How do I use AI to automate boring parts of my job?

Start by logging your week before automating anything. Spend five days noting every repetitive task and roughly how long it takes. Then automate the one with the highest frequency-times-duration, not the one that's most annoying — those are usually different, and people reliably automate the annoying-but-rare task first and save nothing. The tasks that automate well with current tools: 1. **Drafting repetitive written communication.** Status updates, standard replies, meeting follow-ups. Build a prompt template with your context and tone, paste in the specifics, edit the output. Saves most of the drafting time and all of the blank-page friction. 2. **Summarising documents and meetings.** Genuinely reliable, because it's grounded in text you provide rather than recalled. Meeting transcripts to action items is one of the highest-value uses available right now. 3. **Reformatting and restructuring data.** Messy text to a structured table, one format to another, extracting specific fields from unstructured documents. Very reliable and very common. 4. **First-pass analysis.** 'Here are 200 customer comments, group them by theme and give me counts.' Excellent as a starting point that you then verify. 5. **Research scaffolding.** Drafting an outline, generating the questions to investigate, producing a first structure you rewrite. The approach that makes it stick: - **Build reusable prompt templates rather than starting fresh each time.** Keep a document of prompts that worked, with placeholders. This is the difference between saving five minutes occasionally and saving hours every week, and almost nobody does it. - **Include your context in the template** — your role, your audience, your style constraints, an example of a good output. Generic prompts produce output you have to rewrite entirely, which saves nothing. - **Automate the draft, not the decision.** The reliable pattern is AI produces, you review and own. Anything that goes out without human review needs to be a task where errors are cheap. - **Chain simple steps** rather than asking for one complex transformation. More reliable and easier to debug when it goes wrong. What not to automate: anything requiring accountability or judgement about people, anything where a subtle error is expensive and hard to spot, and anything involving confidential data unless you're on an approved tool. One warning worth taking seriously: measure whether you actually saved time. It's easy to spend forty minutes engineering a prompt for a task that took fifteen minutes and happens monthly. Automate what's frequent, and be honest about the payback period — the same discipline you'd apply to any other efficiency project.