<br>(Cognitive Services)"] tts["Text-to-speech<br>(Cognitive Services)"] end app -- Translate static TTS --> translation app -- Sezarch RAG data --> search app -- Generate completion --> gpt gpt -. Answer with completion .-> app app -- Generate voice --> tts tts -. Answer with voice .-> app app -- Get cached data --> redis app -- Save conversation --> db app -- Transform voice --> sst sst -. Answer with text .-> app app <-. Exchange audio .-> communication_services app -. Watch .-> queues communication_services -- Load sound --> sounds communication_services -- Notifies --> eventgrid communication_services -- Transfer to --> agent communication_services <-. Exchange audio .-> agent communication_services <-. Exchange audio .-> user eventgrid -- Push to --> queues search -- Generate embeddings --> ada user -- Call --> communication_services
[!NOTE] This project is a proof of concept. It is not intended to be used in production. This demonstrates how can be combined Azure Communication Services, Azure Cognitive Services and Azure OpenAI to build an automated call center solution.
Prefer using GitHub Codespaces for a quick start. The environment will setup automatically with all the required tools.
In macOS, with Homebrew, simply type
make brewFor other systems, make sure you have the following installed:
bashzshapt install makeyum install makebrew install makeThen, Azure resources are needed:
ccai-customer-aNow that the prerequisites are configured (local + Azure), the deployment can be done.
A pre-built container image is available on GitHub Actions, it will be used to deploy the solution on Azure:
ghcr.io/clemlesne/call-center-ai:mainghcr.io/clemlesne/call-center-ai:0.1.0Fill the template from the example at
. The file should be placed at the root of the project under the name config-remote-example.yaml
config.yamlaz login
[!TIP] Specify the release version under the
parameter (default isimage_version). For example,mainorimage_version=16.0.0. This will ensure any future project breaking changes won't affect your deployment.image_version=sha-7ca2c0c
make deploy name=my-rg-name
Wait for the deployment to finish.
make logs name=my-rg-name
If you skiped the
make brewFinally, run
make installIf the application is already deployed on Azure, you can run
make name=my-rg-name sync-local-config[!TIP] To use a Service Principal to authenticate to Azure, you can also add the following in a
file:.envAZURE_CLIENT_ID=xxx AZURE_CLIENT_SECRET=xxx AZURE_TENANT_ID=xxx
If the solution is not running online, fill the template from the example at
. The file should be placed at the root of the project under the name config-local-example.yaml
config.yamlExecute if the solution is not yet deployed on Azure.
make deploy-bicep deploy-post name=my-rg-name
[!IMPORTANT] Tunnel requires to be run in a separate terminal, because it needs to be running all the time
# Log in once devtunnel login # Start the tunnel make tunnel
[!NOTE] To override a specific configuration value, you can use environment variables. For example, to override the
value, you can use thellm.fast.endpointvariable:LLM__FAST__ENDPOINTLLM__FAST__ENDPOINT=https://xxx.openai.azure.com
[!NOTE] Also,
script is available to test the application without the need of a phone call (= without Communication Services). Run the script with:local.pypython3 -m tests.local
make dev
http://localhost:8080Call recording is disabled by default. To enable it:
recordingsrecording_enabledtrueTraining data is stored on AI Search to be retrieved by the bot, on demand.
Required index schema:
| Field Name | | Retrievable | Searchable | Dimensions | Vectorizer |
|---|---|---|---|---|---|
| answer | | Yes | Yes | ||
| context | | Yes | Yes | ||
| created_at | | Yes | No | ||
| document_synthesis | | Yes | Yes | ||
| file_path | | Yes | No | ||
| id | | Yes | No | ||
| question | | Yes | Yes | ||
| vectors | | No | Yes | 1536 | OpenAI ADA |
Software to fill the index is included on Synthetic RAG Index repository.
The bot can be used in multiple languages. It can understand the language the user chose.
See the list of supported languages for the Text-to-Speech service.
# config.yaml conversation: initiate: lang: default_short_code: fr-FR availables: - pronunciations_en: ["French", "FR", "France"] short_code: fr-FR voice: fr-FR-DeniseNeural - pronunciations_en: ["Chinese", "ZH", "China"] short_code: zh-CN voice: zh-CN-XiaoqiuNeural
If you built and deployed an Azure Speech Custom Neural Voice (CNV), add field
custom_voice_endpoint_id# config.yaml conversation: initiate: lang: default_short_code: fr-FR availables: - pronunciations_en: ["French", "FR", "France"] short_code: fr-FR voice: xxx custom_voice_endpoint_id: xxx
Levels are defined for each category of Content Safety. The higher the score, the more strict the moderation is, from 0 to 7. Moderation is applied on all bot data, including the web page and the conversation. Configure them in Azure OpenAI Content Filters.
