How to use AI at work: 8 practical ways to get started

You don't need a complicated workflow to start using AI at work. You need a task that occurs often, has a clear result, and won't cause much damage if the first output is imperfect.

These eight strategies apply across most roles and can be used with any tool permitted by your employer. Each includes a starting point and a validation check to ensure efficiency doesn't compromise accuracy.

How to get started with AI at work

AI can save time, but the first output often creates more work. In Workday's 2026 Beyond Productivity study, 85% of active AI users said they saved 1–7 hours a week, while nearly 40% of those savings went back into correcting, rewriting, and checking AI output. Only 14% consistently reported a clear positive result.

Those figures come from active AI users at large organizations, so they aren't a promise about what you'll save. They do show why choosing the task, writing a clear prompt, and planning the review matter as much as the tool.

[[cta-big]]

Choose the task before the tool

Start with a task you already understand. A good first use case is small, recurring, easy to describe, and easy for you to judge. Write down the result you want before you open an AI tool. “Turn these notes into a list of decisions and owners” gives you a visible finish line; “make me more productive” does not.

Use four questions to screen a task:

There are also two useful ways to work with AI:

• Use it as an assistant to transform material, organize information, or create a rough draft.
• Use it as an advisor when you want a critique, another perspective, or practice before a difficult conversation.

Write better prompts with CRAFT

Useful AI output starts with a good brief. A framework we use in our courses is CRAFT: Context, Role, Action, Format, and Tone. It helps you give the model the same direction you would give a capable new teammate.

You don't need to write a long prompt for every task. Include enough CRAFT to remove the model's guesswork. The prompt callouts below use all five parts where they add value, along with guardrails that specify what the model must not invent.

Check the rules and the data

Using an unapproved tool is common, but common does not mean permitted or safe. BlackFog's 2026 Shadow AI research found that 49% of 2,000 surveyed workers had used AI tools their employer had not sanctioned. Among that group, 58% relied on free versions.

Before you put work into an AI tool:

• Check your employer's acceptable-use, security, and data-handling policies.
• Use an approved account or tool when the policy requires one.
• Remove client names, employee data, personal identifiers, and confidential information.
• Ask IT, security, or your manager when the rule is unclear.

Decide how you'll review the answer

Ask one question: Did the model have the relevant source material in front of it, or did it have to recall or infer the answer?

Rewriting your paragraph or summarizing a document you supplied is usually easier to check because the source is available for comparison. A statistic, definition, policy detail, or citation needs more scrutiny because the model may be relying on incomplete training data or filling a gap. In either case, increase the review when the cost of an error is high.

Danoosh Kapadia, an instructor and subject-matter expert in our AI Essentials course, offers a useful mental model: “Think of your LLM as a really, really smart intern.” Give it real work and enough context, then read the result before it goes out under your name.

How to use AI at work across everyday tasks

1. Rewrite something for a different reader

Start with text you've already written. Give the model the original text, name the reader, and explain what matters to them. Specifying the audience and its priorities gives you a more useful result than asking the model to “make this professional.”

Prompt to try

Act as an experienced business editor. Rewrite the text below for [reader or audience], who cares most about [priority]. Lead with [most important point], preserve every fact, deadline, decision, and commitment, and do not add claims or reassurance that are not in the original. Format the result as [an email, an update, or a brief] of no more than [length]. Use a [direct, warm, concise, or confident] tone. After the rewrite, list any changes in meaning or details you removed. [Paste text.]

Because you supplied the facts, your main review is for drift. Check whether the rewrite added a promise, softened a deadline, changed the decision, or removed a necessary caveat.

Best for: emails, updates, proposals, and explanations that are accurate but poorly matched to the reader.

2. Turn notes or a long document into an action summary

Give the model the source and request a specific output. Ask for decisions, open questions, owners, and due dates rather than a generic summary. For a report, ask which sections affect your work and where those points appear in the source.

Prompt to try

Act as a project coordinator reviewing the source material below for [team or audience]. Using only the information provided, extract the decisions, action items, owners, due dates, open questions, and risks. Format the result as a table with one row per item and include the page, section, or note where each item appears. Write “not stated” when an owner or date is missing, and do not infer one. Use a concise, neutral tone. End with three questions the team still needs to resolve. [Paste or attach the source.]

