How to use AI at work: 5 tasks to automate and 7 to leave alone

Using AI at work means assigning specific tasks to AI tools while you retain the review and the final call. The skill that separates strong users from average ones is deciding which tasks to delegate, how much context to supply, and where to place the checkpoints.

Most advice on how to use AI at work still stops at summaries and email drafts. If you already use AI every day, you’ve outgrown it. This guide covers what changes at your level: setting up before you prompt, auditing what you already delegate, automating multi-step work, and recognizing the tasks that should stay yours.

How do you audit the tasks you already hand to AI at work?

You already use AI daily, so audit current use first. A 20-minute pass shows where you over-delegate and where you under-delegate.

Step 1: List what you already hand to AI. Check your chat history, saved prompts, and connected tools from the past two weeks. Add recurring tasks you haven’t delegated yet from your calendar, sent folder, and project tool. Aim for 15 to 20 tasks.

Step 2: Score each task on three questions.

1. Is it repeatable? Do you do it in roughly the same way each time?
2. Can you check the result cheaply, at a defined checkpoint?
3. Can you undo the damage if the result is wrong?

Step 3: Sort by score.

Compare the buckets with your habits. Tasks you already delegate that land in the Keep bucket are your over-delegation risks. Tasks in the Automate bucket that you still do by hand are your best next experiments. Start with two or three. Tasks that repeat on a schedule can later become an AI workflow.

Which 5 tasks should you automate with AI at work?

Most employees still use AI occasionally. Gallup found that 15% of US employees use it daily and 30% a few times a week or more as of May 2026. These five tasks go beyond one-off prompts. Each chains several steps, scores three yes answers on the audit, and comes with a guardrail. Each one below shows what the workflow looks like in practice.

1. Turn meetings and email threads into tracked work

A summary tells you what happened. This workflow turns it into work that moves. Say a 45-minute planning call ends. You give AI the transcript and ask it to list each decision, the person who owns the next step, and any deadline mentioned. It then creates a task for each item in your project tool and drafts one follow-up message that lists them. You review the list, fix anything it got wrong, and send.

Guardrail: Compare each owner and deadline with the transcript before tasks are assigned to anyone else. Check for invented owners and for firm dates built from a casual “maybe next week.”

2. Schedule recurring digests and monitors

Most AI waits for you to type. A scheduled job doesn’t. Say you want a Monday summary of your sales pipeline. You give AI read access to the pipeline data and a template: new deals, stalled deals, and deals that changed stage. It runs every Monday morning and sends the digest to your inbox. The same pattern works for customer feedback themes, or for a metric you want flagged when it moves more than 10% week over week.

Guardrail: Pin the sources it may use, require links to the underlying records in every digest, and review the first four runs before you trust the schedule. Tell it to flag low-confidence items instead of smoothing them over.

3. Prepare briefs from across your apps

Preparing for a client call usually means searching your inbox, a shared drive, and a chat history, then stitching notes together. Hand that search to AI. Ask it to pull the last three months of emails with the client, the latest proposal, and any open support tickets, then write a one-page brief covering the relationship history, open issues, and the two questions to raise. Because it works from your own files, the brief reflects your actual context instead of general knowledge.

Guardrail: Require a source link behind each claim, and open the two that carry the most weight before the call.

4. Extract and compare data across document sets

Say you’re choosing between six vendor proposals. Each lists pricing, contract length, and support terms in a different place and in different wording. AI can read all six, pull the same fields from each into one table (price per seat, term, renewal clause, support hours), and note where a proposal leaves a field blank. What used to take an afternoon of copying becomes a table you check. The same method works for contracts, invoices, and research reports.

Guardrail: Ask for the page or section reference behind each extracted value, and spot-check every row that affects the decision.

5. Run routine triage through agents with an approval gate

Triage means sorting incoming items and deciding what happens to each one, and an agent can do it on its own schedule. For example, it reads each new inbound email, labels it (sales inquiry, support request, contract, or noise), and proposes an action: create a lead, route to support, file the document, or archive. At first it only proposes, and you approve in a batch. Once its proposals match yours consistently, you can let low-risk actions, such as archiving newsletters, run automatically.

Guardrail: Keep a human approval step before any action that can’t be undone, and audit a sample of the agent’s decisions each week.

Which 7 tasks should you leave alone?

These tasks score one or zero on the audit. AI can still support your thinking, but the data, the decision, and the work product stay with you. The first four protect your company’s security.

1. Confidential data in any unapproved tool

Customer records, contracts, financial statements, board materials, unreleased plans, HR files, passwords, and API keys all fall here. Pasting them into a tool your company hasn’t approved can breach data protection laws and customer agreements, and you can’t pull the data back out. Anonymize records first, or use an approved tool with data controls.

