website SEOWebsite Traffic
This report is a collaborative effort by Alexander Sukharevsky, Dave Kerr, Klemens Hjartar, Lari Hämäläinen, Stéphane Bout, and Vito Di Leo, with Guillaume Dagorret, representing views from QuantumBlack, AI by McKinsey and McKinsey Technology.
by Arthur Mensch, CEO of Mistral AI
We’re at a moment when gen AI has entered every boardroom, but for many enterprises, it still lingers at the edges of actual impact. Many CEOs have greenlit experiments, spun up copilots, and created promising prototypes, but only a handful have seen the needle move on revenue or impact. This report gets to the heart of that paradox: broad adoption with limited return.
The current diagnosis is this: Today, AI is bolted on. But to deliver real impact, it must be integrated into core processes, becoming a catalyst for business transformation rather than a sidecar tool. Most deployments today use AI in a shallow way—as an assistant that sits alongside existing workflows and processes—rather than as a deeply integrated, engaged, and powerful agent of transformation.
Agentic AI is the catalyst that can make this transition possible, but doing so requires a strategy and a plan to successfully power that transformation. Agents are not simply magical plug-n-play pieces. They must work across systems, reason through ambiguity, and interact with people—not just as tools, but as collaborators. That means CEOs must ask different questions: not “How do we add AI?” but “How do we want decisions to be made, work to flow, and humans to engage in an environment where software can act?”
Redefining how decisions are made, how work is done, and how humans engage with technology requires alignment across goals, tools, and people. That alignment can only happen when openness, transparency, and control are central to your technology and implementation—when builders have an open, extensible, and observable infrastructure and users can easily craft and use agents with the confidence that the work of agents is safe, reliable, and under their control. That alignment creates the trust and effectiveness that is the currency of scalable transformation that delivers results rather than regrets.
The technology to build powerful agents is already here. The opportunity now is to deploy agents in ways that are deeply tied to how value is created and how people work. That requires an architecture that is modular and resilient and, more importantly, an operating model that centers on humans—not just as users but as co-architects of the systems they will be living and working with.
This report lays out the playbook not for tinkering but for reinvention. ROI comes from strong intent: define the outcomes, embed agents deep in core workflows, and redesign operating models around them. Organizations that win will pair a clear strategy with tight feedback loops and disciplined governance, using agents to rethink how decisions are made and how work gets done—and turning novelty into measurable value.
QuantumBlack, McKinsey’s AI arm, has been helping businesses create value from AI since 2009, expanding on McKinsey’s technology work over the past 30 years. QuantumBlack combines an industry-leading tech stack with the strength of McKinsey’s 7,000 technologists, designers, and product managers serving clients in more than 50 countries. With innovations fueled by QuantumBlack Labs—its center for R&D and software development—QuantumBlack delivers the organizational rewiring that businesses need to build, adopt, and scale AI capabilities.
Even before the advent of gen AI, artificial intelligence had already carved out a key place in the enterprise, powering advanced prediction, classification, and optimization capabilities. And the technology’s estimated value potential was already immense—between $11 trillion and $18 trillion globally2“The economic potential of generative AI: The next productivity frontier,” McKinsey, June 14, 2023.—mainly in the fields of marketing (powering capabilities such as personalized email targeting and customer segmentation), sales (lead scoring), and supply chain (inventory optimization and demand forecasting). Yet AI was largely the domain of experts. As a result, adoption across the rank and file tended to be slow. From 2018 to 2022, for example, AI adoption remained relatively stagnant, with about 50 percent of companies deploying the technology in just one business function, according to McKinsey research (Exhibit 1).
Gen AI has extended the reach of traditional AI in three breakthrough areas: information synthesis, content generation, and communication in human language. McKinsey estimates that the technology has the potential to unlock $2.6 trillion to $4.4 trillion in additional value on top of the value potential of traditional analytical AI.3“The economic potential of generative AI: The next productivity frontier,” McKinsey, June 14, 2023.
Two and a half years after the launch of ChatGPT, gen AI has reshaped how enterprises engage with AI. Its potentially transformative power lies not only in the new capabilities gen AI introduces but also in its ability to democratize access to advanced AI technologies across organizations. This democratization has led to widespread growth in awareness of, and experimentation with, AI: According to McKinsey’s most recent Global Survey on AI,4“The state of AI: How organizations are rewiring to capture value,” McKinsey, March 12, 2025. more than 78 percent of companies are now using gen AI in at least one business function (up from 55 percent a year earlier).
