McKinsey: 62% of Companies Are Experimenting with AI Agents. Are You in the 38%?

McKinsey published their Q3 2025 Global AI Survey last week, and one data point deserves more attention than the rest: 62% of companies surveyed are now experimenting with AI agents — autonomous or semi-autonomous AI systems that perform tasks without continuous human direction.

Eighteen months ago, that number was 11%.

This is not a gradual adoption curve. This is a step function, and the businesses sitting in the remaining 38% need to understand what is happening on the other side.

What the Data Actually Shows

McKinsey surveyed 1,800 companies across 14 industries and 22 countries. The headline number — 62% experimenting — breaks down into more revealing segments:

18% have agents in production — running live, handling real business operations, generating measurable ROI. These are not science projects. These are agents managing customer interactions, processing documents, handling bookkeeping tasks, qualifying leads, and generating content.

44% are in active experimentation — they have allocated budget, assigned teams, and are running pilots. Most expect production deployment within six to twelve months.

24% are in evaluation mode — they are researching, attending conferences, and talking to vendors, but have not committed resources.

14% have no AI agent initiative — no experimentation, no evaluation, no plans.

The 14% group is the most at risk. Not because AI agents are universally necessary today, but because the learning curve for effective agent deployment is steep. Businesses that start the learning process in 2027 will be competing against businesses that have been refining their agent operations for two to three years.

Why Agent Adoption Accelerated So Sharply

Three factors drove the 5.6x increase in eighteen months:

Model Quality Crossed the Reliability Threshold

In early 2024, AI models were impressive but unreliable enough that agents required constant human supervision. By mid-2025, models like Claude Sonnet 4.5, GPT-5, and Gemini 2 Pro produce outputs reliable enough that agents can operate with periodic rather than continuous oversight. That shift from "supervise every action" to "review a daily summary" changed the labor economics from "more work" to "less work."

Cost Dropped Below the Hire Threshold

When running an AI agent cost $2,000-$5,000 per month (frontier model pricing at production volume), only high-margin businesses could justify it. With speed-tier models like Sonnet and Haiku, the same agent costs $200-$800 per month. That is less than any human alternative, including outsourcing to low-cost markets.

The Tool Ecosystem Matured

Building an agent in 2024 required significant custom engineering — API integrations, prompt chains, error handling, monitoring infrastructure. In 2025, platforms and frameworks have abstracted much of that complexity. What took a development team three months now takes two weeks.

Where Agents Are Being Deployed First

McKinsey's data shows clear clustering in early agent deployment by function:

Customer service (47% of adopters). Chatbots that can actually resolve issues, not just deflect them. The new generation of agents can access customer records, process returns, update accounts, and schedule appointments — the tasks that previously required a human with system access.

Content operations (38% of adopters). Generating, scheduling, and optimizing marketing content across channels. This includes blog content, social media, email sequences, and AI auto-blogging systems that maintain a consistent publishing cadence.

Sales and lead management (34% of adopters). Lead scoring, qualification, follow-up sequences, and meeting scheduling. Agents that monitor inbound leads and respond within minutes, 24 hours a day.

Internal operations (28% of adopters). Document processing, data entry, report generation, compliance checking. The back-office tasks that consume employee time without generating direct revenue.

The Compounding Advantage Problem

The competitive dynamic that McKinsey's data implies — but does not explicitly state — is compounding advantage. A business that deployed agents twelve months ago has:

• Refined their prompts through hundreds of iterations
• Built proprietary training data from their specific operations
• Developed institutional knowledge about what works and what fails
• Optimized their human-AI workflow for their specific business context
• Reduced costs each quarter as they learn to use agents more efficiently

A business starting from zero today cannot buy that knowledge. It has to be earned through deployment experience. And the business that started twelve months ago is not standing still — they are deploying the next generation of agents while their competitors are still figuring out the first generation.

This is the classic compounding advantage dynamic, and it is why the McKinsey data should concern the 38% more than it reassures the 62%.

What the 38% Should Do Now

If your business has not started experimenting with AI agents, the priority is not to deploy everything at once. It is to start the learning process with a single, well-chosen use case.

The ideal first agent deployment has these characteristics:

• High volume (the agent handles enough tasks to justify the investment)
• Low risk (errors are correctible, not catastrophic)
• Measurable (you can quantify the before and after)
• Currently expensive (the human cost of the task is high enough that improvement is obvious)

For most businesses, that first deployment is one of: review response management, lead follow-up sequences, or content scheduling. Each of these can be operational within two to four weeks and produces measurable results within the first month.

What This Means for Your Business

The McKinsey data confirms what we have been observing in our client work: AI agent adoption has crossed from early-adopter territory into early-majority territory. The question is no longer whether agents will be a standard part of business operations — it is how quickly the laggards can close the gap.

If you are in the 62%, the priority is moving from experimentation to production and expanding to additional use cases. If you are in the 38%, the priority is starting — even with a single agent — before the compounding advantage gap becomes insurmountable.

We help businesses at both stages. For those starting out, our AI adoption strategy process identifies the highest-ROI first deployment. For those scaling, our agent infrastructure practice handles multi-agent architectures and cross-functional deployment.

The 62% is going to be 80% by this time next year. The question is which side of that number your business will be on.