Deloitte Predicts Generative AI Adoption by 2027

Deloitte has projected that by 2027, up to 50% of companies utilizing generative artificial intelligence will have initiated agentic artificial intelligence pilots or proofs of concept. This anticipated wave of adoption is significant enough that the term "agentic" is increasingly being used to describe various systems that incorporate large language model calls, ranging from simple pipelines to fully autonomous systems.

The article distinguishes between agentic workflows and autonomous agents, emphasizing the importance of control flow ownership. In agentic workflows, a human writes the code in advance, while autonomous agents are systems that make decisions on their own in real-time. The distinction is crucial to avoid over-engineering simple tasks or under-engineering complex problems.

Anthropic has drawn a foundational line between workflows and agents in their widely referenced document, "Building Effective Agents." Workflows are defined as systems where large language models and tools are coordinated through predefined paths, whereas agents dynamically control their processes and tool usage.

The article further explores the spectrum of deterministic workflows to autonomous multi-agent systems, with examples illustrating the differences in control flow. It argues that predictability versus autonomy is a more relevant framework than merely considering whether a system uses artificial intelligence.

Deterministic workflows are characterized by known sequences of steps decided at design time by humans, where an LLM can perform tasks without influencing subsequent actions. In contrast, orchestrated workflows depend on runtime decisions while still being predefined by humans. This leads to a clearer understanding of how systems operate and where autonomy truly begins.

The ReAct pattern, introduced by Yao et al. in 2022, marks the beginning of true autonomy. This model allows systems to decide actions based on observations from previous actions, without predefined paths for every scenario. The agent engages in an iterative loop until certain conditions are met, showcasing the dynamic nature of genuine autonomy in artificial intelligence applications.