The Skills a Team Needs to Manage AI Agents (and the Training That Builds Them)

McKinsey now works with 25,000 AI agents alongside 40,000 human employees. Its CEO expects about as many agents as people by the end of 2026. Roughly one agent per person.

That's not a pilot. It's a new operating model, and it's coming to your organization whether your team is ready or not.

Agentic AI means AI systems that act on your behalf: they make decisions, run multi-step workflows, and finish tasks with little human oversight. The technology is not the hard part. Preparing your people to work with it is. Most agentic AI training still teaches people to write better prompts, and that covers about a quarter of the job.

How Fast Is Agentic AI Arriving?

The software is arriving fast. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025.

Most organizations have started too. In McKinsey's 2025 State of AI survey, 62% of respondents said their organizations were at least experimenting with AI agents: 23% were scaling an agentic AI system somewhere in the business, and another 39% had begun experimenting.

Experimenting is not the same as being ready. Deloitte's Tech Trends 2026 report puts the gap plainly: while 38% of organizations are piloting agentic solutions, "only 14% have solutions that are ready to be deployed and a mere 11% are actively using these systems in production." Another "42% of organizations report they are still developing their agentic strategy road map, with 35% having no formal strategy at all."

Capgemini surveyed 1,500 executives at organizations with more than $1 billion in revenue and found just 2% had deployed AI agents at scale, with another 12% partly deployed. Trust went backwards: "Only 27% of organizations express trust in fully autonomous AI agents, from 43% 12 months ago."

Trust fell as experience grew. That is what a skills gap looks like from the inside.

That last row matters most. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. The pattern is familiar: organizations buy the technology but don't train the people who have to use it.

How Do AI Agents Change Everyday Work?

Writing in Fortune, a Salesforce developer advocate called this the rise of the "supervisor class": professionals whose main value is no longer doing the work by hand but directing autonomous systems. His examples:

• Salesforce says its support agents now handle 96% of cases autonomously, and agents have saved over 50,000 seller hours.
• Lennar runs 1.1 million agentic workflows a month across customer engagement.
• reMarkable launched its first AI agent in three weeks, and it has since resolved more than 10,500 customer inquiries with an NPS matching its human support team.

It isn't limited to tech companies. Employers surveyed by the World Economic Forum expect 39% of workers' core skills to change by 2030. And IDC reports that 66% of enterprises are reducing entry-level hiring as they deploy AI.

The pay gap is already visible. PwC's 2025 Global AI Jobs Barometer found that workers with AI skills command wage premiums of up to 56% over their peers.

What Training Does a Team Need to Use AI Agents Effectively?

A chat assistant proposes. An agent acts. That one change moves the person from author to approver, and approving is a different skill from writing.

Most organizations aren't ready for that shift. In Capgemini's survey, "only half claim sufficient knowledge of AI agent capabilities, and even fewer can pinpoint where agents outperform traditional AI or automation." Only a third clearly understand when to choose an agent over generative AI, machine learning, or plain process automation.

Delegation

Delegation is not prompting. A prompt shapes one response you're about to read. Delegation hands over a multi-step task you won't watch, so it has to carry the success criteria, the boundaries, and the point at which to stop.

A good training exercise: write a brief, hand it to a colleague with no context, and see what comes back. Whatever they got wrong, an agent will get wrong too.

Verification

Verification means checking work you didn't do. Agents produce output fast, so the question is rarely whether they can do the work. It's whether the work is right. That means understanding what the agent did, why, and where it might have gone wrong, and checking its analysis for accuracy, bias, and gaps.

It is also the skill under the most pressure. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of generative AI use. They found that "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking". In their words, generative AI "shifts the nature of critical thinking toward information verification, response integration, and task stewardship." The data is self-reported, so treat it as a signal rather than a measurement.

Exception Handling

Exception handling is judgment about scope. Agents follow patterns well but struggle at the edges. Your team needs to notice when an agent is working outside its competence, and have the confidence to override it. The automation bias research below shows how easily people defer to a confident machine.

The trainable part is spotting the tells: the agent restates the task in narrower terms, loops on a step, or produces fluent output that no longer answers the question. People notice these sooner with practice.

Permission Awareness

Permission awareness is usually treated as an IT concern. It shouldn't be. IT grants the access, but the person using the agent is the one who sees it reach for a system the task didn't need. Knowing what an agent can touch tells you what it could break if it misreads a request.

What Skills Does the Team Hold Together?

Process Design

Many organizations bolt agents onto processes designed for people. Deloitte found that leading organizations are redesigning workflows end to end around what agents can do, rather than automating existing steps. Your team needs to decide which tasks to delegate, what guardrails to set, and where work passes between people and agents.

