The AI Team You Are Hiring Is Already Obsolete

Bob Sternfels, McKinsey’s global managing partner, says he updates one number “almost every month”. On HBR IdeaCast on 6 January 2026 it was 40,000 humans and 20,000 agents, up from 3,000 agents about a year and a half earlier. The next day, at a live podcast taping at CES, Business Insider reported the firm had 40,000 human employees and 25,000 agents as of the week before. A McKinsey spokesperson later confirmed the 25,000 to the publication.

An AI team in the agentic era is the group of people who design, supervise and answer for software agents doing part of a company’s work, together with the product and data roles that feed them. In my experience, most AI hiring plans still staff to build models, while the work that is growing is the supervising, so hire for that first. For a company of about fifty people, that puts a product owner ahead of a data scientist as the first full-time AI hire.

McKinsey isn’t a typical employer. It sells AI transformation work, and it decides for itself what counts as an agent. Read its count as a direction of travel rather than a benchmark. The direction is what a hiring plan written this year has to survive.

What changes when agents do part of the work

McKinsey’s consultants describe the pattern in “The agentic organization” (September 2025): “a human team of two to five people can already supervise an agent factory of 50 to 100 specialized agents running an end-to-end process such as onboarding a customer, launching a product, or closing the books.” That’s the firm’s account of its own client work, and no outside study has measured it yet. The shape it describes is plain enough. A few people own a process that agents mostly run.

PwC’s January 2026 analysis follows the effect through to the shape of a firm. As agents take entry-level tasks, some companies shrink the base of the pyramid and repurpose the middle tier to train, oversee and manage agents, which leaves a diamond. PwC adds that the diamond “should be a choice, not a default”, because cutting early-career hiring can starve a knowledge firm of its future leaders. It says it is building an hourglass in-house instead.

Middle management is where expectations split. In the MIT Sloan Management Review and BCG survey of 2,102 respondents (November 2025), 45% of organizations with extensive agentic AI adoption expected fewer layers of middle management, against 30% of those with no plans to adopt. Gartner predicted in October 2024 that through 2026, 20% of organizations would use AI to flatten their structure, “eliminating more than half of current middle management positions.” Both are expectations, and neither counts a job lost.

For a hiring plan the point is narrower than any forecast. The jobs that grow sit at the boundary between people and agents: deciding what an agent may do, checking what it did, keeping it running, and stopping it. A team built only to build models has nobody whose job that is.

Seven roles that still matter

None of the familiar roles disappears; a plan made only of them is what is out of date. Each one tilts toward that boundary. Titles differ between companies, so read these as patterns of work.

1. Accountable AI executive. Owns the AI portfolio, the risk appetite and the decision to stop an agent. In IBM’s 2026 CEO study of 2,000 chief executives and senior leaders, 76% of organizations had a Chief AI Officer, up from 26% in 2025. At a fifty-person firm it is a named executive who already holds the authority.
2. AI product owner. Turns a business problem into a scoped use case and decides what an agent is for. More of the job becomes defining what an agent may not do, and when it hands work back to a person.
3. Data engineer. In my view, the role that grows most. Agents act on whatever data they can reach, so pipelines, access rules and quality checks carry more weight than they did for a model that only answered questions. Data governance comes before the feature.
4. Data scientist. Less routine analysis, which agents absorb, and more work designing the tests that show whether an agent’s output can be trusted.
5. ML engineer. From shipping models to running agent systems. Writing in CIO in February 2026, Encora’s chief technology officer Lalit Wadhwa described the shift as “from engineers doing every task themselves to engineers orchestrating, reviewing, and governing AI agents.”
6. Governance specialist. Maps each system to the rules that apply, runs reviews and keeps the records an auditor will ask for. CSET researchers found more than 100,000 US job postings a year asking for AI ethics or governance skills by the early 2020s. In the IAPP’s 2025-26 salary report, half of those working across privacy and AI governance earned more than $169,700.
7. Translator. Knows enough of the technology, the business and the people to get a system adopted. It’s often a promotion from inside, because the business knowledge takes years to build. Without one, sound systems sit unused.

Four roles that are new

McKinsey’s October 2025 list of emerging profiles names “agent orchestrators, who design and supervise agent workflows; hybrid managers, who lead blended human-agent teams; and AI coaches, who help employees integrate AI into daily work.” Alongside the orchestrator I see three more specialist jobs, and companies advertise all four under many titles.

1. Agent orchestrator. Designs the workflow: which agent does which step, with which tools and data, within which limits, and where the handoffs to people fall.
2. Agent evaluator. Builds the test sets, sample checks and red lines that decide whether an agent keeps running. It’s the data scientist’s most natural next role.
3. Agent reliability engineer. Site reliability engineering applied to agents: monitoring, drift, rate and spend limits, and the switch that lets a company actually stop an agent.
4. Oversight designer. Decides where a person approves before an agent acts, where a person reviews afterwards, and where the agent acts alone, which shifts as autonomy levels rise. Singapore’s Model AI Governance Framework for Agentic AI (version 1.5, May 2026) asks organizations to define “significant checkpoints” that require human approval, “such as high-stakes or irreversible actions”, and warns that continuous human oversight over all agent workflows “becomes impractical at scale.”

