Featured
Smart AI agents will radically change how we think about autonomization
Frontline AI from QuantumBlack, AI by McKinsey
Recent advances in generative AI (gen AI) are poised to revolutionize businesses and work processes. This article describes the impact of smart AI agents on shaping hybrid workforces and evolving labor models of the future. We’ll introduce Frontline AI by QuantumBlack, which uses smart AI agents to revolutionize customer interactions and boost operational efficiency with tailored, near-human generative AI. We’ll share a real-world illustration of a transformative application for frontline workforce management in operations.
McKinsey research estimates that gen AI could earn the equivalent of between $2.6 trillion and $4.4 trillion annually. Since early 2023, the surge in interest surrounding gen AI and agentic models has been remarkable.
In the latest McKinsey Global Survey on the state of AI, 71 percent say their organizations regularly use gen AI in at least one business function. However, a substantial number of AI initiatives remain at the proof-of-concept (PoC) stage, and fail to transition into full-scale production. This stagnation has been a notable source of frustration for CEOs and technology leaders.
Traditional chatbots vs smart AI agents
Traditional chatbots rely on a pre-defined set of interactions. They require labor-intensive, rule-based programming or highly specific training of machine-learning models. These bots are typically limited in their ability to understand context. They may fail to handle complex or unexpected queries, resulting in a more rigid and less user-friendly experience.
In contrast, gen AI-powered AI agents use foundational models to execute complex, multi-step workflows. This makes them adaptable and capable of more sophisticated understanding. They can use context and logical reasoning to deliver faster, more accurate, and empathetic interactions than traditional chatbots.
AI-powered agents can revolutionize industries by performing complex tasks using capabilities such as:
- Tool usage: Integrating APIs, databases, and automation scripts to retrieve account details, update records, and execute workflows.
- Memory & context awareness: Retaining short-term memory for conversation flow and long-term memory for past interactions, enabling personalized responses.
- Feedback & learning mechanism: Refining actions based on structured feedback, user responses, system validations, and performance tracking.
- Planning & multi-step execution: Dynamically breaking down tasks and adjusting execution based on real-time conditions.
- Guardrails & business logic integration: Operating within policies and validation layers, ensuring reliability, compliance, and safe enterprise system interactions.
Systems using AI agents can go beyond simple automation by integrating reasoning, tool usage, and adaptive workflows. They unlock efficiencies that were previously impractical to automate. AI agents can assist in use cases needing real-time decision-making, multi-step execution, and dynamic coordination. For example, in banking, an agent can handle dispute resolution from start to finish by retrieving transaction records, assessing policies, and submitting claims automatically. In utilities, an agent can dynamically schedule field crews during outages, adjusting in real-time based on priorities and conditions. In healthcare, an agent can monitor patient vitals, detect anomalies, and trigger interventions without human oversight.
The challenge of smart AI agents
One main issue with LLM-based applications is their opaque nature, which makes it hard to understand their decisions. AI “hallucinations,” where models generate wrong or nonsensical information, add to the problem. While all machine learning models can make errors, language generation by LLMs makes the potential for errors riskier, especially in customer-facing applications. Creating appropriate safeguards and managing risks is ongoing, and full control over these systems is still challenging.
Another significant hurdle is the skill shift needed for agentic models. Traditional data science skills are less important, while software engineering skills for API integration and orchestration become crucial. This shift has revealed a skills gap, as teams used to regression models now need to build complex AI applications. Some shortcomings in agentic solutions can be attributed to the lack of adherence to sound product development practices
Commercial off-the-shelf (OTS) agentic solutions and low-code agent-building platforms are available, but results vary. These tools often need heavy customization to fit specific business needs and can be limited in terms of the models and tools they support. Some OTS solutions charge per interaction, making them costly at high volumes. It is inherently challenging to create a one-size-fits-all OTS solution because business processes and data needs vary widely. In summary, while OTS tools are often good for basic tasks, they fall short in handling complex reasoning methods such as multi-agent frameworks necessary for intricate business processes.
