Kief Studio

McKinsey Runs 25,000 AI Agents. Your Team of Four Can Run the Same Play.

Automate

  • Published
  • Updated
  • 8 min read
Glowing violet waves with a pale gold crest at center, cover for a note on how a team of four can run AI agents

TL;DR: McKinsey says it runs about 25,000 AI agents alongside 40,000 people. A small team does not need that scale. It needs a few AI agents, each with one job, connected to the tools it already uses through an open standard like MCP, with the work redesigned around them and a person checking the results.

At CES in January 2026, McKinsey's global managing partner put a number on AI agents inside a consulting firm. The figure is striking, and it says a lot about how large organizations are reorganizing work. For a team of four, the useful lesson is not the count. It is how the work gets wired.

I want to separate what McKinsey actually said from what the internet added, then look at what the same approach looks like when you have one office, a handful of tools and no change management department.

Update, October 2026: we first published this note in March 2026. We have checked each figure against primary sources, corrected the date and details of the MCP donation, and removed statistics we could not trace.

What did McKinsey say about its AI agents?

Business Insider reported that Bob Sternfels said "as of last week, the company had 40,000 human employees and 25,000 personalized agents," and that "He expects McKinsey to have about the same number of AI agents as human employees by the end of the year." He also said the firm saved "1.5 million hours in search and synthesis work last year alone." (Business Insider, via Yahoo Finance (opens in a new tab))

Do the division and you get about 0.6 AI agents per person. McKinsey needs that many because it is a global firm with a very large number of distinct workflows. A four-person studio, practice or shop has far fewer, which is the good news.

Why the count is the wrong number to copy

An agent count tells you how many narrow jobs an organization has automated. It does not tell you whether any of them pay off. A small team that wires four AI agents into its real work can see more change per person than a large firm does with thousands, because each agent covers a bigger share of the week.

Is the plumbing for AI agents ready?

Mostly, yes. The Model Context Protocol, or MCP, is the open standard that lets AI agents call other software: your CRM, your invoicing system, your calendar. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, "a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI," with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. Anthropic reported "more than 10,000 active public MCP servers" and "97M+ monthly SDK downloads across Python and TypeScript." (Anthropic (opens in a new tab))

The Linux Foundation's announcement, dated December 9, 2025, describes the foundation as "a neutral, open foundation to ensure agentic AI evolves transparently and collaboratively." (Linux Foundation (opens in a new tab)) Our March note placed the donation in March 2026. That was wrong, and we have fixed it.

What that means in plain terms: connecting AI agents to common business software no longer requires a custom protocol for every tool. The connection layer is becoming shared infrastructure. The hard part has moved somewhere else.

Why do most AI projects still fail to pay off?

Because the work around the tool does not change. Fortune's coverage of MIT NANDA's 2025 report, The GenAI Divide, says "about 5% of AI pilot programs achieve rapid revenue acceleration; the vast majority stall, delivering little to no measurable impact on P&L." The core issue, it reports, is "Not the quality of the AI models, but the 'learning gap' for both tools and organizations," and "MIT's research points to flawed enterprise integration." (Fortune (opens in a new tab))

Gartner sees the same risk for AI agents specifically. It predicted that more than 40 percent of agentic AI projects "will be cancelled by the end of 2027 due to rising costs, unclear business value, or insufficient risk controls," and estimated "only about 130 of the thousands of agentic AI vendors are real." (The Register (opens in a new tab)) At the same time, Gartner expects 40% of enterprise applications to "feature task-specific artificial intelligence agents by 2026, a significant rise from less than 5%" in 2025. (IT Brief, reporting Gartner (opens in a new tab))

Put those side by side. AI agents are arriving inside the software you already pay for, whether you plan for them or not. Many of the projects built on top of them will be cancelled. The difference is usually the wiring and the redesign, not the model.

What does wiring AI agents actually mean?

Wiring means three decisions for each agent: one job, the two or three tools it can touch, and the point where a person checks the output. Our knowledge base calls the coordinated version multi-agent orchestration, but the idea starts small.

Network switch with a few patch cables plugged in, like a small team wiring a handful of AI agents to its tools
A small team needs a few well-wired AI agents, each connected only to the tools its job requires.

A hypothetical four-person firm

Picture a four-person bookkeeping practice in Grafton. Here is what four AI agents could look like there:

  1. Overdue invoice follow-up. Reads the invoicing system each morning, drafts reminders for anything past 30 days, and queues them for a person to send.
  2. Meeting notes to CRM. Turns call notes into contact updates and next steps, flagged for review before they save.
  3. Monthly client summary. Pulls the numbers into a standard report draft that an accountant edits.
  4. Document intake. Sorts incoming client documents into the right folders and lists what is missing.

