AI Agents Explained: Insights from McKinsey’s Latest Research

• What AI agents really are: McKinsey AI agents represent a shift from AI that answers questions to AI that plans, acts, and completes end-to-end work across enterprise systems.
• Where adoption stands: Most companies use generative AI, but only a fraction have scaled agents, despite trillions in potential value concentrated in customer operations, engineering, sales, and R&D.
• Where agents deliver real impact: The fastest gains appear in support, IT, engineering, operations, and HR, where agents cut cycle time, reduce costs, and improve quality.
• What’s coming next: Enterprises are moving toward multi-agent systems and stronger governance, and platforms like Ema make production-grade agent deployment practical.

A year ago, most companies were testing generative AI in small, low-risk ways, chat interfaces, copilots, and summarizers. Helpful, but still assistants. They sped up tasks; they didn’t run workflows. What’s changed is expectation: leaders now want systems that can reason, plan, take action, and complete work independently.

That change is what puts AI agents at the center of the enterprise conversation. Three forces accelerated this moment:

• Foundation models now understand complex instructions and reasoning.
• Tool integration enables real actions across CRMs, ERPs, browsers, and internal systems.
• Enterprise pressure for real productivity gains has intensified.

This combination explains why McKinsey AI agents have become a major topic in enterprise AI. McKinsey’s research also outlines what this shift requires:

• Greater value potential, as automation extends from tasks to full outcomes.
• Higher operational complexity, because agents interact with core workflows and systems.
• Broader organizational change, as processes, governance, and talent must adapt.

McKinsey defines an AI agent as a software actor that can take a goal, break it into tasks, and carry out those tasks across systems. Instead of responding to prompts, agents plan, execute, and follow through, much closer to how real employees work.

This shift from “responding” to “doing” is the core difference.

1. Agents Don’t Just Respond, They Act.

Chatbots answer questions. Copilots suggest drafts. Agents complete work. They log into systems, fetch and compare data, send follow-ups, update records, and verify outcomes. They own the workflow instead of handing it back to a human.

2. The Manager–Specialist Agent Model.

McKinsey often describes agent networks as small digital teams:

• A manager agent interprets the goal
• Specialist agents handle subtasks
• A shared context layer delivers data, tools, and governance

It mirrors how real teams function. One agent coordinates, others execute, and everyone shares the same information.

3. Agents Close Loops.

This is the real breakthrough. Agents don’t stop at generating output; they drive a task to its end state. They track progress, adjust steps when something breaks, and escalate only when needed.

Earlier AI systems were reactive. Agents are proactive: they plan, act, and deliver outcomes across CRMs, ERPs, browsers, and internal tools.

This is why McKinsey sees AI agents as the bridge from simple task augmentation to full workflow execution. They convert generative AI from a content tool into an operational engine that reduces handoffs, shortens cycles, and produces measurable results. So, how do these agents actually work under the hood?

McKinsey frames this using a simple lifecycle that shows how agents move from a goal to execution inside a real business environment.

The flow typically looks like this:

1. A user or system assigns a goal: The agent receives an outcome to deliver, resolving a support issue, generating a report, updating records, or triggering a workflow. It isn’t following a script; it chooses the steps needed to reach the goal.

2. A manager agent creates the plan: It breaks the request into subtasks, identifies the right tools, and sequences the workflow. This planning step is what separates agents from traditional rule-based automation.

3. Specialist agents execute the work: Sub-agents handle individual actions: calling APIs, pulling data, navigating internal systems, updating apps, sending alerts, or coordinating with other agents.

4. The system evaluates and refines the output: Feedback loops guide reliability. If something fails, the agent retries or adjusts. If confidence drops, it escalates to a human. This keeps outcomes consistent and safe.

Behind this lifecycle is the technical stack that enables agentic behavior:

• A foundation model or LLM for reasoning, interpretation, and planning
• Tooling and integrations to act across CRMs, ERPs, ticketing systems, browsers, and internal APIs
• A memory and context layer storing history, task state, and business data
• An orchestration and policy layer governing permissions, boundaries, and compliance

Together, these components allow agents to participate directly in operational workflows, not just generate responses.

This is the architecture modern platforms like Ema are built around. Ema’s Universal AI Employee connects to existing systems, Salesforce, Zendesk, internal APIs, and uses the same planning, tool-use, and governance layers described above. This lets enterprises deploy agents that don’t just answer questions but actually carry out work across workflows without modifying their tech stack.

Once you see how the architecture works, it becomes easier to understand the different types of agents McKinsey believes leaders should focus on.

The Five Levels of Enterprise AI Agents

McKinsey doesn’t view AI agents as a single category. They map out a range, from basic helpers to highly autonomous digital workers. This matters because companies move through these stages gradually, and each step unlocks more value.

Here are the five levels:

Across all five levels, it’s clear that enterprises will run a mixed portfolio of agents. Different teams need different capabilities, and value accelerates when these agents start working together across workflows.

