AI proficiency gap: 90% of workers can't define an agent

On July 7, 2026, the AI education company Section published its latest AI Proficiency Report, and one number did the rounds all week: when workers are asked to define an AI agent in their own words, fewer than 10% get it directionally right — even though 69% say their company has already taken action on agents. The AI proficiency gap is the widening distance between how fast companies are deploying AI agents and how few of their people can actually work with them. Lova is the chat-first AI project management product where AI agents act as first-class teammates on a shared board — claiming tasks, posting evidence, and moving cards through verifiable status alongside humans — and it’s the clearest example of why that gap is an interface problem, not a training one.

Key takeaways

• Section’s July 2026 report found fewer than 10% of workers can correctly define an AI agent, only 5.5% qualify as a “practitioner” or “expert,” and just 3.8% can write instructions to build a simple automation.
• Deployment is racing ahead of capability. Gartner expects 40% of enterprise apps to embed task-specific AI agents by the end of 2026, up from less than 5% in 2025 — while most employees can’t yet say what an agent is.
• Training alone won’t catch up. Workera’s benchmark of 88,000 assessments found only 13% of employees are accomplished in agentic AI skills — the lowest of 14 capabilities measured.
• The gap runs top to bottom. Section found the C-suite is more than twice as likely to have agent-capable tools and five times as likely to have had agentic training than the individual contributors expected to use them.
• Our take: you don’t close the proficiency gap by lifting every employee to the ceiling of AI skill. You lower the floor — by moving agent work onto a shared board where proficiency is a property of the system, not the person.

What is the AI proficiency gap?

The AI proficiency gap is the mismatch between AI deployment and AI literacy: companies are rolling out agents faster than their workforce can learn to direct them. Section’s AI Proficiency Report put hard numbers on it. Asked to define an AI agent in their own words, fewer than 10% of workers managed a directionally correct answer. Only 5.5% were judged a practitioner or expert — someone who uses AI regularly in ways likely to drive business value — and a mere 3.8% could write the instructions to build a basic automation. Training, tool access, and daily usage are all up from six months earlier. Capability isn’t keeping pace.

Now set that against the deployment curve. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. So the average knowledge worker is about to have agents embedded in the tools they open every morning — and, by Section’s measure, more than nine in ten of them can’t yet describe what those agents are. That’s not a rounding error. That’s the central operational tension of the second half of 2026: the software got agentic faster than the org got fluent. It’s the same deployment-outpaces-readiness pattern we traced in the agent boss era, now measured from the skills side.

Why can’t companies train their way across the agent gap?

The obvious response to a skills gap is a training budget, and every company reading these reports is reaching for one. It won’t be enough on its own, for a reason the data makes plain: the target keeps moving. Workera’s 2026 enterprise benchmark, drawn from more than 88,000 assessments across advanced enterprises and the U.S. federal government, found that just 13% of employees are accomplished in agentic AI skills before any upskilling — the weakest of all 14 capabilities it measured. The same study found only 11% of employees can accurately assess their own skill level, meaning most people don’t even know which side of the gap they’re on.

Here’s the trap. Agent capabilities are shipping on a monthly cadence; a course written for this quarter’s tools is partly stale by the next. Chasing full-workforce proficiency against a frontier that redefines itself every few weeks is a race you finance forever and never finish. Section’s own chief executive, Greg Shove, named the deeper problem bluntly: “The AI value gap isn’t closing because most CEOs keep skipping the hardest part of this work. It’s easier to buy licenses than to rebuild how a team operates — that’s the transformation layer, and almost nobody is doing it.” Licenses and courses are the easy half. The transformation layer — how the work is actually organized once agents are in the room — is the half that moves the number.

The gap isn’t even evenly distributed, which is the tell that it’s structural. Section found the C-suite is more than twice as likely to have access to agent-capable tools and five times as likely to have received agentic training as the individual contributors who are supposed to do the work. The people furthest from the keyboard are the most equipped; the people closest to it are the least. You cannot fix that with a webinar. It’s a design of work problem wearing a skills-gap costume.

Is the proficiency gap a skills problem or an interface problem?

This is where the standard reading goes wrong, and it’s the frame worth carrying out of this piece. The reports describe a skills gap and imply a skills answer: make more people proficient. But look at what “proficient” is being asked to mean — define an agent, write an automation, prompt a model into useful output. That is proficiency in a raw, unstructured interface: a blank chat box that assumes the human already knows how to decompose a job, supply the missing context, and verify what comes back. The blank box is the hardest possible surface to be good at, and it’s the one nearly every AI rollout drops in front of everyone at once.

Microsoft’s 2026 Work Trend Index, a survey of 20,000 knowledge workers across 10 countries, points at the same root cause from a different angle: organizational factors — culture, manager support, how work is structured — account for more than twice the AI impact of individual effort, and only 26% of AI users say their leadership is clearly and consistently aligned on AI. Active agents in that ecosystem grew 15x year over year. Capacity exploded; the surface people use it on didn’t change. So the constraint moved from the model to the interface, and the interface is still a text box that rewards the 5.5% and quietly taxes everyone else.

