What does using AI in your job actually involve?
For most office roles, using AI at work means three kinds of help: drafting from material you supply, transforming material from one shape into another, and acting as a thinking partner that suggests options, questions and gaps. It rarely means letting a tool make a decision or produce a final piece of work unreviewed.
Drafting covers emails, summaries, job adverts, policy explanations and first versions of reports. Transforming covers turning long documents into short ones, notes into structured records, one message into versions for different readers, or a table into plain-English commentary. The thinking-partner role covers questions such as 'what would a sceptical buyer ask about this?' or 'what is missing from this plan?'.
In all three, the quality of the result depends far more on what you give the tool and how you check what comes back than on which tool you use. That is why the same habits appear in every role guide below, adapted to the work and the risks of each job.
How do you choose the first task to use AI on?
Choose a task you do often, already know how to do well and can check quickly. Your existing skill is what lets you spot a confident mistake, so a task you are unfamiliar with is a poor place to start. The criteria below narrow most people's options to two or three candidates.
It recurs
You do it weekly or more, so any improvement is repeated and you get enough practice to judge the tool.
You can judge quality
You know what a good version looks like without needing someone else to tell you.
The source is available and shareable
The material the task depends on exists, and your company's policy allows it in an approved tool.
A mistake is recoverable
An internal draft or a first version for your own editing, not a figure in a board pack or a message to a regulator.
It has a clear finish
You can tell when the output is ready, which makes it easy to compare with doing it yourself.
What should a good work prompt contain?
A good work prompt contains what a capable new colleague would need to do the job without coming back to you: the purpose, the reader, the source material, the constraints, the format and what to flag. Missing information is the main reason AI output sounds generic or invents details.
Compare two prompts for the same task. The weak version is: 'Write an update for the team about the office move.' The stronger version is: 'Write a 150-word update for our 25 staff about the office move. Use only the facts below: the move is on Friday 14 November, the new address is in the notes, desks must be cleared by 4pm Thursday, IT will move monitors and docks. Plain, friendly tone. End with who to ask. Do not add dates, rules or reassurances not in the notes. List anything staff are likely to ask that the notes do not answer.'
The last instruction is worth adopting everywhere. Asking the tool to list what it could not answer turns a gap it would otherwise fill with a guess into a question you can resolve. The prompting guide linked at the end goes further into structure and templates.
How should you check AI output before you use it?
Check the output against the source, not against your sense of whether it sounds right. Fluent text is easy to trust. NIST's Generative AI Profile describes confabulation as the production of confidently stated but erroneous or false content, and lists over-reliance and automation bias among the risks in how people and AI systems work together. Both are reasons to check deliberately rather than by feel.
A short routine covers most work. First, trace every factual statement, name, date and figure to the source you supplied, and remove or confirm anything you cannot trace. Second, recalculate numbers independently. Third, look for what was added: claims, promises, causes or rules that sound normal for this kind of document but are not true for you. Fourth, read it as the intended reader would and adjust tone. Finally, decide who is accountable for the final version, which is usually you.
The depth of the check should follow the consequence of an error. A draft for your own notes needs less than a message to a client, which needs less than a figure that will be reported.
What are good AI use cases by role?
The table below gives a sensible first task for each role, the check that matters most and the part of the work that should stay with a person. Each role has its own guide with a worked example.
| Role | Good first task | Check that matters most | Keep with a person |
|---|---|---|---|
| Marketing | Channel variants from an approved piece | Every claim and figure traced to the brief | Positioning and claim sign-off |
| Sales | Follow-up email from your call notes | No invented outcomes or case studies | Prices, terms and commitments |
| HR | Plain-language summary of an existing policy | Entitlements match the policy word for word | Decisions about individual people |
| Recruiting | Job advert from hiring manager notes | Requirements match; inclusive language | Screening, ranking and rejection |
| Finance | Variance commentary from a checked table | Figures recalculated; no unconfirmed causes | Reported numbers and forecasts |
| Accounting | Client email requesting missing records | No tax rules or deadlines added | Technical positions and sign-off |
| Customer service | Reply draft from the help article and ticket | Policy and refund terms are accurate | Exceptions and upset customers |
| Project management | Status update from task notes | Dates and owners are real | Re-planning and trade-offs |
| Product management | Theme summary of customer feedback | Themes trace to real comments | Prioritisation decisions |
| Operations | Process document from your steps | Steps match how work is really done | Exceptions and approvals |
| People managers | Agenda and notes for a team meeting | Nothing attributed to people wrongly | Feedback and performance judgements |
What changes from one role to another?
The habits stay the same across roles; the risks change. Three questions explain most of the difference: who is affected if the output is wrong, what information the task involves, and whether a rule already governs the output.