However, this enthusiasm has yet to translate into tangible economic results. More than 80 percent of companies still report no material contribution to earnings from their gen AI initiatives.5“The state of AI: How organizations are rewiring to capture value,” McKinsey, March 12, 2025. What’s more, only 1 percent of enterprises we surveyed view their gen AI strategies as mature.6Hannah Mayer, Lareina Yee, Michael Chui, and Roger Roberts, “Superagency in the workplace: Empowering people to unlock AI’s full potential,” McKinsey, January 28, 2025. Call it the “gen AI paradox”: For all the energy, investment, and potential surrounding the technology, at-scale impact has yet to materialize for most organizations.
Many organizations have deployed horizontal use cases, such as enterprise-wide copilots and chatbots; nearly 70 percent of Fortune 500 companies, for example, use Microsoft 365 Copilot.7Satya Nadella, “Microsoft Fiscal Year 2025 First Quarter Earnings Conference Call,” Microsoft, October 30, 2024. These tools are widely seen as levers to enhance individual productivity by helping employees save time on routine tasks and access and synthesize information more efficiently. But these improvements, while real, tend to be spread thinly across employees. As a result, they are not easily visible in terms of top- or bottom-line results.
By contrast, vertical use cases—those embedded into specific business functions and processes—have seen limited scaling in most companies despite their higher potential for direct economic impact (Exhibit 2). Fewer than 10 percent of use cases deployed ever make it past the pilot stage, according to McKinsey research.8New at McKinsey Blog, “McKinsey’s ecosystem of strategic alliances brings the power of generative AI to clients,” April 2, 2024. Even when they have been fully deployed, these use cases typically have supported only isolated steps of a business process and operated in a reactive mode when prompted by a human, rather than functioning proactively or autonomously. As a result, their impact on business performance also has been limited.
What accounts for this imbalance? For one thing, horizontally deployed copilots such as Microsoft Copilot or Google AI Workspace are accessible, off-the-shelf solutions that are relatively easy to implement. (In many cases, enabling Microsoft Copilot is as simple as activating an extension to an existing Office 365 contract, requiring no redesign of workflows or major change management efforts.) Rapid deployment of enterprise chatbots also has been driven by risk mitigation concerns. As employees began experimenting with external large language models (LLMs) such as ChatGPT, many organizations implemented internal, secure alternatives to limit data leakage and ensure compliance with corporate security policies.
The limited deployment and narrow scope of vertical use cases can in turn be attributed to six primary factors:
Despite its limited bottom-line impact so far, the first wave of gen AI has been far from wasted. It has enriched employee capabilities, enabled broad experimentation, accelerated AI familiarity across functions, and helped organizations build essential capabilities in prompt engineering, model evaluation, and governance. All of this has laid the groundwork for a more integrated and transformative second phase—the emerging age of AI agents.10Lareina Yee, Michael Chui, Roger Roberts, and Stephen Xu, “Why agents are the next frontier of generative AI,” McKinsey Quarterly, July 24, 2024.
LLMs have revolutionized how organizations interact with data—enabling information synthesis, content generation, and natural language interaction. But despite their power, LLMs have been fundamentally reactive and isolated from enterprise systems, largely unable to retain memory of past interactions or context across sessions or queries. Their role has been largely limited to enhancing individual productivity through isolated tasks. AI agents mark a major evolution in enterprise AI—extending gen AI from reactive content generation to autonomous, goal-driven execution. Agents can understand goals, break them into subtasks, interact with both humans and systems, execute actions, and adapt in real time—all with minimal human intervention. They do so by combining LLMs with additional technology components providing memory, planning, orchestration, and integration capabilities.
With these new capabilities, AI agents expand the potential of horizontal solutions, upgrading general-purpose copilots from passive tools into proactive teammates that don’t just respond to prompts but also monitor dashboards, trigger workflows, follow up on open actions, and deliver relevant insights in real time. But the real breakthrough comes in the vertical realm, where agentic AI enables the automation of complex business workflows involving multiple steps, actors, and systems—processes that were previously beyond the capabilities of first-generation gen AI tools.