Now the question becomes: how far have organizations actually reached, and where is value showing up today?

What McKinsey’s Data Reveals About AI Adoption

McKinsey’s research shows a clear pattern: generative AI is widely adopted, but true scale remains limited. 88% of organizations use AI in at least one function, yet only a third have embedded it deeply enough to change core workflows or measurable performance.

AI agents follow the same pattern:

• 23% of companies are scaling agentic AI in at least one function.
• 39% are experimenting, mostly in customer operations, IT, and knowledge workflows.

McKinsey ties this adoption picture to the broader economic potential of generative AI, estimating $2.6–$4.4 trillion in annual value. Around 75% of that value sits in customer operations, sales and marketing, software engineering, and R&D, areas with high-volume, repeatable, and measurable knowledge work. These are also the functions where agents are gaining the fastest traction.

A key theme is the “gen AI paradox”: big investment, limited financial impact. Pilots exist everywhere, but scalable value is rare. Agents help solve this by automating full workflows rather than isolated tasks.

When deployed at scale, McKinsey estimates agents can deliver 3% to 5% annual productivity improvement. The implication for leaders is simple: the value is real, but only if processes, governance, data access, and workflows evolve alongside the technology.

These adoption patterns become clearer when you look at where enterprises are already deploying agents and seeing meaningful returns.

AI agents create value by automating work that’s too complex or variable for traditional systems. Many processes have endless permutations, like planning a business trip with changing flights, hotel rules, approvals, and constraints. Rule-based automation breaks in these situations, which is why humans still carry most of the load. Agents close that gap.

Their impact comes from three core strengths:

• They handle complex, unpredictable scenarios.: Foundation models let agents adapt to unusual inputs, incomplete data, and non-linear workflows, where traditional automation fails.
• They can be guided with plain language: Teams can describe tasks without converting every step into code, making automation faster to build and easier for domain experts to influence.
• They work across existing tools: Agents interact with enterprise apps, browsers, APIs, and documents, reducing the integration overhead of older automation approaches.

These capabilities tie directly to measurable business gains. The long-term impact is even broader. Agents help modernize tech stacks by supporting modular architectures, updated frameworks, and more efficient cloud environments. With multiple agents collaborating and learning from human feedback, companies can automate entire processes, not just isolated tasks.

Some results are already visible. In customer service, organizations using AI agents have seen 14% higher issue resolution per hour and 9% lower handling time. These improvements strengthen customer experience and open the door to new models, such as premium human-assisted service powered by smarter automation.

AI agents aren’t appearing everywhere at once. They succeed first in functions where workflows are structured, volumes are high, and outcomes can be measured. Here’s how companies are using AI Agents:

1. Customer Operations and CX

• Agents act as autonomous support handlers, triage engines, or real-time assistants.
• They resolve issues, fetch data across systems, draft responses, and close simpler tickets independently.
• Impact shows up in resolution rates, handle time, CSAT, and deflection.
• Improvements often appear within weeks because the workflows are high-volume and predictable.

2. IT, Service Desk, And Knowledge Management

• IT teams face repeatable, ticket-heavy tasks, ideal for agentic automation.
• Agents provide first-line support by answering common questions, resolving access issues, and routing incidents.
• They surface knowledge articles instantly and reduce queue times.
• Companies see faster resolutions and a meaningful drop in cost per ticket.

3. Software Engineering and DevOps

• Engineering workflows are structured but time-consuming.
• Agents help with code generation, test creation, documentation, deployment steps, and incident routing.
• They orchestrate tools, analyze logs, prep patches, and coordinate responses, not just suggest code.
• Key metric for leaders: velocity gains across the development pipeline.

4. Operations, Finance, and Risk

• Regulated industries benefit from agents that combine automation with structured reasoning.
• Common examples: underwriting, claims processing, loan reviews, KYC tasks, and audit workflows.
• Gains show up in cycle time, error reduction, analyst capacity, and throughput.
• Agents handle heavy data gathering and validation, so humans focus on exceptions.

5. HR and Employee Experience

• HR handles high-volume, rule-based activities like onboarding, policy queries, and document creation.
• Agents act as 24/7 support for employees and streamline back-office workflows.
• The impact is visible in turnaround time, employee satisfaction, and reduced administrative load.

Across all these functions, measurement is the constant. Cycle time, cost per ticket, hours saved, quality, and error rates decide whether an agent moves from pilot to production. And that performance depends less on the LLM itself and more on design, governance, and how deeply the agent connects to real workflows.

But measurable outcomes depend on more than deployment frequency. They depend on how well an agent is designed for enterprise realities.

Great agents aren’t defined by what they can show in a demo, but by how well they handle real workloads, real data, and real constraints inside the enterprise.

Here are the essentials that define an enterprise-ready agent:

1. Deep context & data access: Agents need rich context from both structured data (CRM, ERP, tickets) and unstructured sources (documents, chats, knowledge bases). Without it, they guess; with it, they reason.

