What are AI skills?
AI skills are the abilities a person needs to get useful, safe work out of AI tools in their own job. For a marketer, an account manager or a support agent, that means writing a clear request, choosing which tasks suit a tool, handling information carefully, checking the output and explaining what they changed. It does not mean training models or writing code.
A useful public reference is the AI foundation skills for work benchmark that Skills England published on GOV.UK in January 2026. It groups six foundation skills into three domains: technical, non-technical, and responsible or ethical. The six are writing clear instructions for AI tools, using AI tools to support routine processes and tasks, using simple software that automates work tasks, adjusting how an AI tool works to get better results, understanding the risks and consequences of AI use, and analysing information using AI tools, which the benchmark describes as including checking outputs for accuracy and spotting errors.
That list is a sensible floor. In practice the skills that separate careful users from careless ones sit inside it: supplying the right context, refusing to let a fluent answer stand in for evidence, and keeping accountability with a person. The rest of this guide treats AI skills as observable behaviour on a task, because that is the only version a manager can actually see, support and review.
What skills are needed to learn artificial intelligence?
It depends which kind of AI work you mean. Building AI systems needs mathematics, statistics, programming and machine learning engineering. Using AI tools well in a non-technical job needs none of those. It needs clear writing, domain knowledge, information judgement and the habit of checking work. Most search results mix the two, which is why people often assume they need a coding course before they can use ChatGPT or Copilot at work.
This guide covers the second kind: workplace AI skills for people whose job is marketing, sales, operations, support, HR, finance or management. If you want to become a machine learning engineer, a university course, a data science programme or a provider's developer documentation is a better starting point than anything here, and Tosheo.ai is not built for that goal.
Are AI skills the same as Claude Skills or agent skills?
No. In some AI products, "skills" is also a product feature name. Anthropic's help centre describes Skills in Claude as packages of instructions, scripts and resources that Claude can load to perform specialised tasks more consistently, such as following an organisation's workflow or creating branded documents. Claude checks which skills are relevant to a request and loads only what that task needs.
Those are configurations for the tool, not abilities of the person using it. They can be useful: a team might package its report format so the tool applies it each time. But someone still has to decide what goes into that package, judge whether the output is right and own the result. A packaged skill can make a weak instruction more consistent; it cannot make an unchecked claim true. When this site says AI skills, it means the human abilities.
Which AI skills matter most for employees?
Seven practical skills cover most knowledge work. They are listed roughly in the order a new user meets them, although an experienced user practises all seven on every serious task.
Choosing the task
Deciding whether a task suits an AI tool at all. Drafting a first version, restructuring notes or generating options often does. Making a decision about a person, stating a legal position or sending a figure to a client without checking it does not.
Briefing with context
Giving the tool the audience, purpose, source material, constraints and format. Most weak outputs trace back to a request that left the tool to guess.
Handling information
Knowing what may and may not go into an approved tool under the company's policy, and using fictional or approved material when unsure.
Drafting and transforming
Using the tool to produce, shorten, restructure or adapt text for a channel, then editing it into the writer's own voice.
Analysing and summarising
Asking for a summary or comparison of supplied material and keeping the summary inside what the material actually says.
Checking output
Opening sources, recomputing important figures, removing invented claims and spotting confident errors. NIST calls the underlying risk confabulation: confidently stated but erroneous or false content.
Explaining the edit
Being able to say what was changed and why. This is the skill that makes the other six reviewable by a manager or a colleague.
What does a good AI skill look like in practice?
Take a fictional 14-person design agency, Brightwater Studio. An account manager, Leila, needs a weekly update for a client. The supplied facts are: the homepage design is approved, the pricing page is waiting on the client's legal review, and the launch date of 12 November depends on that review finishing by 3 November.
A weak request is: "Write a client update saying the project is going well." A typical result is upbeat and vague. It might say the project is "on track for launch" and that "all key pages are complete". Both statements are false on the supplied facts. Nothing in the request told the tool what was true, so it filled the gap with plausible reassurance.
A stronger request is: "Draft a short weekly update for a client marketing lead. Use only these facts: [the three facts]. State the dependency on their legal review clearly and ask them to confirm a date. Plain, friendly tone, under 120 words. Do not add any facts that are not listed." The output is now close to usable, but Leila still checks it. Suppose it says the launch "will go ahead on 12 November". She edits that to "can go ahead on 12 November if the legal review finishes by 3 November", because the original turned a condition into a promise.
Every skill in the previous section appears in that small task. Leila chose a task that suits a tool, briefed it with the real facts, kept confidential contract details out, drafted, checked the result against the supplied facts and could explain her one edit. That explanation is what her manager would review. The quality of the first prompt matters, but the check and the edit are where the skill shows.
How do AI skills differ by role?
The seven skills are the same everywhere, but the task, the evidence and the most common failure change by role. That is why generic training often feels abstract: the examples come from someone else's job. The table shows where the emphasis tends to fall. It is an editorial guide, not a finding from research.
| Role | Typical first task | Skill that matters most | Common failure to watch |
|---|---|---|---|
| Marketing | A campaign email from a brief | Checking claims | Invented results, scarcity or endorsements |
| Sales | A follow-up email from call notes | Briefing with context | Commitments the rep never made |
| Customer support | A reply to a policy question | Checking output against policy | A confident but wrong policy detail |
| HR and recruiting | Interview questions from a role brief | Handling information | Personal data pasted into an unapproved tool |
| Finance | A short explanation of a variance table | Analysing and summarising | A cause stated that the data does not show |
| Operations | A process document from rough notes | Explaining the edit | Steps that read well but skip a real control |
Want to see how this connects to a role?
Explore the planned team trial ↗Which skills stay with people?