On the operations side, agents take on routine, data-heavy tasks so humans can focus on higher-value work. But they go further, transforming processes in five ways:
In a complex supply chain environment, for example, an AI agent could act as an autonomous orchestration layer across sourcing, warehousing, and distribution operations. Connected to internal systems (such as the supply chain planning system or the warehouse management system) and external data sources (such as weather forecasts, supplier feeds, and demand signals), the agent could continuously forecast demand. It could then identify risks, such as delays or disruptions, and dynamically replan transport and inventory flows. Selecting the optimal transport mode based on cost, lead time, and environmental impact, the agent could reallocate stock across warehouses, negotiate directly with external systems, and escalate decisions requiring strategic input. The result: improved service levels, reduced logistics costs, and lower emissions.
Agents can also help spur top-line growth by amplifying existing revenue streams and unlocking entirely new ones:
In short, agentic AI doesn’t just automate. It redefines how organizations operate, adapt, and create value.
The following case studies demonstrate how QuantumBlack helps organizations build agent workforces—with outcomes that extend far beyond efficiency gains.
The problem: A large bank needed to modernize its legacy core system, which consisted of 400 pieces of software—a massive undertaking budgeted at more than $600 million. Large teams of coders tackled the project using manual, repetitive tasks, which resulted in difficulty coordinating across silos. They also relied on often slow, error-prone documentation and coding. While first-generation gen AI tools helped accelerate individual tasks, progress remained slow and laborious.
The agentic approach: Human workers were elevated to supervisory roles, overseeing squads of AI agents, each contributing to a shared objective in a defined sequence (Exhibit 3). These squads retroactively document the legacy application, write new code, review the code of other agents, and integrate code into features that are later tested by other agents prior to delivery of the end product. Freed from repetitive, manual tasks, human supervisors guide each stage of the process, enhancing the quality of deliverables and reducing the number of sprints required to implement new features.
Impact: More than 50 percent reduction in time and effort in the early adopter teams
The problem: A market research and intelligence firm was devoting substantial resources to ensure data quality, relying on a team of more than 500 people whose responsibilities included gathering data, structuring and codifying it, and generating tailored insights for clients. The process, conducted manually, was prone to error, with a staggering 80 percent of mistakes identified by the clients themselves.
The agentic approach: A multiagent solution autonomously identifies data anomalies and explains shifts in sales or market share. It analyzes internal signals, such as changes in product taxonomy, and external events identified via web searches, including product recalls or severe weather. The most influential drivers are synthesized, ranked, and prepared for decision-makers. With advanced search and contextual reasoning, the agents often surface insights that would be difficult for human analysts to uncover manually. While not yet in production, the system is fully functional and has demonstrated strong potential to free up analysts for more strategic work.
Impact: More than 60 percent potential productivity gain and expected savings of more than $3 million annually.
The problem: Relationship managers (RMs) at a retail bank were spending weeks writing and iterating credit-risk memos to help make credit decisions and fulfill regulatory requirements (Exhibit 4). This process required RMs to manually review and extract information from at least ten different data sources and develop complex nuanced reasoning across interdependent sections—for instance, loan, revenue, and cash joint evolution.
The agentic approach: In close collaboration with the bank’s credit-risk experts and RMs, a proof of concept was developed to transform the credit memo workflow using AI agents. The agents assist RMs by extracting data, drafting memo sections, generating confidence scores to prioritize review, and suggesting relevant follow-up questions. In this model, the analyst’s role shifts from manual drafting to strategic oversight and exception handling.
Impact: A potential 20 to 60 percent increase in productivity, including a 30 percent improvement in credit turnaround
Realizing AI’s full potential in the vertical realm requires more than simply inserting agents into legacy workflows. It instead calls for a shift in design mindset—from automating tasks within an existing process to reinventing the entire process with human and agentic coworkers. That’s because when agents are embedded into a legacy process without redesign, they typically serve as faster assistants—generating content, retrieving data, or executing predefined steps. But the process itself remains sequential, rule bound, and shaped by human constraints.
Reinventing a process around agents means more than layering automation on top of existing workflows—it involves rearchitecting the entire task flow from the ground up. That includes reordering steps, reallocating responsibilities between humans and agents, and designing the process to fully exploit the strengths of agentic AI: parallel execution that collapses cycle time, real-time adaptability that reacts to changing conditions, deep personalization at scale, and elastic capacity that flexes instantly with demand.
Consider a hypothetical customer call center. Before introducing AI agents, the facility was using gen AI tools to assist human support staff by retrieving articles from knowledge bases, summarizing ticket histories, and helping draft responses. While this assistance improved speed and reduced cognitive load, the process itself remained entirely manual and reactive, with human agents still managing every step of diagnosis, coordination, and resolution. The productivity improvement potential was modest, typically boosting resolution time and productivity between 5 and 10 percent.
