How to use AI at work: a practical operating manual

Using AI well at work is a set of habits, not a piece of software. The professionals who get the most from it treat the assistant like a fast, well-read new hire: they give it context, ask for a specific deliverable, check the work, and keep notes on what worked. This guide is the operating manual for building those habits.

Step 1: pick one tool and learn it well

Start with a single general-purpose assistant and use it every day for two weeks before you add anything else; juggling four tools teaches you none of them.

The short version of the decision:

If you are choosing on your own, the differences between the three matter less than learning one of them properly. Pick the one whose interface you like; the comparison guide can wait until you are ready to specialize.

Whatever you choose, spend the first hour on three things: find the setting that controls whether your conversations are used for training and set it the way your policy requires, learn how to attach a file, and learn how to start a fresh conversation. Long chats drift; a clean new chat usually beats the tenth message in an old one.

The five daily use cases

Almost all office work that AI helps with falls into five buckets. Learn one per week and you will cover most of your job within a month.

1. Writing and rewriting

Emails, status updates, proposals, job descriptions, meeting agendas. Never ask for a blank-page draft when you can give raw material instead: paste your bullet points, the email you are replying to, or a rough voice-memo transcript, and ask for the finished version in a stated tone and length.

Ask for three versions when the stakes are high and pick the best parts of each; one version when they are low.

What to check: names, dates, commitments, and anything you did not put in the raw material. If the draft mentions a deadline you never gave it, delete the deadline.

2. Summarizing and extracting

Long email threads, contracts, reports, meeting transcripts, and industry newsletters. Ask for a summary with a fixed structure: decisions made, open questions, action items with owners, and anything that surprised you. Then ask follow-up questions of the same document instead of re-reading it.

What to check: spot-check two or three claims against the source, and never forward a summary of a document you have not opened yourself. Models sometimes add a plausible detail that is not in the text.

3. Research and learning

Explaining an unfamiliar concept, comparing two approaches, getting up to speed on a regulation, or preparing questions for a specialist. Ask for an explanation at the level of a smart colleague from another department, then ask it to quiz you.

For anything current or factual, use a tool with web search turned on and click the sources. Models have a training cutoff and will not know about last month's rule change unless they can look it up.

What to check: every fact you intend to repeat. Treat the answer as a briefing from someone who is usually right, not as a source you can cite.

4. Spreadsheets and data

Writing formulas, explaining someone else's formulas, cleaning messy exports, drafting pivot-table logic. Paste a few rows of sample data (anything sensitive removed) and the exact output you want, and ask for the formula plus a one-paragraph explanation of how it works.

What to check: run the formula on rows where you already know the answer. For any arithmetic the model does in prose rather than in a formula or code, recompute it yourself.

5. A thinking partner for plans and decisions

Pressure-testing a plan, listing what could go wrong, turning a vague goal into milestones, or arguing the other side of a decision you have already half-made. This is the use case people discover last and value most.

Ask it to be critical on purpose: "List the five strongest objections a skeptical CFO would raise to this plan, then tell me which two I have no good answer for."

What to check: whether the objections are real in your context. The model does not know your company's politics, history, or constraints unless you tell it.

A prompting routine that works every time

You do not need clever tricks, only a repeatable structure. Every good work prompt answers five questions:

1. Role: who should the AI be? ("You are an experienced operations manager who writes clearly.")
2. Task: what exactly do you want? ("Draft a status update for the leadership team.")
3. Context: what does it need to know? (Paste the notes, name the audience, give the history.)
4. Format: what should the output look like? (Length, headings, bullets, tone, a table.)
5. Constraints: what must it not do? ("Do not invent figures. Ask me before assuming anything.")

The last item is the one most people skip and the one that prevents the most damage: tell the model to ask clarifying questions when your inputs are thin, and it will, instead of guessing.

A worked example

Here is a weak prompt a project manager might type on a Monday morning:

"Write a status update for my project."

The result will be generic and padded. The same request with the five parts filled in:

You are an experienced project manager who writes crisp, honest status updates for senior leaders.

Task: Draft this week's status update for the [PROJECT NAME] project.

Context:
- Audience: [WHO READS IT, E.G. THE VP OF OPERATIONS AND TWO DIRECTORS]
- My raw notes from this week: [PASTE YOUR NOTES]
- Last week's update, for continuity: [PASTE LAST WEEK'S UPDATE]
- Overall status this week: [GREEN / YELLOW / RED] because [ONE-LINE REASON]

Format: Under 200 words. Four labeled sections: Progress, Risks, Decisions needed, Next week. Plain sentences, no jargon, no exclamation marks.

Constraints: Use only facts from my notes. Do not invent dates, numbers, or names. If anything important seems missing (an owner, a due date, a decision), ask me up to three questions before writing.

The first version takes ten seconds to type and twenty minutes to fix. The second takes two minutes to fill in and usually needs one small edit. Save it: it is the first entry in your prompt library.