Domain Expertise

This is the part agents can't replicate. A compliance analyst who understands regulatory nuance. A sales lead who reads client relationships. A nurse who spots what a triage agent missed. Domain expertise becomes more valuable with agents, not less, because it turns raw agent output into good decisions.

Governance

When agents act on their own, someone has to set the limits. Which decisions can agents make alone? Which need human approval? What data can they access? Close to three-quarters of companies plan to deploy agentic AI within two years, but Deloitte's State of AI in the Enterprise 2026 found only 21% of them have a mature model for agent governance. Your team needs the judgment to know where autonomy should stop.

Once you run many agents instead of one, ownership, permissions, and oversight become problems of their own. Managing AI agents at scale covers that layer.

What Do the EU AI Act and NIST Require?

EU AI Act Article 4 requires providers and deployers of any AI system to take measures supporting AI literacy among their staff and others who operate AI on their behalf. It has applied since February 2, 2025, whatever the risk class. Our Article 4 guide covers the scope and the July 2026 amendment in detail.

Article 26(2) goes further for high-risk systems, and it is one sentence: deployers "shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support." The July 2026 Digital Omnibus pushed the high-risk rules back to December 2027 at the earliest, which gives teams time to train, not a reason to wait.

Article 14(4) then lists what that person must be able to do, and it reads like a training syllabus:

• (a) understand the system's capacities and limitations, and monitor it for "anomalies, dysfunctions and unexpected performance"
• (b) "remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias)"
• (c) correctly interpret the system's output
• (d) decide not to use the system, or "disregard, override or reverse" its output
• (e) intervene in the system or halt it safely

The EU wrote automation bias into binding law. That shows how central it is.

NIST's AI Risk Management Framework reaches the same place on a voluntary basis. Its GOVERN 2.2 subcategory is a training requirement in one line: "The organization's personnel and partners receive AI risk management training to enable them to perform their duties and responsibilities consistent with related policies, procedures, and agreements." The companion Playbook says the training should differ by role, with trainings suitable "for AI actors carrying out technical tasks (e.g., developers, operators, etc.) as compared to AI actors in oversight roles (e.g., legal, compliance, audit, etc.)."

ETSI EN 304 223, the European Standard on baseline cyber security requirements for AI models and systems, uses the word shall: "Organizations' cyber security training programme shall include AI security content which shall be regularly reviewed and updated" and "AI security training shall be tailored to the specific roles and responsibilities of staff members."

How Do You Train Against Automation Bias?

A systematic review in JAMIA screened 13,821 papers down to 74 studies and measured the effect: "The risk ratio was 1.26 (95% CI 1.11 to 1.44); erroneous advice was more likely to be followed in the CDSS groups than in the control groups and when in error the CDSS increased the risk of an incorrect decision being made by 26%." One study the review cites, by Friedman and colleagues, is more striking still: "in 6% of cases, clinicians over-rode their own correct decisions in favor of erroneous advice from the DSS."

Trained professionals, in their own field, talked out of correct answers by a machine.

A NASA-funded experiment measured the cost to attention. Across 8 simulated flights with 80 student participants, people with an automated aid missed far more unprompted events than people without one: "participants in the automated condition missed more of these six events (M = 2.44) than those in the non-automated condition (M = 0.18)". The twist is that the automated group scored better overall (88.67 versus 83.68 out of 100). Automation helps and blinds at the same time, so the training has to target the blind spot, not the tool.

NIST's Generative AI Profile lists it among twelve risks, under "Human-AI Configuration": "due to the complexity and increasing reliability of GAI technology, over time, humans may over-rely on GAI systems... Automation bias can exacerbate other risks of GAI, such as risks of confabulation or risks of bias or homogenization."

What helps is a sense of ownership. In a study of pilots in high-tech cockpits, accountability imposed by the researchers made no significant difference. But in a follow-up analysis, pilots "who reported an internalized perception of 'accountability' for their performance and strategies of interaction with the automation were significantly more likely to double-check automated functioning against other cues and less likely to commit errors than those who did not share this perception."

That translates into three training designs:

1. Practice on cases where the agent is confidently wrong. A curriculum built only on good outputs trains approval, not verification. Seed wrong answers on purpose.
2. Make a person own the outcome, explicitly and by name. "The agent drafted it" should never be an acceptable answer.
3. Check against an independent source, not the agent's own confidence. Fluency is not evidence, and models are fluent by design.

What Security Training Do Agent Users Need?

LLM01: Prompt injection. The OWASP definition includes the part users most need to understand: "These inputs can affect the model even if they are imperceptible to humans, therefore prompt injections do not need to be human-visible/readable, as long as the content is parsed by the model." An agent that reads a document, a webpage, or an inbox can take instructions from that content. OWASP is also clear that the problem is unsolved: "research shows that [RAG and fine-tuning] do not fully mitigate prompt injection vulnerabilities."

