More than 150 million Americans are on a payroll, according to the Bureau of Labor Statistics. If you’re one of them, these are the AI skills to learn for 2027: the ones that make you faster at the job you already have, and then make you the person who builds AI for everyone else at your company.

There are eight, in the order I’d learn them. The first five make you excellent at delegating your own work to an agent. The last three are the skills of a Business Automation Manager, the person who builds AI systems for the rest of the team. For each one you’ll get what it is, why it pays off, which tools to use, and one rule to remember.

Part 1: Delegating your own work
1. Briefing
2. Context engineering
3. AI integrations
4. Output templating
5. Agent skill development

Part 2: Building AI for your company
6. Agentic coding
7. Plugin development
8. AI workflow automation

What changed in 2026, and why these are the AI skills to learn

In 2025, working with an AI agent meant building one yourself in a tool like n8n, Lindy or Relevance. That started to change towards the end of 2025 with Claude Code, and then changed for everyone with what we call the Cowork category: Claude Cowork and ChatGPT Work for business teams, and Claude Code, Cursor and Codex for engineers. Now almost anyone at work can have an agent as a co-worker and delegate real work to it.

That one change drove three shifts in how people use AI at work this year:

• Prompting became context engineering. The agent does the work, so the result depends on what it knows before you type anything.

• The chat box became a workspace. AI moved into your folders, your tools and a browser of its own.

• One-off answers became repeatable assets. Skills let the agent know what it can do without reading how until it needs to.

In short, 2026 moved the work from the conversation to the setup. For 2027, I expect three more. Personal agents arrived in September, when OpenAI launched Dots, an always-on agent that works towards your goals in the background, and they move you from running AI to supervising it. Agents are also moving to the cloud, so a task keeps running when your laptop is shut. And the work is going organisational: Today’s agents are single-player, so the outputs one person gets are hard for a colleague to reuse. Every skill below comes out of one of these shifts.

Part 1: Five AI skills for delegating your own work

Delegating work to an agent is quickly becoming an expectation of the job, which is why we call AI the new Excel. Automating your own work is now part of your role, because the tools let you do it yourself.

1. Briefing

Brief an agent the way you’d brief a new starter. You wouldn’t tell a junior on day one to “prepare the industry report”. You’d say: Here’s the research we did, here’s what we’re trying to achieve, prepare everything for that.

The common mistake is using Claude Cowork or ChatGPT Work like a chatbot. Agents take a while to run, so every back-and-forth costs you minutes. Start with “let’s work on the monthly report” and it’ll come back asking for last month’s examples, then for where the data lives, and you’ll spend half an hour feeding it. Do the work up front instead, send the brief, and do something else while it spends 20 minutes on the report. Workday’s research found that only 14% of employees consistently get clear, positive outcomes from AI, and a thin brief is where a lot of that gap starts.

OpenAI’s own prompting guide splits a good brief into four parts: Goal, Context, Output and Boundaries. I put the files in one project folder and dictate the rest.

Tools to use: Claude Cowork, ChatGPT Work. The rule: Don’t just say it the way you’d say it to a colleague. Prepare it the way you would for one.

2. Context engineering

Briefing is gathering what the agent needs for one task. Context engineering is making sure everything it needs, and only what it needs, is already there next time. In practice that means a sensible folder structure plus a plain-text instructions file the agent reads at the start of every session, called CLAUDE.md in Claude and AGENTS.md in most other agents.

Here’s one we use every week. Every 9x Live workshop gets its own folder holding the event details, the description and every asset it needs. When I ask “Can you create the invite email for next week’s live workshop?”, the agent finds next week’s folder, the speaker’s photo and the description on its own. The folder structure does the clever part.

The first run takes longer than doing it yourself, because the context is in your head. Set it up once and every run after that’s a one-line request.

Tools to use: Claude Cowork, Notion, Google Drive, GitHub, Obsidian. The rule: Stop explaining your business to AI. Build the library so it can look things up itself.

3. AI integrations

For most popular tools, connecting AI takes one click: Open the connectors section in Claude or ChatGPT, find the tool and connect it. The skill is what you do when there’s no connector, which happens a lot with niche tools and more traditional companies, or when the connector doesn’t offer the action you need.

There are three ways round it:

1. Build a custom connector with a tool like Make or Airtop. It uses MCP (Model Context Protocol), the open standard agents use to talk to other software.

2. Have the agent write a short script against the tool’s API, which most tools have.

3. Skip the direct connection and let the agent work through the browser.

Tools to use: Claude and ChatGPT connectors, Make MCP, Zapier MCP. The rule: There’s almost always a workaround to get the result you need.

4. Output templating

Sometimes the workaround beats the connector. We make our decks in Pitch, and its connector turned out to be one AI prompting another: Claude sent a prompt to Pitch’s built-in agent, so we paid for AI twice. Then we noticed that slides copied out of Pitch and pasted into a text file are just code, and models are good at writing code. So we taught Claude and ChatGPT to write it. We paste the result back into Pitch and get a fully editable deck, with our fonts, colours and layouts intact.

That’s the whole skill: Work out what format the finished output needs to be in, for the tool it lands in and for the result you’re after. If you want a PDF, have the agent build a well-designed A4 HTML page first and convert that. It’s worth the effort, because Workday found that nearly 40% of the time AI saves is lost to rework: correcting, rewriting and checking its output.

Tools to use: Claude and ChatGPT skills with your brand templates, Figma, PowerPoint. The rule: Don’t ask AI for words. Ask it for the finished file.

