Not 007, But a New Kind of AI Agent

Overview:

• The Generative Artificial Intelligence (AI) revolution is moving from Act I to Act II aided by key breakthroughs in Large Language Models (LLMs)
• New AI Agent use cases are more powerful, intelligent and autonomous
• Agentic AI will shift more innovation up the tech stack and software companies will be more critical in the AI value chain leading to a boost in AI spend capture.
• We think inflection starts to build in 2026 to 2027 as software incumbents evolve their architectures and packaging strategies and as enterprises shift spend from do-it-yourself (DIY) to off the shelf.

The TD Cowen Insight

Agentic AI is the new and more powerful wave of the revolutionary Generative Artificial Intelligence (GenAI) compute cycle. As use cases shift from AI Assistants to AI Agents, enterprise software stands to play a much more critical role. While AI spend at the software layer has been meager to date, we think the AI value capture will start to move up-stack, and we look for emerging inflection to take hold in 2026 to 2027.

Our Thesis

Given breakthroughs in reasoning and memory performance from frontier model providers, the Generative AI revolution is moving from Act I (AI Assistants) to Act II (AI Agents), promising to unlock much greater AI value creation. Agentic AI enables LLMs to perform higher-order thinking and work in a multistep, autonomous fashion, expanding beyond the simpler AI-powered query-answer chatbot apps. To effectively operate in enterprise settings, AI Agents need access to proprietary data and external tools. This is where software companies become much more critical in the AI value-chain, facilitating application programming interface (API) and function-calling, cross-enterprise data access and frameworks to manage and govern how Agents interact with enterprise services.

As the Agentic AI cycle builds over the next five to 10 years, we think this will put upward pressure on today's estimated US$820 billion in enterprise software spend as new digital AI workers consume more software and data, human productivity gains get re-invested back into tech and Agentic AI eats into the large estimated US$20 trillion in information technology services spend. As software vendors transform their architectures, pricing and packaging strategies to embrace Agentic AI, we expect the AI mix of software spend to rapidly rise. This will in turn drive up infrastructure consumption meters with implications for Cloud platform vendors and downstream compute and storage hardware vendors alike.

We think enterprise spend on Agentic AI will kick into a new gear in 2026 to 2027. To date, GenAI spend capture has been dominated by hyperscalers and frontier model providers. But given the relative complexity of operating new AI Agent apps, software companies will play a much bigger role in value-creation in this GenAI 2.0 cycle. We estimate spend on Agentic AI software to increase from US$3.4 billion in 2025 to US$51.5 billion in 2028 as the adoption curve gets off the ground.

As incumbents build out new Agentic capabilities, and as enterprises shift spend from DIY to packaged solutions, we think AI spend capture by public software companies will begin to inflect, which is a positive development for public software investors that are eager to find greater exposure to the AI movement.

What Is Proprietary?

We introduce new AI revenue estimates for key enterprise incumbents with established monetization strategies and highlight some preliminary views on AI revenue implications for two companies in the space. We also introduce our estimate of total Agentic AI software spend over the next three years as well as a forecast of total frontier model revenue, expected to grow at an astounding rate. Lastly, we discuss our view of the key foundational technologies needed to make Agentic AI effective in the enterprise and how we think software stack will evolve over time.

Financial & Industry Model Implications

We estimate that spend on Agentic AI software will increase from US$3.4 billion in 2025 to US$9.6 billion in 2026, heading to US$23.6 billion in 2027 and US$51.5 billion in 2028, representing an approximate 150% 3-yr compound annual growth rate (CAGR) as the S-curve for Agentic AI adoption begins to steepen in 2026 to 2027. We also estimate revenue from frontier model vendors will increase from approximately US$6.5 billion in 2024 to approximately US$191 billion in 2028, a remarkable estimated 133% CAGR at scale, which includes revenue from both enterprise and consumer use-cases.

What to Watch

We think key catalysts to watch include

• advancements in LLM reasoning capabilities and cost of deployment;
• compute workloads shifting from training to inference as production apps ramp up;
• vendors evolving pricing and packaging strategies from seat-based to usage-based models;
• application and data layers converging to power LLMs with source-agnostic enterprise data and
• enterprises shifting from internal development and DIY GenAI projects to packaged software solutions from incumbent or AI-first start-up vendors.

For Agentic AI to proliferate in the enterprise, software vendors will be tasked to ground LLMs with proprietary data, set guardrails for models and agents and manage cross-domain orchestration of agentic workflows. Such fast-paced changes in the landscape will likely drive more tuck-in mergers and acquisitions (M&A) activity. We think key targets could be in foundational tech areas like

• data stores (i.e. vector indexing, data discovery and data cataloging);
• data retrieval tools (i.e. retrieval-augmented generation (RAG), unstructured API connectors);
• memory and contextualization tools (i.e. semantic data governance and authentication);
• agent orchestration services (i.e. agentic planning and process management);
• model development tools (i.e. benchmarking, versioning and fine-tuning) and
• model deployment and management tools (i.e. evaluations, monitoring & Large Language Model Operations (LLMOps)).

In terms of gauging how quickly the slope of Agentic AI adoption will take hold, we list five key questions:

• Are enterprise data estates mature enough to effectively adopt AI Agents, and can they support Agents working across different systems?
• How long will it take for model accuracy and hallucination rates in enterprise settings to improve enough to be trusted in autonomously powering mission-critical workflows?
• When will buyers and sellers settle on the best pricing approach for Agentic AI use-cases, and what will the mix of consumption pricing be?
• Will companies use AI Agents from every point vendor, or will there be a battle for a new layer in the software stack with new crowned winners?
• What are the net impacts of the use of AI Agents when it comes to the organizational structure and culture, and how well can companies navigate these changes?