LLM06: Excessive agency. OWASP defines it as "the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM, regardless of what is causing the LLM to malfunction." Its three root causes give a whole team a shared vocabulary: excessive functionality, excessive permissions, excessive autonomy.

Users don't build the controls, but they are best placed to notice the symptoms. Someone who knows the phrase "excessive permissions" is the one who asks why the scheduling agent can read the finance drive.

Why Aren't Most Organizations Ready?

The cost is real. EY found that 40% of potential productivity gains are being missed because companies lack a talent training strategy.

1. Training lags behind deployment. 77% of employers plan to reskill and upskill their workforce to work better alongside AI. Yet 44% of companies don't use proven learning methods such as mentorship, coaching, or peer support, and have no plans to start.

For many people, training barely exists. Slack's Workforce Index, a survey of 17,372 desk workers in 15 countries, found 30% had received no AI training at all, not even self-directed learning, and 61% had spent under five hours learning to use AI.

In a later Slack survey, 29% of desk workers said their company hasn't issued any formal guidance, and half said AI use isn't explicitly encouraged. The same survey found the lever: "workers at companies that actively promote AI use are nearly three times more likely to become power-users."

2. Most training targets the wrong skills. Traditional AI training teaches people about AI. Agents need people who can work with AI. That's the difference between AI literacy and AI fluency, and most programs stop at literacy.

KPMG's Q2 2026 AI Pulse, a survey of 204 US leaders at organizations with $1 billion or more in revenue, found "65% say teaching prompt and instruction skills remain the most important way to train employees to work productively with AI agents." Set that against the four skills above: two-thirds of leaders are aiming their programs at the first one.

3. Completion rates hide the problem. HRMorning reports that employees are already using AI agents to breeze through compliance training and parts of upskilling programs. The completion rate looks fine. Whether anyone learned anything is another question.

How Do You Build an Agent-Ready Team?

The organizations getting this right build AI fluency first and add agent skills on top. They don't do it in a one-off workshop: the subject changes faster than a course can be revised, and the skills are habits, not facts.

Start With Daily AI Fluency

AI fluency, the ability to evaluate, apply, and reason about AI in daily work, comes before everything else. People who can't yet judge a model's output won't suddenly be able to judge an agent's multi-step plan.

Build it in short daily sessions, not intensive workshops. Microlearning courses average 80% completion rates, versus 20% for long-form modules. And organizations that pair AI investment with structured upskilling are nearly twice as likely to see strong returns.

Keep It Continuous

The target keeps moving. The World Economic Forum's Future of Jobs Report 2025, drawing on more than 1,000 employers with over 14 million workers across 55 economies, puts the scale this way: "if the world's workforce was made up of 100 people, 59 would need training by 2030." It expects 11 of those 59 not to get it. A curriculum fixed at the start of a quarter describes a different generation of tools by the end of it.

Make It Role-Specific

A finance analyst supervising a forecasting agent needs different skills than a recruiter managing a screening agent. The signs that an agent has gone off track look nothing alike for a paralegal and a support lead. Generic "AI for everyone" programs don't build that judgment. One industry analysis puts the gain from training set in the learner's own context at up to 40% for comprehension and retention.

Practice on Real Work

Nobody learns verification by reading about it. Every session should end with real work, not a quiz. That is the core design principle behind why most AI training programs fail.

Measure Competence, Not Completion

Track what people can do with AI agents: scenario-based assessments, on-the-job performance, and decision quality, not how many modules they clicked through. The same HRMorning analysis advises shifting KPIs from completion rates to demonstrated competence. To show a program worked, see how to measure AI training ROI.

Build Team-Level Capability

Individual fluency is necessary but not enough. Teams need a shared vocabulary, shared judgment, and shared rules for when to trust an agent and when to step in. When teams learn together, discussing edge cases and sharing mistakes, agent adoption sticks.

The Window Is Open, Not Forever

Gartner expects generative AI to require 80% of the engineering workforce to upskill through 2027. Engineering is not the only function affected. Agents that execute, analyze, and recommend are reshaping finance, HR, legal, operations, and customer success too.

The organizations that get value from agents won't be the ones that bought them first. They'll be the ones whose people can tell, quickly and reliably, when an agent is wrong. The rest will keep wondering why their agent projects stall.

Your people don't need to become AI engineers. They need to become good supervisors of AI, one short session at a time.

Go Deeper

• Managing AI agents at scale: what changes when you run many agents instead of one, from ownership and permissions to oversight and measurement.
• EU AI Act Article 4: the legal minimum for AI literacy in the EU.
• What is AI fluency: the foundation under every skill on this page.