5. Agent skill development

A long-running conversation gets worse every time you scroll back and ask it to “do that again”. Turn the process into a skill instead: a saved set of instructions the agent loads only when the task comes up.

Here’s how I build them. Do the task with the agent once, start to finish. Review it until the output meets your standard. Only then ask the built-in skill creator to save it, and when it goes wrong later, correct the agent first and then have the skill creator update the skill. I haven’t written a single word of the skills I use. The mistake I see most is opening the skill creator to describe a skill you’ve never actually run.

Skills compound. Our 9x Live workshop assets take about 30 seconds with a skill, against 10 to 15 minutes by hand in Figma. Our newsletter skill and our workshop skill both call the same marketing email skill, so a template change is one update, not 30.

Tools to use: Claude skills, ChatGPT Work skills. The rule: If you’ve explained it twice, make it a skill.

Part 2: Three AI skills for building AI for your whole company

Everything so far makes you excellent at your own job. The next three make you the person who builds AI systems for everyone else, which is where the 2027 shifts point.

That’s the Business Automation Manager role: somebody with the domain expertise of the business side and just enough technology to act on it. One company asked us to train 30 people, and nine weeks later five of them had been promoted into that new job title.

6. Agentic coding

If your job has nothing to do with engineering, this one can look out of place, but it’s the skill that makes the next two possible. Agentic coding means describing what you want to a coding agent and letting it build it. We used to edit our website by hand in Framer. Now it’s all done through Claude Code, and a new case study page takes 10 minutes instead of half a day.

You don’t need to be an engineer, but you do need a few of the concepts engineers work with: a repository on GitHub so your code is versioned, hosting so it has somewhere to run, and APIs so it can talk to the rest of your stack. Claude can walk you through all three.

Tools to use: Claude Code, Codex, Cursor. The rule: You don’t need to be able to code. You need to describe exactly what you want.

7. Plugin development

A plugin is how you take your skills and share them with a whole organisation. It bundles the connections to the tools your colleagues need with the skills they should use, so it’s an entire AI setup in one install. You can roll it out to everyone from their first day and control who gets what.

At 9x we run on about a dozen, covering things like 9x Live, operations, Pitch, Attio and Notion, all shipped from one GitHub repository. A new team member installs them and has the same skills and connectors as everyone else on day one. Improving the skills inside needs a feedback loop too: We sync ours into Notion so colleagues who’d rather not open GitHub can comment on them there.

Tools to use: Claude plugins, ChatGPT Work plugins, GitHub. The rule: Ship it once, and everyone gets your best work.

8. AI workflow automation

This is where AI moves into the background. A workflow runs on a schedule or a trigger, and a person reviews the result, or in some cases nobody needs to. What changes is who you’re building for: the rest of the company, not just yourself.

Automating a process used to need the business team, a product manager and the tech team in a room. Now the subject-matter expert can build it with a coding agent. Some people call Make and n8n dead because Claude Code can write the code, but that code still needs hosting. The visual platforms handle hosting, logging and safe credential storage, and Claude Code builds on them through their MCP connectors. None of our n8n workflows or Make scenarios are built by hand any more, including the one that shares our live workshop recordings.

The payoff shows up in McKinsey’s 2026 State of AI survey: Nearly three-quarters of AI high performers have fundamentally redesigned workflows, against a quarter of everyone else.

Tools to use: n8n, Make, Zapier, Claude Code. The rule: Be the person who automates the work, not the person whose work gets automated.

Which AI skills are in demand, and who needs which

The AI skills gap shows up in pay. PwC’s 2026 Global AI Jobs Barometer puts the average wage premium for workers with AI skills at 62%, up from 57% a year earlier, and Workday found that 79% of employees getting positive outcomes from AI had received more skills training.

Not everyone at a company needs all eight. We train teams in three tiers of AI skills:

• AI literacy, for everyone: a working understanding of what AI can and can’t do. This list assumes it.

• Cowork skills 1 to 5, for about half the workforce: delegating day-to-day work to an agent.

• Automation skills 6 to 8, for about 5%: the builders.

HBX Group took around 3,000 people through the literacy tier with us in six weeks, and 95% finished on time. That’s the order that works: a shared baseline first, then the Cowork skills, then a small group who build for everyone.

Where to start if you’ve done none of this

Download ChatGPT Work or Claude Cowork and delegate one real task to it this week. Pick something you do every month, like a report, write the brief properly, and let it run while you do something else. That’s skill one, and every other skill on this list builds on it.

If you’d like to learn these skills with other people doing the same, join a free 9x Live session. We run live workshops every week, and you can bring the task you’re stuck on.

Common questions

What AI skills are in demand in 2027? The skills employers pay for are the ones that turn AI into finished work: briefing an agent, giving it the right context, connecting it to your tools and turning processes into skills. Above that, the people building AI for others, through coding agents, plugins and automated workflows, are the scarcest.

Do I need to learn to code? Not for skills 1 to 5. For skills 6 to 8 you’ll work with a coding agent, so you need the concepts (repositories, hosting, APIs) rather than the ability to write code yourself.

Which tool should I learn first, Claude Cowork or ChatGPT Work? Whichever your company already pays for. Both handle every skill in part 1, and skills you build in one work in the other.

How long does it take to learn these AI skills? You can delegate your first task in an afternoon. A company-wide literacy baseline took HBX Group six weeks. The automation skills take longer: In the programme where five of 30 people were promoted, that happened nine weeks in.