Cede-Agency: The Verification Skills SWP and TA Teams Need

Consider what happened to the business email in the space of a single year. At the start of 2025, many of us concluded that AI could handle it entirely. The formatting was clean, the structure was sound, and the speed was extraordinary. We were enamored. By December of the same year, most of us could tell within ten seconds of opening a message that a machine had written it — and, more damagingly, that no original thinking had gone into it. From enamorment we went to dismay.

That arc is worth sitting with, because it is not really a story about email. Enterprises are targeting super-intelligent machines for 2026 and beyond. Between narrow AI and anything approaching general intelligence, a great many initiatives are already underway, some of them at genuine scale, and much of what they will produce remains unknown. What is becoming clear is that the human interaction with these systems — and the skills required to operate them — will matter more than the systems themselves.

The email episode gave us an early, low-stakes demonstration of the mechanism. The machine produced something fluent. The human accepted it without adding judgment. The recipient detected the absence of judgment immediately. Nothing failed technically. What failed was the human step that was supposed to sit between generation and transmission, and everybody could see it.

The humble pen and the durability of writing

In Genesis: Artificial Intelligence, Hope, and the Human Spirit (Kissinger, Schmidt, and Huttenlocher, 2024), there is a paragraph I have found myself returning to:

Humans have long desired to extend the range of what we can observe to both the very tiny and the very far. The microscope and the telescope are quintessential tools of human observation. Less appreciated is the humble pen. Writing, invented four thousand years ago, remains an outstanding tool for the codification and transmission of complexity. That includes mathematics, perhaps the purest and most universal of human languages, and enough in itself to facilitate the transfer of abstruse ideas and collaboration on technological projects. On a per-byte basis, language in all its multifarious and beautiful forms is unusually dense — among the most efficient data structures that have been invented.

Writing was among the first capabilities we assumed AI would fully absorb. It is turning out to be one of the indisruptible ones. Not because machines cannot generate prose — they obviously can, at volume — but because the value of writing was never the prose. It was the compression. Language is how a human being takes a complex, partly-formed judgment and encodes it densely enough that another human can decompress it and act on it. A machine can produce the surface of that transaction without the substance underneath, which is precisely why readers detected the difference so quickly.

The same logic applies well beyond writing. Any capability whose real function is to encode judgment, rather than to produce an artifact, is far more durable than its output format suggests.

Butler's warning about ceded agency

Samuel Butler is best known for Erewhon, which contains one of the earliest serious discussions of the evolution of machines. In the section titled "The Book of the Machines," Butler argued that machines evolve much as biological organisms do: becoming more complex, more autonomous, and increasingly indispensable to the humans around them. His concern was not that machines would become intelligent overnight. It was that humans might gradually cede agency, judgment, and purpose to them.

Gradually is the operative word. Nobody in the email example decided to stop thinking. The tool was good enough that thinking felt optional in that instance, and then in the next one, and the habit set before anyone examined it. That is what ceding agency looks like in practice — not a decision, but an accumulation of small non-decisions, each individually defensible.

Cede-agency is a term those of us responsible for skills development and workforce preparation in the enterprise now have to take seriously. It names a risk that does not appear in any capability assessment, because the systems are working exactly as designed while it happens.

Verification is the root skill

This is why the skills that matter most for Strategic Workforce Planning and Talent Acquisition professionals in this cycle are verification skills. They are not tech-stack skills. We have written a good deal about those elsewhere, and they remain necessary — knowing what a model can do, what a platform is architected for, how data flows between systems. Verification skills sit underneath all of that. They are the root skills for working effectively with machines.

Verification is the discipline of holding a machine's output at arm's length long enough to establish whether it should be acted on. It has a few distinct components, and they are learnable.

• Knowing what the answer should roughly look like before you ask, so that a plausible but wrong output has something to fail against rather than arriving into a vacuum.
• Being able to trace an output back to the data and assumptions that produced it, and recognizing when that trace cannot be completed — which is itself a finding, not a technicality.
• Distinguishing fluency from correctness, particularly in domains where a confident, well-formatted answer is the most persuasive thing a system can produce.
• Being willing and able to overrule the machine, and to write down the reasoning for the override in a form a colleague can evaluate.

The last of these is the one enterprises most often neglect, and it is the one that decides whether the other three are worth anything. A team that can detect a bad output but has no standing to reject it has an audit function, not a judgment function.

Why these skills are hardest exactly where they are most needed

SWP and TA work is unusually exposed to this problem. Both disciplines involve making consequential recommendations about people and capability under time pressure, using evidence that is incomplete by nature. That combination is exactly where a fluent machine output is most tempting and least easy to check. A generated shortlist, a generated skills adjacency, a generated demand forecast — each arrives looking finished, and each is being handed to a professional who has more requests than hours.

Verification is also invisible when it works. Nobody notices the recommendation that was quietly corrected before it went out. That makes it hard to fund and easy to skip, which is another reason it needs to be named as a skill and treated as one, rather than assumed to be an attribute of experienced people.

Where to start this year

My suggestion for the year ahead is narrower than a training program. Take the decisions your SWP and TA teams already make with machine assistance, and for each one write down two things: what a good answer would look like independent of the tool, and who is accountable for overruling the tool when it is wrong. Then check whether that second person has ever actually done it. If the override has never been exercised in a workflow that has been running for months, you do not have a well-calibrated system. You have a team that has quietly stopped verifying.

The machines will keep improving, and the pressure to accept their output will keep rising with their fluency. The capability worth building deliberately in 2026 is not the ability to use these systems. Most people acquire that quickly. It is the ability to hold judgment in the loop while using them — and, unlike fluency, that one does not develop on its own.