Verification habits

AI output is a draft, never a source. The failure mode that hurts professionals is not obvious nonsense; it is fluent, confident, specific text that happens to be wrong: invented citations, numbers rounded into existence, misattributed quotes, gaps in your data filled with plausible guesses. AI hallucinations explained covers why.

Build these habits from day one:

• The signature test. Before anything AI touched goes out, ask whether you would sign your name to every sentence. If you cannot vouch for a claim, check it or cut it.
• Check the five risky categories. Names, numbers, dates, quotes, and citations. These are where models fail most and where errors cost the most.
• Ask the model to mark its uncertainty. Add "Flag anything you are not certain about and tell me what I should verify" to prompts where accuracy matters.
• Use grounded tools for facts. Search-enabled assistants and Perplexity show sources; NotebookLM stays inside your documents. Click the sources anyway.
• Never verify with the same model that wrote it. Asking "are you sure?" produces a confident yes far too often. Check against a primary source instead.

The full process is in How to fact-check AI output.

Privacy rules

The rule that matters most: never paste confidential data into a consumer AI tool unless your organization has approved it. "Confidential" includes customer lists, financials that are not public, employee information, contracts, source code, and anything covered by an NDA.

On some consumer plans your conversations may be used to improve the vendor's models unless you opt out; business and enterprise plans typically include contractual commitments not to train on your data, plus admin controls and retention settings. Check the vendor's current documentation and your company's policy rather than assuming.

Practical habits:

• Strip identifiers before pasting. Replace names with roles ("the client", "Employee A"), remove account numbers, and round figures that are not needed for the task.
• Know your regulated categories. Health information (HIPAA and PHI), personal data (PII), privileged legal material, student records (FERPA), and financial data under SOX or GLBA each carry their own rules. If your role touches any of these, ask your compliance or legal team which tools are approved before you paste anything.
• Prefer the tools your employer already licenses. Copilot inside Microsoft 365 or Gemini inside Google Workspace inherit your company's data agreements.
• Assume chats are discoverable. Treat anything you type as something that could be read in a records request or an audit.

If your team has no written policy, How to write a team AI policy and the AI policy generator will get you a one-page version in an afternoon.

Build a personal prompt library

The difference between people who "tried AI once" and people who use it every day is usually a document. A prompt library is a plain text file, a note, or a spreadsheet where you save prompts that worked, with enough notes to reuse them.

Keep four fields for each entry:

1. Name. "Weekly status update", "Rewrite for a customer", "Explain this formula".
2. When to use it. One line.
3. The prompt. With [BRACKETED PLACEHOLDERS] for the parts that change.
4. Notes. What to check afterward and any tweak that improved results.

Start with five entries: the status update above, an email rewrite, a meeting-notes summarizer, a formula helper, and a "poke holes in my plan" prompt. Add one entry a week. By the end of the quarter you will have a toolkit that fits your job exactly.

Most assistants also let you save standing context: Projects in ChatGPT and Claude, Gems in Gemini, custom GPTs in ChatGPT. Put your role, audience, writing preferences, and standard checks in there once, and every conversation starts warmed up. Custom GPTs vs. Claude Projects vs. Gemini Gems compares them; the prompt builder will structure new prompts for you.

Your 30-day plan

The plan is deliberately light. The goal is a habit, not a certification.

Week 1: setup and writing

• Choose your tool using the table above. Set the data-use settings your policy requires.
• Use it for every email longer than a paragraph. Paste your bullets, ask for the draft, edit, send.
• Save your first two prompts to your library.

Week 2: summarizing and research

• Every long document or thread you receive goes into the assistant first, with the four-part summary structure (decisions, open questions, actions, surprises).
• Use a search-grounded tool for one real research question, and click every source it gives you.
• Add two summarizing and research prompts to your library.

Week 3: spreadsheets and thinking

• Bring one messy spreadsheet task to the assistant. Test the formula on known rows.
• Run one plan or decision through the "strongest objections" prompt before a meeting.
• Set up a Project, Gem, or custom GPT with your standing context.

Week 4: review and share

• Look back at the month: which tasks got faster, which outputs needed the most fixing, which prompts you reused most.
• Write a checklist for doing one recurring task with AI, including the verification step. That checklist is the seed of a team practice.
• Share your three best prompts with a colleague. Teaching them is the fastest way to find the gaps in your own process.

At the end of the month, run the time-savings calculator with honest numbers to see where the gains came from.

Next steps

• Learn the underlying ideas in What is generative AI? and How large language models work.
• Go deeper on prompting with the prompt engineering guide and the ready-made prompts for writing and email.
• Tighten your checking with How to fact-check AI output and your data habits with AI privacy at work.
• Compare tools in Best AI tools for work, or read the ChatGPT, Claude, and Gemini reviews.
• Find your role's playbook under AI for your job, or take the AI readiness quiz.
• The course walks through all of this with exercises.