Until quite recently, using AI at work meant one thing: an empty chat box. You typed a request, got something back, and the quality depended almost entirely on how well you phrased it. Two colleagues doing the same task got two different results, and every conversation started from zero. The model knew nothing about how your team works, what your templates look like or what a finished piece of work should contain, so you told it again. Every time.
That period left a mark. A large share of the workforce tried an early version of ChatGPT or Copilot in 2024 or 2025 and asked it something that mattered to their job. They got back something generic, or confidently wrong, and reasonably concluded it wasn't for them. They went back to working the way they always had and haven't looked since. People who stopped early had no reason to keep checking, so the tools got better without them.
And the tools have changed in a way that alters the basic shape of the thing. In October 2025, Anthropic introduced Skills for Claude. A skill is a package of instructions, templates, reference material and scripts that the model picks up automatically when a task calls for it. In December, Anthropic published the format as an open standard, and within months the rest of the industry had adopted it. ChatGPT added Skills for business workspaces in March 2026. Microsoft uses the same format in Copilot for Excel and in Microsoft 365 Copilot agents. Mistral's Vibe, now available to every civil servant in Luxembourg, lets administrators switch on shared skills for whole teams.
The difference sounds modest, but it isn't. A skill is a team's way of doing something, written down once and reused by everyone. Think of the monthly variance report, the contract review checklist or the house style for a briefing note. Instead of each person coaxing the model towards the right result, the person who knows the process best writes it into a skill, and the model follows it for everyone. Anthropic's own finance team runs around 150 of them. The accountants and analysts wrote them themselves rather than handing the job to engineers, and they're version-controlled like software. A budget-versus-actuals comparison that used to take a week now takes minutes, and travel compliance checks have gone from four hours to three and a half minutes. This is what AI as a utility looks like: shared infrastructure that behaves consistently whoever is using it, rather than a clever tool that only some people are good at.
This is also where the early sceptics matter most. A good skill has to capture the judgement of the people who know a team's work best. Those are often exactly the people who decided two years ago that AI couldn't do their job. Many of them don't know skills exist, and nobody has asked them to write one. That's a loss on both sides. The organisation misses their expertise, and they miss a version of the technology that bears little resemblance to the one they gave up on.
The teams getting real value treat skills as something to manage continuously. They're careful about what's worth turning into a skill, usually repeated tasks where a consistent output matters. They test each skill against doing the task without it because a skill that doesn't beat that baseline is just overhead. They give each one an owner, revisit it when the model or the process changes, and retire it when it stops earning its place. None of this is glamorous, and it takes time out of already full weeks. But it's the difference between a skill library that keeps getting more useful and one that quietly goes stale.
AI at work has moved from prompting to packaging. The gap between organisations that get value from it and those that don't is no longer mostly about the model. It's about whether teams are building, testing and looking after the know-how that sits on top of it. It's also about whether they invite back the people who gave up two years ago to see what's changed.