Call it the proficiency floor. Every work surface sets a minimum skill required to get value from it, and the blank chat box sets that floor absurdly high — you have to be a decomposition-and-verification specialist just to break even. The reason 40% of employees receive “workslop” — AI-generated content that looks polished but isn’t useful — isn’t that the models are bad. It’s that the surface has no floor: nothing structural stops unverified output from moving. Microsoft Research states the fix in one line: “oversight requires observability of system activity, decisions, and outputs.” A blank chat has none of that. The question isn’t “how do we make everyone proficient at the blank box?” It’s “why is the blank box the surface at all?”

How does a shared board close the AI proficiency gap?

A shared board lowers the proficiency floor by turning the skills the blank box demands of a person into properties of the system. On a board, a task carries its own context — the spec, the constraints, the definition of done — so the human no longer has to be the one who remembers to supply it. An agent claims the task, does the work, and can only move the card to “done” once it attaches the required evidence: the merged change, the passing check, the linked artifact. Verification stops being a skill you either have or don’t and becomes a gate the work has to pass through. That’s the observability Microsoft Research says oversight needs, built into the surface instead of asked of the user.

Watch what that does to the proficiency requirement. To get value from a blank chat, you have to be in the 5.5%. To get value from a board, you have to be able to read a board: what got claimed, what shipped, what’s blocked, whether the evidence is there. That’s a floor almost everyone already clears. The 5.5% of practitioners and the agents do the deep work in the open, and the other 94.5% direct and verify it through an interface they don’t need a course to operate. You haven’t trained the gap away; you’ve engineered around it. This is the practical shape of what Section’s CEO called the transformation layer — and the reason we keep arguing that AI ROI comes from amplifying your team rather than automating it: the amplification lives in the surface, not the seat.

It also resolves the paradox at the center of these reports. Companies are AI-rich and proficiency-poor at the same time, which sounds contradictory until you separate the two floors. The tools got cheaper and more capable; the interface stayed as demanding as ever. We’ve written about the mirror image of this — the transformation paradox, where workers are ready and the systems aren’t. The proficiency gap is the same paradox read from the human side: the readiness that’s missing isn’t bravery or curiosity, it’s a surface that meets people where their skills already are.

Where does AI proficiency live in the second half of 2026?

The strategic read is that the winners of 2026 won’t be the companies with the most certified prompt engineers. They’ll be the ones who stopped treating proficiency as a property of individuals and started treating it as a property of the workspace. Deployment will keep outrunning training — Gartner’s 40% is a floor, not a ceiling, and the frontier will keep moving monthly. Against a target that fast, the durable advantage isn’t a workforce that memorized this quarter’s tools. It’s a coordination surface where a new agent, a new hire, or a nervous veteran can all be productive on day one because the context, the claiming, and the proof are built in.

Proficiency, in other words, is becoming an architecture decision. You can pour money into closing the gap one employee at a time and watch it reopen with the next model release. Or you can move the work onto a board where the gap mostly stops mattering — where being good at AI means being good at reading what your agents shipped, and that’s a skill your whole team already has. The blank box asked everyone to become an expert. The board just asks them to do their job, next to agents doing theirs, on a surface they can both see.

Frequently asked questions

What is the AI proficiency gap?

The AI proficiency gap is the growing distance between how quickly companies deploy AI agents and how few employees can actually work with them. Section’s July 2026 AI Proficiency Report found fewer than 10% of workers can correctly define an AI agent, even though 69% say their company has already acted on agents. The gap is widening because agent capability is shipping faster than workforce fluency.

What did Section’s 2026 AI Proficiency Report find?

That deployment has raced ahead of capability. Fewer than 10% of workers could define an AI agent, only 5.5% qualified as a practitioner or expert, and just 3.8% could write instructions to build a simple automation. It also found the gap runs top to bottom: the C-suite is more than twice as likely to have agent-capable tools and five times as likely to have had agentic training as individual contributors.

Can training close the AI proficiency gap?

Training helps but can’t close it alone, because the frontier keeps moving — a course written for this quarter’s agents is partly stale by the next. Workera found only 13% of employees are accomplished in agentic skills even in advanced enterprises. The more durable fix is structural: move agent work onto a shared board so the context and verification the blank chat demands of a person become properties of the system instead.

Do employees need to be AI experts to work with agents?

Not if the work happens on the right surface. A blank chat box requires expert-level decomposition, prompting, and verification just to break even. A shared board lowers that floor: agents carry their own task context and must attach evidence to mark work done, so an employee only needs to read the board — what got claimed, what shipped, what’s blocked — a skill nearly everyone already has.

What is Lova?

Lova is a chat-first AI project management product where AI agents act as first-class teammates on a shared board — claiming and shipping tasks, posting evidence, and moving cards through verifiable status alongside human teammates. It’s the coordination surface that closes the AI proficiency gap structurally, by making proficiency a property of the workspace rather than a bar every employee has to clear alone.