One year of agentic AI: Six lessons from the people doing the work

Source: McKinsey / QuantumBlack (Lareina Yee, Michael Chui, Roger Roberts, with Stephen Xu); McKinsey article (PDF) · 2025-09

A year into the agentic AI revolution, one lesson is clear: It takes hard work to do it well.

An agentic enterprise transformation holds the promise of unmatched productivity. While some companies are enjoying early successes with such activities, many more are finding it challenging to see value from their investments. In some cases, they are even retrenching—rehiring people where agents have failed.

Introduction. Deploying agentic AI successfully isn’t easy. Here’s what we’re learning about how to get it right.

These stumbles are a natural evolution of any new technology, and we’ve seen this pattern before with other innovations. To understand the early lessons, we recently dug into more than 50 agentic AI builds we’ve led at McKinsey, as well as dozens of others in the marketplace. We’ve boiled down our analysis results to six lessons to help leaders successfully capture value from agentic AI (see sidebar “What is agentic AI?”).

Related work. What is agentic AI?

Agentic AI is a system based on gen AI foundation models that can act in the real world and execute multistep processes. AI agents can automate and perform complex tasks, often using natural language processing, that would normally require human effort (for more, see “What is an AI agent?,” from our McKinsey Explainers series).

Method. Agentic AI efforts that focus on fundamentally reimagining entire workflows—that is, the steps that involve people, processes, and technology—are more likely to deliver a positive outcome. Understanding how agents can help with each of these steps is the path to value. People will still be central to getting the work done, but now with different agents, tools, and automations to support them.

An important starting point in redesigning workflows is mapping processes and identifying key user pain points. This step is critical in designing agentic systems that reduce unnecessary work and allow agents and people to collaborate and accomplish business goals more efficiently and effectively. That collaboration can happen through learning loops and feedback mechanisms, creating a self-reinforcing system. The more frequently agents are used, the smarter and more aligned they become.

Before rushing into an agentic solution, business leaders should take stock of the task’s demands. In practice, that means getting clear on how standardized the process should be, how much variance the process needs to handle, and what portions of the work agents are best suited to do.

Developing effective agents is challenging work that requires harnessing individual expertise to create evaluations (or “evals”) and codifying best practices with sufficient granularity for given tasks. This codification serves as both the training manual and performance test for the agent, ensuring that it performs as expected.

These practices may exist in standard operating procedures or as tacit knowledge in people’s heads. When codifying practices, it’s important to focus on what separates top performers from the rest. For sales reps, this might include how they drive the conversation, handle objections, and match the customer’s style (see sidebar “Eval types”).

Another big lesson from our experience is that companies should be deliberate in redesigning work so that people and agents can collaborate well together. Without that focus, even the most advanced agentic programs risk silent failures, compounding errors, and user rejection.

Discussion. It’s not about the agent; it’s about the workflow Achieving business value with agentic AI requires changing workflows. Often, however, organizations focus too much on the agent or the agentic tool. This inevitably leads to greatlooking agents that don’t actually end up improving the overall workflow, resulting in underwhelming value.

Consider an alternative-legal-services provider that was working to modernize its contract review workflows. Legal reasoning in the company’s domain was constantly evolving, with new case law, jurisdictional nuances, and policy interpretations, making it challenging to codify expertise.

To account for natural variance, the team designed its agentic systems to learn within the workflow. Every user edit in the document editor, for example, was logged and categorized. This provided the engineers and data scientists with a rich stream of feedback, which they could then use to teach the agents, adjust prompt logic, and enrich the knowledge base. Over time, the agents could codify new expertise.

Focusing on the workflow instead of the agent enabled teams to deploy the right technology at the right point, which is especially important when reengineering complex, multistep workflows (exhibit). For example, insurance companies often have big investigative workflows that span multiple steps (such as claims handling and underwriting), with each step requiring different types of activities and cognitive tasks. Companies can redesign these types of workflows by thoughtfully deploying a targeted mix of rule-based systems, analytical AI, gen AI, and agents, all underpinned by a common orchestration framework (such as open-source frameworks like AutoGen, CrewAI, and LangGraph). In these cases, the agents are the orchestrators and the integrators, accessing tools and integrating outputs of other systems into their context. They are the glue that unifies the workflow so it delivers real closure with less intervention needed.

Agents aren’t always the answer AI agents can do a lot, but they shouldn’t necessarily be used for everything. Too often, leaders don’t look closely enough at the work that needs to be done or ask whether an agent would be the best choice to perform that work.

To help avoid wasted investments or unwanted complexity, business leaders can approach the role of agents much like they do when evaluating people for a high-performing team. The key question to ask is, “What is the work to be done and what are the relative talents of each potential team member—or agent—to work together to achieve those goals?” Business problems can often be addressed with simpler automation approaches, such as rules-based automation, predictive analytics, or large language model (LLM) prompting, which can be more reliable than agents out of the box.

One year of agentic AI: Six lessons from the people doing the work 3 On one level, these issues are straightforward. For example, low-variance, high-standardization workflows, such as investor onboarding or regulatory disclosures, tend to be tightly governed and follow predictable logic. In these cases, agents based on nondeterministic LLMs could add more complexity and uncertainty than value.

By contrast, high-variance, low-standardization workflows could benefit significantly from agents. For example, agents were deployed at a financial-services company to extract complex financial information, reducing the amount of human validation required and streamlining workflows. These tasks demanded information aggregation, verification checks, and compliance analysis—tasks where agents can be effective.

Conclusion. The world of AI agents is moving quickly, so we can expect to learn many more lessons. But unless companies approach their agentic programs with learning in mind (and in practice), they’re likely to repeat mistakes and slow their progress.