Why Accenture Calls for Knowledge Graphs to Build AI Trust

Investment in AI technologies reached record levels in 2024, with businesses across sectors seeking to transform operations through machine learning and automation systems.

However, the transition from pilot projects to production environments has exposed significant challenges.

Many organisations are struggling to demonstrate measurable returns on AI investments, while technical complexity and integration difficulties have slowed adoption rates.

As a result, the gap between AI’s theoretical capabilities and practical business applications has become a defining issue for technology leaders.

However, Accenture believes there is a way around these challenges with knowledge graphs.

These graphs represent information as interconnected networks of entities and relationships, with structured data systems enabling AI applications to access contextual information and maintain accuracy when processing complex business queries.

In response, major consulting firms, software vendors and technology providers are now emphasising integration services and data governance frameworks to support sustainable AI adoption – as well as prioritising reliability and trust over raw capability.

“If your AI strategy doesn’t start with data governance and semantic integration, you’re skipping a critical step,” says Kristalyn Warren Mumaw, Accenture’s AI Evangelist.

Physical AI emerging as a catalyst for autonomous operations

A concept called “Physical AI” gained prominence at this year’s Consumer Electronics Show, representing the convergence of foundational models with robotics systems.

This approach enables autonomous actions through natural language commands, moving beyond traditional robotic process automation (RPA) systems that require predetermined logic.

Chris McDivitt, Global Solution Lead for Autonomous Supply Chain at Accenture, illustrates the shift with a practical example: “You can ask a robot to ‘find all the green widgets in Aisle 3,’” he explains.

The system can execute this task even when physical locations change or when “green” refers to product names rather than colours.

The advancement is a departure from industrial controllers and RPA systems, which historically required users to issue commands with pre-programmed responses.

Physical AI systems can now process higher-level, goal-oriented instructions by combining foundational models’ planning capabilities with physical robots’ mobility and actuation.

For example, supply chain and manufacturing sectors face particular transformation opportunities – and this technology enables businesses to reconsider physical operations and workflows – introducing new options for human-machine interaction within these industries.

Accenture identifying trust as limiting factor for AI adoption

Trust emerges as the primary constraint on AI application expansion, according to Chris’ analysis – partly because user confidence varies significantly depending on the application context and potential consequences of incorrect responses.

Chris poses a series of questions to illustrate the trust gradient: “Would you trust ChatGPT to tell you why the sky is blue? How about diagnosing a medical condition? What about explaining why you're short 30 units for a critical outbound order?”

He says that the likelihood of accepting AI responses depends on two factors: confidence in the system’s accuracy and the cost of incorrect answers.

Users typically accept general knowledge responses but hesitate when specialist knowledge or current enterprise data is required.

Trust levels directly influence user engagement, the degree of autonomy granted to AI agents and the complexity of tasks assigned to Physical AI systems. 

This creates a feedback loop where trust limitations constrain AI utility, potentially limiting the technology’s business impact.

Human-AI collaboration models evolve towards strategic input

Current AI interactions predominantly involve users guiding chatbots or issuing commands to AI agents.

This model places humans in supervisory roles, making final decisions on action execution.

Chris predicts evolution towards strategic workflows where humans direct multi-agent systems to solve complex tasks.

This progression from operational to strategic human input requires technological maturation, including improvements to large language models (LLMs) and agent frameworks.

The transition also demands enhanced policies and guardrails alongside greater trust and proof that agents deliver appropriate outcomes.

Chris believes that these developments will redefine collaboration between humans and AI agents across enterprise environments – and that the future workforce must recognise AI as a dynamic system rather than a static tool. 

Increased user engagement generates additional data, improving AI performance and driving further engagement – and Chris identifies this virtuous cycle as the primary value driver, extending beyond initial AI model capabilities.

Digital Brain concept addressing enterprise knowledge requirements

To increase accuracy and trust, Accenture proposes placing AI agents within what it terms the “Digital Brain” environment.

This concept addresses the need for AI systems to access enterprise knowledge and current data when making decisions.

The approach recognises that plain LLMs, while capable of general reasoning, lack the specialist knowledge and current data required for enterprise applications.

For instance, medical diagnosis requires specialist medical knowledge, while operations diagnosis demands enterprise-specific knowledge and real-time data.

Chris says: “To increase accuracy, usefulness and subsequently trust of AI, we need to put AI agents in the right environment to operate in,” he says.

The Digital Brain concept is Accenture’s framework for addressing enterprise AI implementation challenges, particularly in supply chain and manufacturing contexts where accuracy and trust are essential for operational success.

“If you want to learn more about building secure, autonomous supply chains that enhance resilience and efficiency, let’s have a conversation,” he concludes.