Customization of the data schema is fully supported. You can add or remove fields as needed, depending on the requirements.
By default, the schema of composed of:
caller_emailemailcaller_nametextcaller_phonephone_numberValues are validated to ensure the data format commit to your schema. They can be either:
datetimeemailphone_numberE164textFinally, an optional description can be provided. The description must be short and meaningful, it will be passed to the LLM.
Default schema, for inbound calls, is defined in the configuration:
# config.yaml conversation: default_initiate: claim: - name: additional_notes type: text # description: xxx - name: device_info type: text # description: xxx - name: incident_datetime type: datetime # description: xxx
Claim schema can be customized for each call, by adding the
claimPOST /callThe objective is a description of what the bot will do during the call. It is used to give a context to the LLM. It should be short, meaningful, and written in English.
This solution is priviledged instead of overriding the LLM prompt.
Default task, for inbound calls, is defined in the configuration:
# config.yaml conversation: initiate: task: | Help the customer with their insurance claim. Assistant requires data from the customer to fill the claim. The latest claim data will be given. Assistant role is not over until all the relevant data is gathered.
Task can be customized for each call, by adding the
taskPOST /callConversation options are represented as features. They can be configured from App Configuration, without the need to redeploy or restart the application. Once a feature is updated, a delay of 60 secs is needed to make the change effective.
By default, values are refreshed every 60 seconds. Refresh is not sync across all instances, so it can take up to 60 seconds to see the change on all users. Update this in the
app_configuration.ttl_sec| Name | Description | Type | Default |
|---|---|---|---|
| Time waiting the LLM before aborting the answer with an error message. | | 15 |
| Time waiting the LLM before sending a waiting message. | | 4 |
| The timeout for a callback in hours. Set 0 to disable. | | 3 |
| Amount of silence in secs to trigger a warning message from the assistant. | | 20 |
| TThe maximum number of retries for voice recognition. Minimum of 1. | | 3 |
| The timeout for STT completion in milliseconds. | | 100 |
| Whether call recording is enabled. | | false |
| Whether to use the slow LLM for chat. | | false |
| The cutoff timeout for voice activity detection in milliseconds. | | 250 |
| Silence to trigger voice activity detection in milliseconds. | | 500 |
| The threshold for voice activity detection. Between 0.1 and 1. | | 0.5 |
To use Twilio for SMS, you need to create an account and get the following information:
Then, add the following in the
config.yaml# config.yaml sms: mode: twilio twilio: account_sid: xxx auth_token: xxx phone_number: "+33612345678"
Note that prompt examples contains
{xxx}{bot_name}# config.yaml prompts: tts: hello_tpl: - : | Hello, I'm {bot_name}, from {bot_company}! I'm an IT support specialist. Here's how I work: when I'm working, you'll hear a little music; then, at the beep, it's your turn to speak. You can speak to me naturally, I'll understand. What's your problem? - : | Hi, I'm {bot_name} from {bot_company}. I'm here to help. You'll hear music, then a beep. Speak naturally, I'll understand. What's the issue? llm: default_system_tpl: | Assistant is called {bot_name} and is in a call center for the company {bot_company} as an expert with 20 years of experience in IT service. # Context Today is {date}. Customer is calling from {phone_number}. Call center number is {bot_phone_number}. chat_system_tpl: | # Objective Provide internal IT support to employees. Assistant requires data from the employee to provide IT support. The assistant's role is not over until the issue is resolved or the request is fulfilled. # Rules - Answers in {default_lang}, even if the customer speaks another language - Cannot talk about any topic other than IT support - Is polite, helpful, and professional - Rephrase the employee's questions as statements and answer them - Use additional context to enhance the conversation with useful details - When the employee says a word and then spells out letters, this means that the word is written in the way the employee spelled it (e.g. "I work in Paris PARIS", "My name is John JOHN", "My email is Clemence CLEMENCE at gmail GMAIL dot com COM") - You work for {bot_company}, not someone else # Required employee data to be gathered by the assistant - Department - Description of the IT issue or request - Employee name - Location # General process to follow 1. Gather information to know the employee's identity (e.g. name, department) 2. Gather details about the IT issue or request to understand the situation (e.g. description, location) 3. Provide initial troubleshooting steps or solutions 4. Gather additional information if needed (e.g. error messages, screenshots) 5. Be proactive and create reminders for follow-up or further assistance # Support status {claim} # Reminders {reminders}
The delay mainly come from two things:
From now, the only impactful thing you can do is the LLM part. This can be acheieve by a PTU on Azure or using a less smart model like
gpt-4.1-nanoThe application is natively connected to Azure Application Insights, so you can monitor the response time and see where the time is spent. This is a great start to identify the bottlenecks.