Compare the result with the original. Look especially for omitted caveats, dissenting views, and unclear ownership. A summary is a shortcut to orientation, not a substitute for reading a contract, policy, or document you're accountable for understanding closely.

Best for: meeting notes, long email threads, research reports, and background documents.

3. Build a first draft from your own bullet points

When the blank page is slowing you down, give AI the raw material already in your head. Name the audience, format, purpose, length, required points, and anything it must not claim. Ask for a rough draft you can reshape.

Prompt to try

Act as a first-draft writing partner. I need a [document type] for [audience] that helps them [desired outcome]. Use only the facts and bullet points I provide below. Organize the draft with [headings, sections, or other format], keep it to approximately [length], and use a [tone] voice. Do not invent names, numbers, dates, policies, examples, or commitments. If the draft needs information I have not supplied, insert [DETAIL NEEDED] instead of filling the gap. After the draft, list the three details I should verify first. [Paste bullet points.]

This is riskier than rewriting because the model has room to invent connective details. Review every specific name, number, date, policy, and commitment. If you can't identify the source of a detail, remove it or verify it independently.

Best for: briefs, proposals, emails, and outlines when you have the facts but no first draft.

4. Practice a difficult conversation

Describe the conversation, the other person's likely priorities, and the objections you expect. Ask the model to play that person. Run the conversation once, then ask it to be more skeptical or direct on the second attempt.

Prompt to try

Act as [the other person's role] in a practice conversation about [topic]. The situation is [brief context], my goal is [desired outcome], and your likely priorities are [priorities]. Role-play the conversation with me one response at a time. Raise realistic objections, ask the difficult follow-up questions I may hear, and do not invent private motives or facts about the person. Keep your tone [skeptical, direct, calm, or frustrated]. When I type “debrief,” stop the role-play and give me feedback on my opening, clarity, listening, and response to pushback.

The value is rehearsal. You get to test your wording and hear possible pushback before the real conversation. The model's interpretation of your colleague is still a guess based on your description, so don't treat the role-play as proof of what that person thinks.

If you've been avoiding a conversation, spend 10 minutes rehearsing the opening and the most likely difficult question.

Best for: salary conversations, deadline negotiations, client pushback, and other conversations where rehearsal helps.

5. Pressure-test work from a stakeholder's point of view

Name the reviewer and their priorities. Instead of asking for general feedback, ask the model to examine your work through the lens of that stakeholder's specific concerns and surface the strongest argument against your recommendation.

Prompt to try

Act as a [stakeholder role] reviewing the [proposal, plan, or presentation] below. Your priorities are [priorities], and you will be skeptical of [likely concern]. Identify the five places where you would push back. Format the response as a table with the objection, why it matters, evidence needed, and a question I should be ready to answer. Then give the strongest argument against my recommendation and one revision that would make the case more credible. Use a direct, rigorous tone. Treat your objections as hypotheses, and do not claim to know what the real stakeholder thinks. [Paste or attach the work.]

Treat each objection as a hypothesis. Your experience tells you which concerns are realistic, which need evidence, and which a real stakeholder should confirm.

Best for: proposals, presentations, plans, and decisions that will be reviewed by a busy or skeptical reviewer.

6. Translate jargon you missed in a meeting

Give the model the unfamiliar phrase and enough context to distinguish a useful explanation from a dictionary definition. Then ask a question you can use to confirm your understanding with a colleague.

Prompt to try

Act as an expert familiar with [industry or function]. In a meeting about [topic], I heard the phrase “[term or sentence].” Explain what it most likely means in plain language for someone who understands [what you already know]. Give one short example, identify any part of the meaning that may vary by company or context, and suggest one natural question I can ask a colleague to confirm my understanding. Format the answer in three short sections and use a clear, nonjudgmental tone. Do not present an internal or uncertain meaning as fact.

Definitions can be wrong or outdated, especially when a term is internal to your company or industry. Use the explanation to get oriented, then verify it before repeating it in a client conversation or formal document.