Now imagine that the call center introduces AI agents but largely preserves the existing workflow—agents are added to assist at specific steps without reconfiguring how work is routed, tracked, or resolved end-to-end. Agents can classify tickets, suggest likely root causes, propose resolution paths, and even autonomously resolve frequent, low-complexity issues (such as password resets). While the impact here can be increased—an estimated 20 to 40 percent savings in time and a 30 to 50 percent reduction in backlog—coordination friction and limited adaptability prevent true breakthrough gains.
But the real shift occurs at the third level, when the call center’s process is reimagined around agent autonomy. In this model, AI agents don’t just respond—they proactively detect common customer issues (such as delayed shipments, failed payments, or service outages) by monitoring patterns across channels, anticipate likely needs, initiate resolution steps automatically (such as issuing refunds, reordering items, or updating account details), and communicate directly with customers via chat or email. Human agents are repositioned as escalation managers and service quality overseers, who are brought in only when agents detect uncertainty or exceptions to typical patterns. Impact at this level is transformative. This could allow a radical improvement of customer service desk productivity. Up to 80 percent of common incidents could be resolved autonomously, with a reduction in time to resolution of 60 to 90 percent (Exhibit 5).
Of course, not every business process requires full reinvention. Simple task automation is sufficient for highly standard, repetitive workflows with limited variability—such as payroll processing, travel expense approvals, or password resets—where gains come primarily from reducing manual effort. In contrast, processes that are complex, cross-functional, prone to exceptions, or tightly linked to business performance often warrant full redesign. Key indicators that call for reinvention include high coordination overhead, rigid sequences that delay responsiveness, frequent human intervention for decisions that could be data driven, and opportunities for dynamic adaptation or personalization. In these cases, redesigning the process around the agent’s ability to orchestrate, adapt, and learn delivers far greater value than simply speeding up existing workflows.
To scale agents, companies will need to overcome a threefold challenge: handling the newfound risks that AI agents bring, blending custom and off-the-shelf agentic systems, and staying agile amid fast-evolving tech (while avoiding lock-ins).
These challenges cannot be addressed by merely bolting new components, such as memory stores or orchestration engines, on top of existing gen AI stacks. While such capabilities are necessary, they are not sufficient. What’s needed is a fundamental architectural shift: from static, LLM-centric infrastructure to a dynamic, modular, and governed environment built specifically for agent-based intelligence—the agentic AI mesh.
The agentic AI mesh is a composable, distributed, and vendor-agnostic architectural paradigm that enables multiple agents to reason, collaborate, and act autonomously across a wide array of systems, tools, and language models—securely, at scale, and built to evolve with the technology. At the heart of this paradigm are five mutually reinforcing design principles:
The emerging architecture for agentic AI relies on seven interconnected capabilities:
The agentic AI mesh acts as the connective and orchestration layer that enables large-scale, intelligent agent ecosystems to operate safely and efficiently, and continuously evolve. It allows companies to coordinate custom-built and off-the-shelf agents within a unified framework, support multiagent collaboration by allowing agents to share context and delegate tasks, and mitigate key risks such as agent sprawl, autonomy drift, and lack of observability—all while preserving the agility required for a rapid technology evolution (see sidebar “Seven interconnected capabilities of the AI agentic mesh”).
A few characteristics are key for LLM providers to take into account in the agentic era:
Beyond this architectural evolution, organizations will also have to revisit their LLM strategies. At the core of every custom agent lies a foundation model—the reasoning engine that powers perception, decision-making, and interaction. In the agentic era, the requirements placed on LLMs evolve significantly. Agents are not passive copilots—they are autonomous, persistent, embedded systems. This creates five critical categories of LLM requirements, each aligned with specific deployment contexts, for which different kinds of models will be relevant (see sidebar “Foundational models for agents: Five requirements”).
Finally, to truly scale agent deployment across the enterprise, the enterprise systems themselves must also evolve.
In the short term, APIs—protocols that allow different software applications to communicate and exchange data—will remain the primary interface for agents to interact with enterprise systems. But in the long term, APIs alone will not suffice. Organizations must begin reimagining their IT architectures around an agent-first model—one in which user interfaces, logic, and data access layers are natively designed for machine interaction rather than human navigation. In such a model, systems are no longer organized around screens and forms but around machine-readable interfaces, autonomous workflows, and agent-led decision flows.