Judgement about what is true, what matters and who is accountable stays with people. A tool can draft a performance review, but a manager has to decide what is fair. It can summarise a contract, but someone qualified has to decide what the contract means. It can suggest a reply to an upset customer, but the agent has to decide whether the reply is honest and kind.
Domain knowledge also becomes more valuable, not less. The person who knows the product, the client or the regulation is the one who can spot that a fluent draft is wrong. Relationships, negotiation, ethical choices and the decision to send something are not tasks to hand over. Good AI skills training makes those human responsibilities more visible rather than hiding them behind a faster first draft. Any prediction about which jobs will change most is speculation, and this guide does not make one.
How do you learn AI skills?
You learn them the way you learn any working skill: by doing a real kind of task, getting specific feedback against a clear standard and trying again on a new example. Reading about prompting helps for an hour. After that, practice with review is what changes behaviour.
Pick one recurring task
Choose something you do every week that has a clear right answer to check against, such as a status update, a customer reply or a summary of a meeting.
Use safe material first
Practise with fictional or already approved information until you understand what your tool and your company policy allow.
Write down the standard
Before you start, note what a good result must contain and what it must not contain. That becomes your rubric.
Compare a weak and a strong brief
Run the same task twice, once with a vague request and once with full context, and note what changed in the output.
Check and edit
Verify every factual claim against your source, fix what is wrong and write one line explaining each change.
Get feedback
Ask a colleague or a reviewer who knows the work to read your final version and your explanation, not just the output.
Repeat on a new example
A second attempt on the same task partly tests memory. A new but similar task shows whether the skill transfers.
What free AI skills training is available?
Several free or low-cost options exist, and for an individual they can be a good start. On 28 January 2026 the UK government announced an expansion of its AI Skills Boost programme, describing free online AI foundations courses that take under 20 minutes each, open to every UK adult, with a virtual badge for learners who complete a course benchmarked against all six AI foundation skills.
Google's AI Essentials page, checked on 16 October 2026, describes a self-paced course on Coursera that it says takes under five hours, with modules on AI tools, prompting and responsible use, and a Google certificate on completion. It lists a subscription of 49 US dollars per month after a 7-day free trial in the US and Canada, with prices varying elsewhere. The help centres of the tools themselves, such as OpenAI, Anthropic, Google and Microsoft, also publish practical guidance at no cost.
These options are good at general foundations. What they cannot do is use your team's roles, tools and review standards, or tell a manager whether a particular person can now handle a particular task. That gap is why a team usually needs some form of role-specific practice after, or instead of, a general course.
How should a manager build AI skills across a team?
Start with the conditions for practice, then the practice itself. A team cannot build AI skills if nobody knows which tool is approved, what information may go into it or who reviews the output. Settle those first, even in a one-page policy, and make them easy to find.
Then choose one task per role, set a visible standard and give people short, regular time to practise. Ten to fifteen minutes between meetings is easier to sustain than a half-day workshop that is forgotten by the following week. Review the work, not the logins. If your organisation is in scope of the EU AI Act, there is also a general AI literacy obligation to consider; the AI literacy guide explains what the European Commission says about it.
Expect uneven progress. In any team some people already use AI daily without checking it, some avoid it entirely, and a few use it carefully. They need different starting points, which is the argument for a short diagnostic before assigning anyone a path. The AI upskilling guide sets out a step-by-step programme.
How do you know whether someone has an AI skill?
You know by looking at a piece of work they did on a defined task, against criteria they could see in advance. A self-rating tells you about confidence. A completion record tells you a course was finished. Neither shows whether someone removed the invented claim from Leila's client update.
Keep those signals separate in any report, and state what a work sample does not show: it is evidence about one task under certain conditions, not a prediction of performance in every workflow. The guide to assessing AI skills covers task design, rubrics, repeat attempts and how to report results honestly.
How is Tosheo.ai planned to help?
Tosheo.ai is being built as self-serve AI skills training for teams of 5 to 200 people. The product is pre-build, so everything here is planned scope. A manager is intended to pick the team's roles and the AI tools it uses, such as ChatGPT, Claude, Gemini, Copilot, Notion AI or Canva AI, optionally upload 1 to 5 company documents for context, and invite the team.
Each learner is planned to take a 5-minute diagnostic and then follow an adaptive, role-specific path of 10 to 15 minute lessons: a short explanation, a worked example from their role and a hands-on task with a real AI tool, graded by AI against a visible rubric with specific feedback. Learners will be able to flag a score they disagree with for review. Managers are intended to see skill progress and completion, not raw chat logs. A final practical assessment is planned to lead to a recognised certificate with a verifiable URL.
Tosheo.ai is for building AI skills in people you already employ. Testing candidates during hiring is a different job with different consequences, and it belongs in a fair, defined hiring process rather than in a training product.
When is this approach not the right answer?
If your goal is technical AI work, such as building models, fine-tuning or engineering AI features into software, workplace skills training is the wrong tool. Look for programming and machine learning courses instead.
If you are an individual learning for your own career, a free public course or a general certificate may serve you better than a team product. Tosheo.ai is planned for teams with a manager, and its pricing has a five-seat minimum.
If your organisation needs single sign-on, LMS integration, procurement contracts or lessons in languages other than English, those are not included in the planned launch scope. And if the team has no approved tool or no basic policy yet, fix that first. Training people to use a tool they are not allowed to use with real work creates confusion rather than skill.
Sources and scope
Skills England on GOV.UK, AI foundation skills for work benchmark, 28 January 2026 ↗
GOV.UK, Free AI training for all as government and industry programme expands, 28 January 2026 ↗
Anthropic Help Centre, What are Skills?, checked 16 October 2026 ↗
Google, Google AI Essentials course page, checked 16 October 2026 ↗