How Can AI Be Used at Work?
AI tools like ChatGPT, Claude, and Gemini are, in a word, a "fast assistant you can ask in plain words." Even without specialized knowledge, if you talk to it in ordinary language—"summarize this," "draft this email"—an answer comes back in seconds. Beyond text, you can also show it an image, PDF, or figure and ask "make a table from this." This article organizes things so a first-time user can concretely grasp "what it can be used for and how / where it is better not to leave things to it."
There Are Broadly "2 Ways to Use" AI
Even with the same AI, what comes back changes depending on how you ask. Knowing this difference first makes everything much clearer.
| Chat | Agent | |
|---|---|---|
| What it does | You ask a question → AI answers in text | You ask "do this for me" → AI actually does the work and finishes it |
| What you get | Advice or a draft in text | A finished file, or the result of the action it performed |
| Mental image | A knowledgeable colleague you can consult | An assistant who carries out the work for you |
Chat examples: "Summarize these minutes in 3 lines," "Draft a polite apology email," "List 5 risks of this plan." You copy the returned text and use it yourself.
Agent examples: Ask "look up candidate flights for next week's business trip and put them in a comparison table," and the AI browses the web itself and returns the comparison-table file itself. It can also take on "hands-on work" such as making an Excel file with formulas, actually running a program it wrote and producing the result, or filling out an application form (ChatGPT's agent mode, Claude's code execution, etc.). For important actions like sending or purchasing, it is common for it to ask for confirmation before executing.
There is also a third, embedded type. Inside the apps you always use—Notion, Microsoft 365 (Excel, etc.), Gmail / Google Docs, Slack—there is an AI button, and you can summarize or draft right there.
What Helps at Work Right Away (Strengths)
- Text work: drafting emails/reports, summarizing minutes and long documents, translation, wording adjustment, proofreading
- Prep work for thinking: recent AI has become good at "thinking step by step (reasoning)," useful for building a first draft of complex consultations, analysis, and comparisons
- Looking things up: when needed it searches the web on the spot and answers about the latest news and prices with source links (default in ChatGPT/Gemini). For deeper digging, "Deep Research" can produce a research report with citations
- Materials and data: the skeleton of a proposal, comparison tables, aggregating survey results and surfacing insights. It can also read images, PDFs, and figures and organize them
- Simple automation/code: Excel formulas and scripts for repetitive work even with no experience; code creation is an area that has become especially accurate in recent years
- Idea generation: drafts, counterarguments, alternatives, lots of naming options at once
Things Better Not Left Entirely to AI (Weaknesses / Cautions)
- Accuracy of facts and numbers: AI sometimes makes mistakes with full confidence (hallucination). Even smart models still misstate basic facts. For important numbers, proper nouns, and dates, open the source and check; for calculation, "have it build a formula and re-check."
- An entire very long document at once: if you hand over long text in one go, it can drop content from the middle. It is more stable to narrow to key points or split it up.
- Leaving a long sequence of steps entirely to it: Agent is convenient, but carrying many steps through end-to-end is still a weak point—it can lose sight of the goal partway. Proceed with a human checking at important milestones.
- Internal circumstances / final judgment: it doesn't know department history or unspoken premises. For conclusions involving hiring, investment, contracts, legal, or health, use it as a draft and let a human decide and take responsibility.
- Confidential information: don't carelessly paste customer lists, personal data, or confidential materials. For company use, choose business settings that are "not used for training and auditable."
Tips to Not Fail (3)
- Use it as a "draft": don't output AI's answer as-is; a human always reviews it last.
- Verify facts with sources: don't swallow search results either; look at the primary source behind the link.
- Decide what info is OK to input: make it a rule from the start—no personal or confidential data.
Let's Try It First
The shortcut is to make an account on one of ChatGPT / Claude / Gemini and try two things in real work: (1) a Chat like "summarize this text in 3 lines," and (2) a look-up like "find 5 latest pieces of info on XX with sources and put them in a table." In 10–30 minutes you'll get a feel for "where in my work I can use this." The next article, "The Difference Between Chat and Agent," looks at the two ways of using it in more detail.