Roles that make or influence decisions about people, such as HR, recruiting and management, carry the most scrutiny. The EU AI Act lists AI systems used for recruitment and selection, and for decisions on promotion, termination, task allocation and performance evaluation, among its high-risk uses in Annex III. Roles that publish claims, such as marketing and sales, answer to advertising and direct marketing rules that apply however the text was produced. Roles that report numbers, such as finance and accounting, answer to professional standards and to the people who rely on the figures.
Roles that handle customers, such as support and account management, carry a different risk: an AI-drafted reply that misstates a policy becomes a promise the company may have to honour. In each case, the role guide explains which checks matter most and why.
Want to see how this connects to a role?
Explore the planned team trial ↗How does the same task look different in three roles?
Take one fictional source: release notes for a new feature at Fernhill Software, a small company selling scheduling software to clinics. The notes say the feature lets clinics send automatic appointment reminders by text message, is included in the Professional plan, launches on 1 December and does not yet support languages other than English.
A marketer asks AI for a launch email to existing customers. The draft says reminders 'reduce no-shows dramatically'. The release notes contain no data on no-shows, so the marketer cuts the line and describes what the feature does instead.
A salesperson asks AI for a follow-up to a clinic that asked about reminders on a call. The draft says the feature is 'available on all plans'. It is not. The salesperson corrects it to the Professional plan and checks the customer's current plan before sending.
A support agent asks AI to draft a help centre answer about reminders. The draft says customers can 'choose any language'. The notes say the opposite. The agent corrects it and adds who to contact for a language request.
Same source, same tool, three different errors, each one typical of its role. That is why practising on tasks from your own job teaches more than generic examples.
What information should you keep out of AI tools at work?
Follow your company's AI policy first; if there is none, ask before pasting anything that is not already public. As a general rule, be most careful with personal data about customers, candidates or colleagues, confidential client material, unpublished financial information and anything covered by a contract or non-disclosure agreement.
The answer also depends on the tool and the plan. A personal account, a company workspace and an AI feature inside another application can each have different terms on storage, access and whether content is used to improve models. Removing names does not by itself make a document anonymous, particularly in a small organisation. The AI policy guide covers how to set these boundaries for a team.
How do you know you are getting better at using AI at work?
Look at your edits, not your output volume. Someone becoming skilled with AI at work tends to rewrite less of each draft because their prompts improve, catches invented details faster, and can explain why they changed what they changed. They also become quicker at deciding when not to use AI at all.
A simple way to track this is to keep a short log for two or three weeks: the task, the prompt, what you changed and why. Patterns appear quickly, such as a tool that always adds a reassurance you have to remove or a type of task where the first draft is consistently usable. Those patterns are more useful than any general claim about time saved, because they come from your own work.
How should a team learn to use AI in their jobs together?
Give the team a shared, realistic task, let everyone attempt it with AI and then compare the edits rather than the drafts. Comparing edits shows who checks sources, who catches invented claims and who keeps the voice, and it spreads those habits faster than a slide deck.
Keep practice on fictional or approved low-risk material, agree which tools are approved, and make it safe to report a mistake. For organisations within scope of the EU AI Act, Article 4 asks providers and deployers of AI systems to take measures to support AI literacy among staff, taking into account their knowledge and the context of use. This is general information, not legal advice. The employee training guide and the 30-day rollout plan linked below set out a fuller programme.
When is AI the wrong tool for the job?
AI is the wrong tool when you cannot check the result, when the decision affects a person's job, pay or candidacy, when the output is a final reported figure, and when the conversation needs a human who is accountable for what is said. It is also the wrong tool when the task is quicker to do directly, such as a two-line reply you already know how to write.
Be cautious too when there is no source. Asking a general assistant what your company's policy, your client's contract or the current tax rule says produces a plausible answer, not the document. In those cases, find the source first and use AI, if at all, to help you read it.
Where does Tosheo.ai fit, and where does it not?
Tosheo.ai is being built as self-serve AI training for teams of roughly 5 to 200 people. A manager picks roles and the AI tools the team uses, each learner takes a short diagnostic and then follows a role-specific path of 10 to 15 minute lessons, each ending in a hands-on task with a real AI tool, graded against a visible rubric. Managers are planned to see skill progress and completion, not raw chat logs. The product is pre-build, so all of this is planned scope.
It is not the right choice for every buyer. It is not designed for large organisations whose purchasing requires single sign-on (SSO), LMS integrations or enterprise contracts, none of which are included at launch. It is not hands-on engineering training, and lessons are planned in English only. If you are an individual learning on your own, a general course or the free guides on this site may suit you better. Planned team pricing is on the pricing page.
Sources and scope
European Commission AI Act Service Desk, Regulation (EU) 2024/1689 Annex III, high-risk AI systems ↗
European Commission AI Act Service Desk, Regulation (EU) 2024/1689 Article 4, AI literacy ↗