This shift is already underway. Microsoft is embedding agents into the core of Dynamics 365 and Microsoft 365 via Copilot Studio; Salesforce is expanding Agentforce into a multiagent orchestration layer; SAP is rearchitecting its Business Technology Platform (BTP) to support agent integration through Joule. These changes signal a broader transition: The future of enterprise software is not just AI-augmented—it is agent-native.
As agents evolve from passive copilots to proactive actors—and scale across the enterprise—the complexity they introduce will be not only technical but mostly organizational. The real challenge lies in coordination, judgment, and trust. This organizational complexity will play out most visibly across three dimensions: how humans and agents cohabit day-to-day workflows; how organizations establish governance over systems that can act autonomously; and how they prevent unchecked sprawl as agent creation becomes increasingly democratized.
Agents unlock the full potential of vertical use cases, offering companies a path to generate value well beyond efficiency gains. But realizing that potential requires a reimagined approach to AI transformation—one tailored to the unique nature of agents and capable of addressing the lingering limitations they alone cannot resolve. This approach is the subject of our next chapter.
Unlike gen AI tools that could be easily plugged into existing workflows, AI agents demand a more foundational shift, one that requires rethinking business processes and enabling deep integration with enterprise systems. McKinsey has a proven Rewired playbook for AI-driven transformations.11Eric Lamarre, Kate Smaje, and Rodney Zemmel, “Rewired to outcompete,” McKinsey Quarterly, June 20, 2023. To capitalize on the agentic opportunity, organizations must build on that, fundamentally reshaping their AI transformation approach across four dimensions:
Redesigning the approach to AI transformation is an important step, but it is not enough. To unlock their full potential at scale, organizations must also activate a robust set of enablers that support the structural, cultural, and technical shifts required to integrate agents into day-to-day operations. These enablers span four dimensions—people, governance, technology architecture, and data—each of which is a foundation for scalable, secure, and high-impact deployment of agents across the enterprise.
The rise of AI agents is more than just a technological shift. Agents represent a strategic inflection point that will redefine how companies operate, compete, and create value. To navigate this transition successfully, organizations must move beyond experimentation and pilot programs and enter a new phase of scaled, enterprise-wide transformation.
This pivot cannot be delegated—it must be initiated and led by the CEO. It will rely on three key actions:
Like any truly disruptive technology, AI agents have the power to reshuffle the deck. Done right, they offer laggards a leapfrog opportunity to rewire their competitiveness. Done wrong—or not at all—they risk accelerating the decline of today’s market leaders. This is a moment of strategic divergence.
While the technology will continue to evolve, it is already mature enough to drive real, transformative change across industries. But to realize the full promise of agentic AI, CEOs must rethink their approach to AI transformation—not as a series of scattered pilots but as focused, end-to-end reinvention efforts. That means identifying a few business domains with the highest potential and pulling every lever: from reimagining workflows to redistributing tasks between humans and machines to rewiring the organization based on new operating models.
Some leaders are already moving—not just by deploying fleets of agents but by rewiring their organizations to harness their full disruptive potential. (Moderna, for example, merged its HR and IT leadership13Julien Dupont-Calbo, “L’IA n’est plus un outil, c’est un collègue”: Moderna fusionne sa DRH et sa DSI, [“AI is no longer a tool, it’s a colleague”: Moderna merges its HR and IT departments], Les Echos, May 15, 2025.—signaling that AI is not just a technical tool but a workforce-shaping force.) This is a structural move toward a new kind of enterprise. Agentic AI is not an incremental step—it is the foundation of the next-generation operating model. CEOs who act now won’t just gain a performance edge. They will redefine how their organizations think, decide, and execute.
The time for exploration is ending. The time for transformation is now.
Alexander Sukharevsky is a senior partner in McKinsey’s London office, where Dave Kerr is a partner; Klemens Hjartar is a senior partner in the Copenhagen office; Lari Hämäläinen is a senior partner in the Seattle office; Stéphane Bout is a senior partner in the Lyon office; Vito Di Leo is a partner in the Zurich office; and Guillaume Dagorret is a senior fellow with the McKinsey Global Institute and is based in the Paris office.
The authors wish to thank Alena Fedorenko, Annie David, Clarisse Magnin, Lareina Yee, Larry Kanter, Michael Chui, Roger Roberts, Sarah Mulligan, Thomas Vlot, and Timo Mauerhoefer for their contributions to this report.
This article was edited by Larry Kanter, a senior editor in the New York office.