Best for: cross-functional meetings, unfamiliar acronyms, and low-stakes concepts that need clarification.

7. Learn one part of a tool or skill

Tell the model exactly where you're stuck and what you already know. Ask for an explanation at that level, try the step yourself, and then ask the model to quiz you or have you explain it back. That last step shows whether you understood the idea or only recognized the wording.

Prompt to try

Act as a patient tutor for [tool, process, or skill]. I already understand [current knowledge], but I am stuck on [specific problem]. Help me complete [desired action] using the real, sanitized example below. Explain one step at a time, pause after each step so I can try it, and adjust as needed. Use plain language and a supportive tone. If a menu path, formula, or procedure may have changed, say so and point me to the current official documentation instead of guessing. When we finish, quiz me with three questions and ask me to explain the process back in my own words. [Add example or context.]

Bring a real example when policy permits, but remove sensitive information first. Verify software menu paths, formulas, and procedures in the current official documentation. Don't use trial and error for legal, financial, compliance, or safety-critical instructions.

If you've been working around the same skill gap for months, start with the last specific moment it blocked you.

Best for: software features, spreadsheet formulas, repeatable processes, and concepts you can test as you learn.

8. Explore a file or spreadsheet you already have

If your approved tool supports file uploads, begin with a question you genuinely want answered. Ask which themes recur in survey responses, which rows break the pattern, or what you should investigate next. “Analyze this” is too broad to provide a clear finish line for the work.

Prompt to try

Act as an exploratory analyst. Using only the attached, sanitized file, help me answer [specific question] for [audience or decision]. First, describe the columns or source material you used and flag missing data or limitations. Then identify the main patterns, exceptions, and follow-up questions. Format the findings as a table with the observation, supporting rows or source location, confidence level, and verification step. Show the calculation behind every total or percentage. Use a precise, neutral tone, separate observations from hypotheses, and do not infer sensitive characteristics or facts that are not in the file.

Remove personal, client-identifying, and confidential information before uploading. If you can't sanitize the file, wait until you have an approved environment designed for that data.

Summaries and theme extraction can be checked against the supplied file. Totals, percentages, formulas, and other derived figures need independent spot checks. Never act on a surprising number until you can reproduce it.

Best for: sanitized survey exports, project trackers, and spreadsheets you understand well enough to audit.

How to make AI use stick after the first few tries

An AI workflow won't last if opening the tool, rebuilding the context, and checking the answer take longer than the task. Keep the setup small enough to repeat.

• Reduce the setup. Bookmark or pin the approved tool and keep one reusable prompt for the task.
• Preserve the context. Save the sanitized prompt, useful output, and source links with the work.
• Keep the review checkpoint. Decide who will check the output and what they will check before the deadline arrives.

Review is part of productive AI use. In the Workday study, 77% of daily users said they reviewed AI-generated work at least as carefully as human-produced work. The NIST Generative AI Profile likewise recommends acceptable-use guidance, risk-based controls, and human review that matches the use case.

As Kapadia explains, “There should always be a point where a human verifies the machine's output.” Your first durable AI habit should include that point by design.

Pick one low-risk task from this list that will recur this week. Try it once, note what you had to correct, and use that correction to improve the prompt or the review next time.

FAQs: Using AI at work

Do I need my employer's approval before using AI at work?

Follow your employer's policy. If it requires an approved tool or explicit permission, then yes. When the policy is unclear, ask before entering work information, and start only with non-sensitive material in an environment your employer permits.

Can my employer see what I type into an AI tool?

Keep confidential, client-identifying, employee, and personal data out of unapproved AI tools. Exact visibility depends on the account, device, network, and monitoring controls, so don't assume a personal login makes activity private. Ask your IT or security team what is logged in your workplace.

How do I know when to trust an AI answer?

Check whether the model used source material you supplied or had to recall and infer the answer. Increase your review when the source is missing or the cost of an error is high, and always verify names, numbers, dates, citations, and policy details.

What is the easiest AI task to try first?

Rewrite a short piece of text you already wrote for a specific reader. You control the source facts, the result is easy to compare with the original, and the review is limited to catching changes in meaning, tone, deadlines, or commitments.