Feel free to raise an issue or propose a PR if you have any idea to optimize the response delay.
Enhance the LLM’s accuracy and domain adaptation by integrating historical data from human-run call centers. Before proceeding, ensure compliance with data privacy regulations, internal security standards, and Responsible AI principles. Consider the following steps:
Application send traces and metrics to Azure Application Insights. You can monitor the application from the Azure portal, or by using the API.
This includes application behavior, database queries, and external service calls. Plus, LLM metrics (latency, token usage, prompts content, raw response) from OpenLLMetry, following the semantic sonventions for OpenAI operations.
Additionally custom metrics (viewable in Application Insights > Metrics) are published, notably:
call.aec.dropedcall.aec.missedcall.answer.latencyFor a monthly usage of 1000 calls of 10 minutes each. Costs are estimated for 2024-12-10, in USD. Prices are subject to change.
[!NOTE] For production usage, it is recommended to upgrade to SKUs with vNET integration and private endpoints. This can increase notably the costs.
This totalizes $720.07 /month, $0.12 /hour, with the following breakdown:
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| West Europe | Audio Streaming | $0.004 /minute | $40 |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | gpt-4.1-nano global | $0.15 /1M input tokens | $35.25 | 8k tokens for conversation history, 3750 tokens for RAG, each participant talk every 15s |
| Sweden Central | gpt-4.1-nano global | $0.60 /1M output tokens | $1.4 | 400 tokens for each response incl tools, each participant talk every 15s |
| Sweden Central | gpt-4.1 global | $2.50 /1M input tokens | $10 | 4k tokens for each conversation, to get insights |
| Sweden Central | gpt-4.1 global | $10 /1M output tokens | $10 | 1k tokens for each conversation, to get insights |
| Sweden Central | text-embedding-3-large | $0.00013 /1k tokens | $2.08 | 1 search or 400 tokens for each message, each participant talk every 15s |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | Serverless vCPU | $0.000024 /sec | $128.56 | Avg of 2 replicas with 1 vCPU |
| Sweden Central | Serverless memory (average of 2 replicas) | $0.000003 /sec | $32.14 | Avg of 2 replicas with 2GB |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | Basic | $73.73 /month | $73.73 | Has 15GB of storage /index, should be upgraded for big datasets |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| West Europe | Speech-to-text real-time | $1 /hour | $83.33 | Each participant talk every 15s |
| West Europe | Text-to-speech standard | $15 /1M characters | $69.23 | 300 tokens for each response, 1.3 tokens /word in English, each participant talk every 15s |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | Multi-region write RU/s /region | $11.68 /100 RU/s | $233.6 | Avg of 1k RU/s on 2 regions |
| Sweden Central | Transactional storage | $0.25 /GB | $0.5 | 2GB of storage, should be upgraded if more history is needed |
Not included upper:
[!NOTE] Azure Monitor costs shouldn't be considered as optional as monitoring is a key part of maintaining a business-critical application and high-quality service for users.
Optional costs totalizing $343.02 /month, with the following breakdown:
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| West Europe | Call recording | $0.002 /minute | $20 |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | text-embedding-3-large | $0.00013 /1k tokens | $0.52 | 10k PDF pages with 400 tokens each, for indexing |
| Region | Metric | Cost | Total (monthly $) | Note |
|---|---|---|---|---|
| Sweden Central | Basic logs ingestion | $0.645 /GB | $322.5 | 500GB of logs with sampling enabled |
Quality:
Reliability:
Maintainability:
Resiliency:
Security:
Responsible AI:
At the time of development, no LLM framework was available to handle all of these features: streaming capability with multi-tools, backup models on availability issue, callbacks mechanisms in the triggered tools. So, OpenAI SDK is used directly and some algorithms are implemented to handle reliability.
gpt-4o-realtimegpt-4o-realtime