Despite these challenges, some companies have successfully used agentic solutions in areas like knowledge management, contact center automation, and helpdesks. These successes show that building agents, while different from traditional machine learning models, isn’t necessarily harder. Understanding agents’ unique features can help identify which processes to improve and how to implement these systems effectively.
Introducing QuantumBlack Frontline AI
QuantumBlack Frontline AI is a suite of AI-enabled tools designed to enhance workforce performance across industries. It includes AI-driven tools for demand forecasting, scheduling optimization, and generative AI-powered chatbots, with smart AI agents being the latest addition. Traditional AI and Machine Learning (AI/ML) techniques are integrated with cutting-edge generative AI to provide real-time insights, personalized coaching, and decision-making support. This leads to better efficiency, productivity, and satisfaction for frontline employees and their customers.
The conversational agent platform is a gen AI-powered system designed to enhance the entire conversation journey. At its core is an advanced “Brain” that supports various products like a Contact Center Autopilot and Agent Copilot, which automate customer interactions, and AI Coach and Digital Twin Tower, which provide personalized coaching and analytics.
Users interact with the smart AI agents through a user interface for bidirectional communication.
The Brain
- Profile: Defines the agent’s purpose, such as handling customer requests in a call center.
- Memory: Stores conversation data to understand interactions more effectively. Uses both short-term and long-term memory to deliver personalized recommendations. Frontline AI uses knowledge graphs to manage data quality and relevance from historical interactions. This enhances the AI agents’ ability to comprehend context by uncovering implicit connections between pieces of information
- Reasoning: Uses Large Language Models (LLMs) to understand user requests and make decisions. It includes: recognizing intent to determine the appropriate actions for AI agents; verifying relevance between prompts and actions to maintain adherence to guardrails; extracting key information from unstructured text into standardized templates; and selecting the next best step based upon prior actions. Frontline AI balances LLM-based reasoning with deterministic methods for tasks requiring high precision. By intelligently combining these approaches, the platform ensures the most appropriate methodology is applied to achieve both efficiency and accuracy.
- Planning: Manages workflows using a state machine framework. Static state machines follow predefined logic, and dynamic state machines use a multi-agent framework to handle complex workflows with greater flexibility. By adhering to state machine logic, AI agents ensure that interactions progress only when predefined conditions are satisfied, maintaining structure and control throughout the conversation.
- Expert Toolbox: Includes tools like the Knowledge Extractor to search and summarize documents with a multimodal Retrieval Augmented Generation (RAG) system; Data Manager to convert text to SQL to fetch data; and the Data Transformer, which converts free-form data into structured formats.
Interactive interface with multi-channel communication
The AI agent uses a flexible interface for interactions. It supports various avatars, such as voice-based agents, text chatbots, or command-line tools, enabling engagement across voice channels and digital platforms like WhatsApp and Slack.
Agnostic solution with seamless integration
Smart AI agents are practical, scalable, and reliable. They operate on cloud-native platforms like Kubernetes and can be deployed across any cloud provider. They seamlessly integrate with external databases and business systems like Salesforce and SAP, aligning with existing infrastructure.
The real-world application and impact of QuantumBlack Frontline AI
QuantumBlack partners with clients across industries to collaboratively design tailored solutions using Frontline AI. From creating a detailed blueprint to co-developing, testing, and deploying these tools in client environments, these solutions are highly customizable and aligned with each client’s unique needs. Alongside McKinsey’s expertise in change management, QuantumBlack supports AI adoption and capability building, enabling clients to achieve sustained value from the Frontline AI tool implementation.
The impact of Frontline AI in the utilities sector
Current challenges faced by the utilities sector globally include the following:
- Low automation in call centers: Only 15–30 percent of call centers use automation, leading to high costs and difficulties in hiring and retaining agents
- Call volume spikes: Events like extreme weather or price changes cause unpredictable call volume spikes, complicating staffing
- High volume of low-complexity customer needs: A high volume of customer calls are simple queries about usage information or billing. Typically, up to 40–60 percent of customer calls can be automated.
