State of the Art of Agentic AI Transformation

Artificial intelligence continues to surge ahead at an unprecedented pace—catapulting some companies forward and leaving others far behind.

In 2023 and 2024, tech-forward enterprises broke through the pilot phase, achieving 10% to 25% EBITDA gains by scaling information retrieval and single-task AI. In doing so, they established a repeatable playbook that others can now follow—grounded in robust methodology, analytic tools, and clear benchmarks.

Yet a year on, most organizations remain stuck in experimentation mode, satisfied with minor productivity gains that haven’t delivered significant value. But as the leaders move forward, falling behind risks ceding competitive ground that may be difficult to recover.

And things are about to get even more interesting: In the first half of 2025, major players—including Anthropic, Alphabet, Microsoft, OpenAI, Salesforce, and others—debuted their visions of agentic AI. Tech-forward enterprises are already turning their focus from automating tasks to redesigning entire workflows. Early adopters are figuring out how agents will coexist safely, discover each other, gather the context and data they need, and collaborate productively. As the vision meets reality, they’ll also wrestle with data silos, informal context, intellectual property, privacy, security, and vendor profit motives.

A purist view of architecture won’t meet the moment. Given the current pace of AI innovation, the most likely scenario is that we’ll see rapid, fitful, and hard-to-predict progress. Companies should maintain an architectural North Star but sustain progress with fit-for-purpose, domain-specific, and human-in-the-loop builds for the foreseeable future.

Maximizing value demands focused attention on a few key priorities:

• Keep up the pace. The most important tasks remain redesigning processes and workflows while cleaning up data. Falling further behind is dangerous.
• Follow the taillights of the enterprise leaders. The playbook is established, with methodology, analytic tools, and benchmarks available.     
• Take a principled but flexible view of architecture. Balance long-term vision with flexible, domain-specific solutions to keep pace with AI’s progress. 

Leaders cracked the code

In our 2023 Tech Report, we wrote about how some leaders were beginning to unlock the secrets of AI productivity (see “You’re Out of Time to Wait and See on AI”). In 2024, we could already see clear patterns about where AI ROI could reliably be captured (see “Five Functions Where AI Is Already Delivering”). While transformation styles varied, these early adopters have delivered a roadmap with proven methodology, analytic tools, and benchmarks. We can summarize this roadmap with five critical actions:

• Set ambitious goals based on top-down diagnostics, not trials and pilots.
• Charge general managers with meeting these targets, not the CIO or CTO.
• Redesign entire workflows, not siloed activities or use cases.
• Curate and clean the data and application environment as needed, not holistically.
• Make, buy, or partner to build capabilities for each major workflow, rather than waiting for enterprise-wide solutions.

The biggest insight from these transformations is that the most important aspects of the transformation are process redesign and cleaning up the data and application environment. Because of this, it doesn’t make sense to wait for the dust to settle on technology. There’s no way to cut corners on process, data, and application cleanup. Every day a company waits is another day it’s left behind.

AI leaders pursue agentic AI

AI innovation increasingly focuses on enabling models to work with much more complex reasoning, context, and unstructured data while communicating with SaaS applications and other agents (see Figure 1). 

Figure 1

As noted earlier, many tech giants debuted their visions of agentic AI in the first half of 2025. While flavors vary, the underlying progression of capabilities crystallizes into four levels.    

• Level 1:  LLM-powered information retrieval agents (e.g., knowledge assistants, copilots)
• Level 2:  Single-task agentic workflows (e.g., task-doers with self-contained action loops)
• Level 3:  Cross-system agentic workflow orchestration (e.g., complex workflow execution, supervised agents)
• Level 4:  Multi-agent constellations (e.g., any-to-any agent, loosely coupled collaborative agents) 

Tech-forward companies scaled Level 1 tools in 2023 and 2024, with varying degrees of success. When deployed in a diffuse way, they delivered microproductivity—what we might call “grab-a-coffee" time-savers. But when deeply embedded in functional workflows in areas such as sales, development, and product management, the gains compounded—especially after heavy data cleaning and curation as well as continuous high-quality governance.

Levels 2 and 3 are now where capital, innovation, and deployment velocity are converging (see Figure 2).  Level 4 is on the whiteboard, held back by several practical realities outlined in the next section.

Figure 2

Practical, not purist architecture

In 2025, leading tech companies turned their attention to making single- and multi-system workflows smarter, powered by agents.

As vendors race to bring their agentic visions to life, enterprise teams face new challenges: How will these agents operate safely, find and connect with each other, gather the context and data they need, and collaborate productively as the vision meets the reality of vendor profit motives, data silos, informal context, enterprise data, IP, and security?

As higher levels of agentic autonomy are pursued, enterprises are encountering a number of thorny issues.

• Human work: Most work happens across multiple systems and organizations, with context and informal processes.
• Technology gaps: These include a lack of communication standards (Model Context Protocol, or MCP, isn’t USB) and compounding errors in multistep tasks.
• Enterprise reality: Data isn’t clean, and privacy, security, and intellectual property are real concerns.
• Vendor motives: These run counter to open standards, shared IP, workflows, and data, leading to battles and walled gardens.

Several architectural visions exist for how higher levels of agentic autonomy can be enabled. Most call for an interconnected fabric or mesh that will register, distribute, and allow communication between agents to enable secure collaboration. Some visions are more libertarian than others, but most resemble Web 3.0: a logical vision for how things should work if no one were greedy and governance and accountability were not thorny issues. Much like Web 3.0, we expect these visions to serve as a useful aspiration, but we don’t expect them to survive contact with enterprise reality unchanged.

For this reason, a rigid approach to architecture falls short of what’s needed. With AI moving at breakneck speed, progress is likely to be rapid, uneven, and tough to forecast. Walled gardens will take the lead. Fit-for-purpose custom builds will dominate enterprise-wide architectures for some time. Human-in-the-loop applications are likely the pragmatic reality for years. Context and graph analytics will remain closely guarded assets. Standards battles will play out at lightning pace (witness the MCP and Agent-to-Agent, or A2A, adoption tipping points); incumbents may try to selectively open source their IP; and domain-specific leaders will emerge.  

Now is the moment to act

Maximizing value in this next phase demands disciplined focus on a small set of high-impact priorities. The organizations that move decisively will extend their lead; those that hesitate risk being left behind.

• Keep up the pace. The critical work remains redesigning processes and workflows while cleaning and standardizing data. Any further delay compounds technical debt and makes catching up much harder.
• Follow the taillights of enterprise leaders. The path forward is well-mapped. Proven playbooks, tested methodologies, advanced analytic tools, and benchmark data are already available. Focus on them to accelerate progress rather than reinventing the wheel.
• Take a principled but flexible view of architecture. Expect domain-specific platforms, not one-size-fits-all enterprise systems—for example, tailored solutions for supply chain, sales, and other key domains. Plan for human-in-the-loop oversight for now—think Iron Man suits, not fully autonomous Iron Man robots. Finally, select vendors strategically to limit (or at least balance) agent lock-in and preserve optionality for future evolution.