For any one agent, six pieces of work need an owner.

The table is my own synthesis, in the form of a RACI chart. Its use is the check it forces: every row needs exactly one name in the column for who answers for it. At a fifty-person company three people may hold all six roles between them, and that works as long as nobody has to ask who stops the agent.

The first AI hire at a fifty-person company

In my experience the usual first hire is a data scientist. It’s often the wrong one. A data scientist who arrives before the data is clean and the use case is chosen often spends the first year cleaning data and looking for a problem. The better order runs from cheap to expensive, and each step should pay for itself before the next one starts.

The data scientist and the ML engineer come when a use case needs a model you can’t buy. Many first use cases at this size run on bought models, so that point often arrives later than the plan assumed. Until then the product owner builds on bought platforms or directs a contractor who does. A firm in finance or health should move governance up to second, because a regulator asks for the missing record before anyone misses the model.

Keeping the people you hire

Two published measures of the AI pay premium sit far apart. The 2026 edition of PwC’s Global AI Jobs Barometer (June 2026) found a 62% premium in advertised salaries for roles that require AI skills over comparable roles that do not, a figure that does not control for experience or location. A Korn Ferry survey of more than 4,000 companies (March 2026) found firms offering an average of 10% above market rate for AI roles. ManpowerGroup’s 2026 talent shortage survey of 39,063 employers in 41 countries put AI skills at the top of the list of skills hardest to find, for the first time.

Pay explains some departures. In my experience three conditions explain more: people spend their week finding and cleaning data, nothing they build reaches users, and no senior business owner understands what they do. Each has a fix inside the company. The cheapest retention move is to show each person the route from today’s role to the next one: data scientist to agent evaluator, ML engineer to agent reliability engineer, product manager to oversight designer. Draw the same map for the staff whose tasks agents take over, starting with the entry-level roles PwC warns about.

Where this could be wrong

The first objection is that McKinsey’s count doubles as marketing for the work it sells. It does, and its rivals told Business Insider that agent counts are not a metric of success. That’s why the argument here rests on the work agents create and not on how many there are.

The second is that many agent projects will fail. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Some companies will hire an orchestrator for an agent they switch off a year later. Whoever holds the evaluator and oversight duties, a hire or the product owner at fifty, makes that call early and cheaply, and a product owner earns their salary with or without agents.

The third applies to a company whose product is the model. A startup selling a model should hire the researcher first. For everyone else, the order above holds.

Sources and changes

• HBR IdeaCast, episode 1060, Adi Ignatius with Bob Sternfels, 6 January 2026.
• Business Insider, Kelsey Vlamis, "McKinsey's CEO breaks down how AI is reshaping its workforce", 7 January 2026; Lakshmi Varanasi, "McKinsey says it has 25,000 AI agents. Its rivals say that's not a metric of success.", 13 February 2026.
• McKinsey, "The agentic organization: Contours of the next paradigm for the AI era", 26 September 2025, and "Six shifts to build the agentic organization of the future", 27 October 2025 (archived copies).
• PwC, "No more pyramids: Rethinking your workforce for the agentic AI era", 29 January 2026, and 2026 Global AI Jobs Barometer, 15 June 2026 (archived copies).
• MIT Sloan Management Review and BCG, "The Emerging Agentic Enterprise", 18 November 2025.
• Gartner, top predictions for IT organizations and users in 2025 and beyond, 22 October 2024, and "Over 40% of agentic AI projects will be canceled by end of 2027", 25 June 2025 (archived copies).
• IBM, "CEOs are reshaping C-suite roles for the AI era", 4 May 2026.
• Lalit Wadhwa, "How agentic AI will reshape engineering workflows in 2026", CIO, 20 February 2026.
• CSET, "AI Ethics and Governance in the Job Market", 20 May 2025; IAPP, Salary and Jobs Report 2025-26, 3 August 2025.
• IMDA Singapore, Model AI Governance Framework for Agentic AI, version 1.5, 20 May 2026.
• Korn Ferry, "Only modest pay bumps for AI skills", 16 March 2026; ManpowerGroup, 2026 global talent shortage release, 26 February 2026.

Changes on 6 October 2026. Cut from about 3,990 words to a Reference and retitled. Two McKinsey agent counts from different days are now given separately. Salary bands with no source, a 2026 to 2031 roadmap whose dates no source gave, a hypothetical startup and a cluster of vendor forecasts are removed. Three figures now follow their sources more closely: the MIT SMR and BCG middle-management figure, the CSET job-postings count and the PwC wage premium. The download is a new worksheet.