- Low customer satisfaction with automated systems: Many organizations have adopted IVR systems offering robotic and key-based customer interactions.
The following are two examples of how FrontlineAI has been used in the utilities sector.
US utility provider
A public utility company in the US implemented smart AI agents that were focused on Voice-to-Voice assistants for customer authentication, information updates, and billing inquiries.
The result was identification of a 40–60 percent reduction in total call-handle time by enhancing self-service options. The implementation process was as follows:
- Analyze call data using LLMs to break down historical call transcripts and calculate average time spent per activity.
- Assess automation complexity and divide into three categories: Simple tasks that verify and relay customer information, yielding high automation success; Moderate tasks that submit and create customer requests, requiring more features; Complex tasks that explain and interpret complex information, involving LLM reasoning and lower automation potential.
- Calculate value potential by evaluating the time spent, automation potential, and call volume to estimate an overall financial impact.
European utility provider
A European utility provider rolled out a smart AI agents solution to their customer base. The initial focus was on low-complexity customer needs that were previously handled by a traditional interactive voice response (IVR) system. The solution returned significant benefits that included an enhanced customer experience and increased operational efficiencies:
- 6pp uplift in customer satisfaction score (CSAT) compared to IVR.
- 2pp uplift in quality of voice (vs. recorded human voice in the IVR).
- 5pp uplift in speed of response (vs. IVR)
- 25 percent acceleration of customer journeys (vs. IVR)
- Up to 75 percent end-to-end automation of customer journeys
- 10 percent reduction of human agent average handling times
A customer was quoted as follows:
“For the very first time an automated voice system worked. The information I provided to the AI agent didn’t have to be repeated once I got through to a real human being.”
Gen AI offers close-to-human interactions and dynamic conversations, surpassing the capabilities of rule-based systems. This can significantly improve the customer experience, while automating interactions and increasing operational efficiencies.
The Future of AI agents
AI agents are set to revolutionize various industries by performing complex, multi-step tasks autonomously. These agents integrate components like specialized tools, structured workflows, and memory, making them more capable, reliable, and adaptable than stand-alone large language models (LLMs).
The challenges and how to overcome them
- Planning and execution: Advances in task decomposition and hybrid LLM-symbolic approaches will help break down complex objectives into manageable steps. Integrating deterministic logic and state-based frameworks could provide more structured execution and reduce failure rates.
- Contextual consistency: Enhanced memory capabilities enable agents to track long-term workflows across multiple sessions. However, designing effective retrieval mechanisms remains a challenge.
- Hallucinations: Reduction through use of self-verification mechanisms, where agents double-check their outputs or consult external knowledge sources. Some systems use “critic agents” to fact-check decisions before execution.
- Efficiency: Current AI agents require multiple LLM calls per decision, increasing costs and slowing execution. Strategies include caching, optimized tool invocation, prompt compression, and hybrid execution models, using smaller task-specific models for simpler operations.
- Business adoption: Successful adoption will depend on establishing governance frameworks, security measures, and auditability to ensure that AI agents operate transparently and within well-defined limits.
QuantumBlack Labs is the R&D and software development hub within QuantumBlack, AI by McKinsey. QuantumBlack Labs has over 250 technologists dedicated to driving AI innovation and supporting and accelerating the work of McKinsey’s 1400+ data scientists. We use our collective experience to develop suites of tools and assets that ensure AI/ML models reach production and achieve sustained impact.
To find out more about QuantumBlack Frontline AI please contact Sohrab Rahimi and Zhekun Xiong.
Thank you to all who contributed to this article: Sohrab Rahimi, Zhekun Xiong, Julian Kleindiek, Nicholas Scharan, Aloizio Macedo, Sujatha Duraikkannan & Jo Stichbury.
