Intelligent Together: GenAI bots set to transform customer experience

Many companies are struggling with both human resource costs and labor shortages, and are looking to automation to improve customer service and support agents. According to Gartner, the cost of employing and recruiting agents can represent up to 95% of contact center costs.

Although contact centers have long used automation to control costs, many customers have been left frustrated by confusing IVR menu systems and ineffective chatbots.

Moreover the idea here is not about reducing costs primarily, it’s to help human agents to make the best use of their time, get access to information instantly thanks to AI, avoid to have to put people on hold, and thus creating an enhanced customer experience resulting in much higher loyalty.

This looks set to be transformed by generative AI. Its ability to understand and generate natural language makes it an excellent fit for automating customer interactions. Gartner says that 85% of customer service leaders surveyed will explore or pilot a customer-facing conversational generative AI (GenAI) solution in 2025.

In addition to their value in customer-facing applications, GenAI-powered conversational bots can automate repetitive tasks to reduce the workload on human agents, especially in support and IT helpdesk scenarios. They can act as agent assist tools to improve productivity by locating relevant information in real time. GenAI can even replace some offshore support functions, allowing companies to bring operations back in-house or to the cloud with lower costs, , and have human agents focusing on higher added value tasks.

For companies in highly regulated industries, conversational bots can be deployed in private cloud environments, ensuring data sovereignty and prevented from retraining on sensitive data. In fact, there is growing demand for on-premise or sovereign cloud solutions, as companies look to protect their data and comply with industry or national regulations. This is the case for public sector deployments, for example.

While text-based chatbots are more familiar to users, providing voice as a channel is essential. Consumers are accustomed to voice commands with tools such as Alexa and Siri, so the concept is already well-accepted in the market, especially with younger consumers.

Deploying voice-based chatbots, however, requires additional work. For a start, voice-to-voice GenAI models are still in their infancy, and processing voice relies on speech-to-text and text-to-speech.

Sound quality can also be an issue. Voice applications are particularly useful when consumers find it hard to interact with text chat, which is likely to happen when they are on the move or doing something else with their hands. For example, driving a car or operating machinery – both are noisy environments that makes voice analysis difficult.

Properly managing the voice channel of a bot therefore requires strong expertise in ASR model fine tuning e.g. or in recommending the proper model. There is no point in using a generalist LLM when the bot is always about reporting an insurance claim or finding the next train from A to B. Model finetuning is essential to offer the best proper answer ratio for a given customer use case. This shows that deploying and managing a bot is not about buying a piece of software running on some cloud, this is about building the proper solution, supported by the relevant experts from the initial design phase.

In terms of deployment, a critical requirement for conversational AI is avoiding hallucinations. Having a contact center bot give wrong information to callers can be very damaging to the brand and even dangerous.

Companies should use guardrails to restrict data sources and prevent bots from using the internet to answer queries. The key is to use retrieval augmented generation (RAG) which connects LLMs with structured knowledge bases. It ensures that bots provide more accurate, context-aware responses by retrieving relevant documents or data before generating answers.

By connecting to multiple knowledge bases, RAG enables chatbots to answer any questions based on company information without predefining them. This is critical for reducing hallucinations and maintaining data integrity.

However, data readiness is an issue for many companies, which continue to have unstructured or outdated documentation in their knowledge base. According to research from Gartner, 61% of companies have a backlog of articles to edit in the database, and more than a third of them don’t have a clear process for revising out-of-date information.

In addition, companies should consider using scripted flows for high-risk use cases, such as gathering information for a complaint. Chatbots should escalate to a human agent when the confidence that they can provide an answer is low, or it needs to make an impactful decision, such as approving an insurance payout.

Chatbots must also be integrated with contact center, CRM and IT systems to avoid silos and fragmented experiences, such as chat-to-agent transfers losing context. This can be particularly challenging with some contact center software deployment, which can include multiple vendors and legacy systems – some of which might be operating on-premise. They may need to be modernized to fully leverage AI capabilities.

Analyst Forrester argues that self-service applications that answer customer questions are helpful, but without the ability to connect to back-end systems, a chatbot is of limited value. “If you can’t check on the status of an order, schedule an appointment, or make a purchase, automation will fall short in customers’ eyes,” it says. 

Orange has developed a range of packaged autonomous agentic AI agents that can undertake standard activities. These agents perform tasks, such as appointment scheduling, trouble ticket creation, or document analysis. Examples include Agentic AI bots for the insurance industry that can analyze accident reports, images, and guide users through structured data submission.

Imagine this scenario. A customer has a car accident and calls their insurance company from the scene. The call is handled by a bot, which pulls up the customer’s CRM file, identifies, collects and analyses the required information. It can request further information if required and send any decision requests to a person. It can call on specialized Agentic AIs to set up appointments with a garage to fix the damage, update the CRM file, and send a detailed accident report to the customer and other relevant parties.

Sentiment analysis adds a new dimension to the chatbot experience. The bots can analyze voice calls and the conversational flow to understand the caller’s sentiment, such as whether they are angry or sad. This then allows the bots to take actions, such as escalating to a human agent or adjusting the tone or call-flow mid-conversation.

In addition, voice analytics can provide useful post-call insights, such as understanding what approach works best in which scenario and this can be fed into the training of both the bots and human agents.

Typical exemples of the value added here: detection of sales call that ended to a voice mail, understand the reasons that prevent x percentage of sales to not finalized.

The market of conversational AI is quite volatile, a number of players are appearing or disappearing every quarter, making customer life difficult when it is about to choose a solution that should be there for several years.

With its multidecade experience of the CX market, Orange Business has decided to position himself as an advisor to its customers, a solution benchmarker, as well as a provider of a very modular conversational AI platform.

Orange Business has developed an AI bot orchestrator to help companies efficiently deploy GenAI-based chatbots and voice bots. The modular, flexible platform integrates various bots and large language models (LLMs), including those already deployed by clients. Orange Business experts can help users to choose the LLM to meet their needs based on language, industry focus, data sovereignty or sustainability criteria.

As a number of our customers come with existing bots, we have taken the unique approach to have a solution that can reuse and extend existing bots, such as adding voice to a chat-only bot. It also supports component-level replacement, enabling updates to specific AI modules (like switching LLMs) without rebuilding the entire system. This modular approach improves performance, cost-efficiency, and flexibility.

In addition to the AI Orchestrator platform, Orange Business provides tools to fine tune the bot behavior over time and for specific use cases. For example, in a public transport application, the bot must be able to perfectly identify street names – even if spoken in a noisy environment like a subway station.

Last but not least, this value proposal is also strongly supported by our professional services teams, with strong skills on data science and conversational bots design, which are critical for our customers to enjoy bots that are performing fine over the long run, and having these teams is a strong differentiator vs a number of conversational AI players who might have a powerful product but very limited people to support.

While conversational AI in chatbots is an exciting opportunity, most projects are generally still in pilot or proof-of-concept stages, with full-scale deployment pending. In some cases, it makes sense to deploy the chatbot first as an employee support tool before rolling it out to be customer-facing. This allows companies to solve any inconsistency or errors before exposing it to the world.

To help companies move forward, Orange Business takes a consultative approach to deployment to understand client needs, infrastructure, and goals. Our team of data scientists and professional experts work with customers to help design an optimized conversational bot experience. We also offer global deployment capabilities with consistent service levels across regions, which is a key requirement for multinational companies.