Each one touches two or three systems the firm already uses. None of them sends anything to a client without a person approving it. That is the redesign: the job changes from doing the task to checking it.

Why fewer tools per agent works better

Model vendors document this. Anthropic's tool use documentation says Claude's ability to pick the right tool degrades once you exceed 30 to 50 available tools, and that tool definitions from several connected servers can consume tens of thousands of tokens before any work starts. (Anthropic documentation (opens in a new tab)) The same principle applies with any model: give each agent the few tools its job needs and nothing else.

Why do small teams have an advantage with AI agents?

Three reasons.

Fewer workflows to map. A small team has a dozen or so core processes. You can write each one down in an afternoon and decide which AI agents belong where.

Faster redesign. Changing how work flows at a large firm takes committees and pilots. At a four-person firm it takes a meeting and a test week.

Bigger share of the week. One well-wired agent that handles follow-ups and reports is a visible slice of a small team's time. At a firm of tens of thousands it barely registers.

The risks are real too. Agents with too much access can do damage quickly, which is why we wrote a guide to AI agent governance for a small business. Start with read access, add write access one action at a time, and keep a person on anything that leaves the building.

How should a small team start with AI agents?

  1. List your weekly repeat work. Anything done the same way every week, with clear inputs and outputs, is a candidate.
  2. Pick one job, not one tool. "Chase overdue invoices" is a job. "Use AI" is not.
  3. Connect only what that job needs. Two or three systems, through official integrations or MCP servers from the vendors you already use.
  4. Decide where a person checks. Write it down before you turn anything on.
  5. Measure for a month. Hours saved, errors caught, and anything the agent got wrong. Then add the second agent.

What to watch in the first month

Keep a simple log next to each agent: what it did, what a person changed before approving it, and what it got wrong. If the reviewer keeps rewriting the same part, the instructions or the data are the problem, not the person. If the reviewer approves everything without reading, the review step has quietly stopped working, and that is the moment to tighten it rather than remove it.

Also watch the edges. AI agents tend to do well on the common case and stumble on the unusual one: the client with two billing addresses, the invoice that was paid in cash, the document in the wrong format. Write those cases down as they appear. They become the rules that make the second month better than the first, and they tell you when a job is ready for more of the routine work.

For the technical side of designing MCP tools that AI agents use correctly, Qurtoo has a practical guide (opens in a new tab), and Brian has written about why every tool we ship has an MCP server (opens in a new tab).

Frequently asked questions

How many AI agents does McKinsey have?

Business Insider reported that McKinsey's global managing partner said the firm had 40,000 employees and 25,000 personalized agents in January 2026, and expected similar numbers of each by year end.

How many AI agents does a small business need?

Start with one, tied to one recurring job. Many small teams can cover their most repetitive work with three to five well-scoped agents.

What is MCP?

The Model Context Protocol is an open standard for connecting AI applications to other software. Anthropic donated it to the Linux Foundation's Agentic AI Foundation in December 2025.

Why do most AI agent projects fail?

Research and analyst forecasts point to integration and value, not model quality: unclear business value, rising costs, weak risk controls and work that was never redesigned.

Should AI agents act without human review?

Not for anything that leaves your business or changes money. Let agents draft and prepare, and have a person approve what goes out.

How long does it take to set up a first agent?

It depends on the tools and the data. Mapping one workflow and deciding its review step is a short job; testing it against real work takes a few weeks.

Where we fit

Kief Studio takes AI ideas out of the demo and into daily work. Our AI implementation work builds quality-gated automation for repetitive work and runs it for you on LTFI, so nothing is installed on your side. When a job has several steps, our multi-agent workflows split it across specialized agents, each with one task, a clear handoff and a point where it stops, connected to your tools.

We will not sell you an agent count. We will help you pick the first job, wire it to the tools you already use, and decide where a person stays in the loop. You can read more about LTFI, our agentic business operating system (opens in a new tab).

First conversation is free. No commitment.

Tell us which job you would hand to an agent first

Related articles

  1. Automate

    AI Implementation Timeline in Worcester: What We Set Up in the First Two Weeks

    Our AI implementation timeline for Worcester and Shrewsbury teams: one workflow, named approvers and a full log, day by day through the first two weeks.

    Amelia Gagne

  2. Automate

    AI Agents Now Hit 66% Task Success. The Gap in Yours Is AI Agent Integration

    AI agents hit 66% task success on OSWorld, yet most small businesses see little. What's missing: AI agent integration with least privilege and human approval.

    Amelia Gagne

  3. Automate

    Prebuilt AI Agents for Small Business: Whose Playbook Do They Run?

    Prebuilt AI agents from your CRM or office suite handle common work well. What you can change, how credits bill, and what to ask before you sign.

    Amelia Gagne