2. Security decisions

Don’t let AI decide who gets access, how a firewall is configured, which vendor passes a security review, or how to respond to an incident. A confident wrong answer here exposes the whole company. Use AI to explain concepts and draft documentation, and leave configuration and approvals to your security team.

3. Sensitive meetings with an AI note-taker running

HR conversations, legal discussions, and performance reviews create transcripts that may be stored, shared, or discoverable. Check your policy, get consent from everyone present, and take your own notes when a transcript creates risk.

4. Actions you can’t undo, with no approval step

Agents that send external emails, accept invitations, or publish content act at machine speed, and a wrong send can’t be recalled. Agents that read untrusted content, such as inbound email, attachments, and web pages, add a security risk: Hidden instructions in that content can hijack the agent. OWASP ranks prompt injection as the top risk for AI applications. Let AI draft and queue the action, and keep the final click yours.

5. Asking AI whether your idea is good

AI models tend to agree with you. A study in Science that tested 11 leading AI systems found they affirmed users’ actions 49% more often than humans did. Ask AI to argue against your plan instead, then take the strongest objection to a colleague.

6. Ranking, scoring, or screening people

Performance ratings, resume shortlists, and bonus allocations carry bias and legal exposure, and many jurisdictions regulate AI in hiring and employee evaluation. Use AI to organize evidence, such as grouping feedback themes, and make each judgment yourself.

7. The first 50 reps of a new skill

Practice builds skill, and AI can take the practice away. In an Anthropic randomized trial, 52 engineers learned a new Python library with or without AI assistance. The AI group scored 17% lower on a mastery quiz. Participants who asked follow-up questions and requested explanations retained more. While you learn, use AI as a tutor. Ask it to explain, quiz you, and review your attempt after you finish one. 

Why shouldn’t you start with a prompt?

A prompt is the last step of using AI well, not the first. Four things go wrong when you start there.

You pick tasks by what the tool does. A tool-first approach asks, “What can this do?”. A task-first approach asks, “Where does my week leak time, and can AI close the gap?”. Only the second question finds work worth automating.

The model fills gaps with the average answer. Without a defined reader, goal, and source material, AI produces text that fits everyone and helps no one. You then spend the saved time editing it into shape.

Tools multiply. Each new tool adds an interface to learn and outputs to check. A Boston Consulting Group study of 1,488 US workers found that self-reported productivity rose with up to three AI tools and fell at four or more.

The review lands at the end. If you haven’t planned how to verify the output, you find the errors after the work is already in front of someone else.

What should you set up before you prompt?

Five steps separate a usable first output from a draft you have to rescue.

1. Ground the model in source material. Attach or connect the files instead of pasting excerpts or trusting the model’s memory. Tell it to answer only from those sources and to say when they don’t cover a point.
2. Show the target. Supply one or two examples of a good output, or a short rubric that defines good. Examples constrain style and format more reliably than adjectives.
3. Fix the output shape. Specify the format, such as a table, a schema, or a section order, so results are comparable across runs and quick to check.
4. Match the model and mode to the task. Use a reasoning-capable model for analysis and multi-step logic, a faster one for formatting and extraction, and decide whether search or tools should be on. Reuse tools you know before adding a new one, and stay under four.
5. Decide how you will verify, and set limits for agents. Choose the check and its time cost before you start. For agents, define permissions, stop conditions, and which actions need your approval.

How do you check AI output without rereading everything?

Rereading every output defeats the purpose of delegating. Build a review that scales.

1. Require evidence in the output. Ask for direct quotes or page references behind each claim, then check the ones that drive a decision.
2. Run a second pass. Use a different model, or a fresh session with a checker instruction, to compare the output with the source and flag unsupported claims. Treat flags as leads, not verdicts.
3. Sample the rest. Spot-check a fixed share of routine outputs, such as one in five, and rotate which ones.
4. Track your error rate by task type. A short log of what you corrected shows which tasks to automate further and which to move to the Keep bucket.
5. Read it as the recipient before anything goes external. If you can’t defend a line, cut it.

For multi-step automations, place a checkpoint before any step that sends, shares, or deletes. Apply one rule to the whole process: If the check takes longer than doing the task yourself, move that task to the Keep bucket.

How do you use AI at work safely and openly?

Hidden AI use skips review and puts company data at risk. Three habits cover most of the exposure: 

• Map your tools to your data. Ask IT which tools and plans are approved and what data each tier may handle. Consumer and business versions of the same product often follow different data terms.
• Know what your connectors can reach. Agents and integrations inherit your permissions, so a connected assistant can read what you can read. Scope access to what the task needs.
• Disclose by stakes. Note AI involvement on client deliverables and analysis others will rely on, and say how you reviewed it.

Frequently asked questions