2. Multi-system integration & orchestration: Enterprise work spans many tools. Agents must read, write, and act across CRMs, ERPs, browsers, APIs, and ticketing systems while coordinating steps end to end.

3. Autonomous reasoning with guardrails: Agents should plan and adapt, but within clearly defined limits. Boundaries keep actions safe, consistent, and compliant.

4. Security, compliance, & governance-first: Role-based access, encryption, audit trails, and data controls are essential. These fundamentals earn trust from IT, risk, and compliance teams.

5. Human-in-the-loop & oversight: Reliable systems spell out when agents act alone and when humans step in. Reviewer agents or approval steps help ensure accuracy.

6. Built-in measurement: Cycle time, accuracy, deflection, cost per ticket, and quality must be tracked from day one. Measurement turns automation into proven ROI.

Together, these principles turn agents from conversational tools into dependable operational systems, exactly why McKinsey stresses the need for updated processes and governance to support them.

With solid design principles in place, the final piece is understanding where this shift is headed and what tomorrow’s agent-powered enterprise looks like.

Agentic AI will reshape how work is organized across the enterprise. McKinsey’s outlook goes well beyond today’s copilots and task bots. Here’s what the next phase looks like:

1. Higher levels of autonomy: Agents will take on a larger share of operational and analytical decisions under human oversight. Not full independence, but meaningful autonomy.

2. Real-time workflow optimization: Instead of following a fixed path, agents will adjust workflows on the fly, surface bottlenecks, and suggest better routes as conditions change.

3. Agent-native products & services: Software will increasingly ship with built-in agents, and service providers will use them to deliver outcomes faster and at lower cost.

4. Redefined human roles: People move from performing each step to supervising agents, shaping goals, validating quality, and handling judgment-heavy work.

5. Multi-agent collaboration: Enterprises will rely on networks of agents that hand off tasks, escalate issues, and coordinate work, mirroring real teams but at machine speed.

6. Greater emphasis on proprietary data: Companies with strong, unique data ecosystems will gain an edge, since agents rely heavily on proprietary context to make accurate decisions.

7. Blurred functional boundaries: Agents won’t stay limited to operations. They’ll extend into product development, go-to-market, finance, compliance, and strategy.

As enterprises prepare for a multi-agent future, the real challenge isn’t imagining use cases; it’s deploying them safely, reliably, and at scale. That’s where Ema makes the shift practical.

Ema is built as a Universal AI Employee designed to run real workflows. It brings agentic intelligence to the core of enterprise operations without forcing you to rebuild your tech stack.

Here’s what makes Ema stand out:

• Pre-built AI agents and workflow engine: Ema uses a Generative Workflow Engine™ with hundreds of integrations. You can configure new agent workflows using the tools you already rely on.
• Enterprise-grade trust and governance: With private models, end-to-end encryption, data redaction, and strong audit controls, Ema meets the compliance standards large organizations expect.
• Scalable, accurate execution: Powered by the EmaFusion™ model, combining public and private AI, Ema delivers accuracy at scale so teams don’t need to trade precision for automation.

In practice, Ema plugs into your CRM, service desk, HR platform, finance tools, and internal APIs. It automates multistep work, coordinates actions across systems, and delivers results without requiring a major rebuild.

If your organization is serious about making agents work, rather than just experimenting, Ema is built for exactly that.

AI agents mark a genuine shift in how work gets done across the enterprise. As the McKinsey AI agents research shows, they’re the missing link between generative AI and real, measurable business outcomes. Organizations that start now, by choosing a high-value workflow, building the right agent setup, and scaling with solid governance, will create an advantage that others will struggle to match.

For teams ready to take that step, Ema offers a proven path forward. The Universal AI Employee ties into your existing systems, respects enterprise guardrails, and runs real workflows across support, HR, sales, operations, and industry functions.

Hire Ema to get started now!

1. What are AI agents according to McKinsey?

McKinsey describes AI agents as software components that can plan tasks, take action, and complete workflows on behalf of users or systems. Unlike chatbots, they go beyond answering prompts; agents execute multistep work across enterprise tools and processes.

2. How are AI agents different from traditional automation?

Rule-based automation breaks when inputs change. AI agents use foundation models, which let them adapt to unpredictable scenarios, use tools, and complete end-to-end processes.

3. Which business functions benefit most from AI agents today?

Customer support, IT service desks, engineering, finance, risk, and HR see the fastest gains because their workflows are structured, high-volume, and easy to measure.

4. What challenges do companies face when deploying AI agents?

Key hurdles include fragmented systems, security concerns, weak data foundations, and operating models that aren’t built for continuous learning and oversight.

5. How can organizations get started with AI agents effectively?

Begin with 2–3 high-impact use cases, define agent boundaries, build a shared governance and security layer, and pilot with real users before scaling.