AI Task Automation

Automate recurring tasks with scheduled AI agents that run without you.

By James Schramko · Updated May 2026

How to automate recurring tasks using AI agents that run on a schedule without you.

This playbook focuses on Claude Code routines as the primary automation layer. Routines are scheduled AI agents that connect to your tools, run on a cron schedule, and deliver results to wherever you specify.

When To Use

You have a task you do weekly or monthly that follows the same steps each time. Pulling reports. Summarising content. Checking platforms. Sending reminders. Extracting data. Auditing files.

If you can describe the steps, a routine can probably do it.

The Setup

Before you automate anything, connect your tools.

Common connections via MCP (Model Context Protocol) servers:

Slack for notifications and delivery
Google Drive for file access and reads
Gmail for email scanning and drafts
Zoom for meeting recordings and transcripts
GitHub for code and content repos
A browser tool for platforms that need a logged-in session

Each MCP connects once and stays connected. Set them up at claude.ai/connectors before building your first routine.

The Process

Step 1: Do the task manually first

Run through the task yourself and note each step. What do you open? What do you look for? What do you do with what you find? Where does the output go?

Do not automate something you have not done manually. You need to know what good looks like before you can check if the automation got it right.

Step 2: Create the routine

Open claude.ai/code/routines and create a new routine. Give it a clear name. Write the instructions as a series of numbered steps.

Be specific about inputs, actions, and outputs.

Example: "Every Monday at 9am: scan my Gmail for newsletters from the past week. Summarise the key points from each one. Post the summary to my Slack channel #digest."

Example: "Every Tuesday: log into Google Analytics, pull the top 10 pages by traffic for the last 7 days, format as a simple list, send to Slack."

Example: "Every Wednesday: list all files modified in the past week in this GitHub repo. Summarise the changes. Email me the digest."

The clearer your description, the better the result. Vague instructions produce vague output.

Step 3: Set the schedule

Routines use a 5-field cron expression in your local timezone.

Common patterns:

`0 9 1` = every Monday at 9:00 am
`0 1 1 ` = 1st of each month at 1:00 am
`0 0 *` = every day at midnight

The cron generator at crontab.guru helps if you need a less common pattern.

Step 4: Pick the connectors

Each routine needs access to whatever tools the task uses. Add the relevant MCP connectors when you create the routine. The routine can only use tools you explicitly grant.

Important: the routine runs with write permission to those connectors. Review what you grant.

Step 5: Review the first few runs

Do not set and forget immediately. Check the output for the first two or three runs. Make sure it is pulling the right data and formatting correctly.

Once you trust the routine, let it run.

Good Tasks To Automate

Newsletter or content summaries. Scan email or RSS feeds, summarise, deliver to Slack or email.

Weekly platform reports. Pull stats from analytics, ad platforms, or dashboards. Format and send.

Reminders and triggers. Scheduled messages, calendar prompts, or check-in nudges.

Data extraction. Pull information from documents, emails, or platforms into a structured summary.

File audits. Scan a folder, identify outdated or duplicate files, report findings.

Content scanning. Check feeds, social accounts, or news sources for relevant updates.

Repo or vault digests. Summarise what changed in a content repo over the past week.

Tasks To Keep Manual

Client communication. Strategic decisions. Content approval before publishing. Anything involving relationship signal.

Financial transactions. Public posting. Account changes. Anything with high consequences if the agent misreads.

Start with low-risk, repetitive tasks. Expand as you build trust in the system.

Routine Hygiene

Name routines clearly. "Weekly Stats Digest" beats "Untitled Routine 47."

Document each routine in one place outside the platform. A simple log of: routine name, schedule, what it does, which connectors it uses. When something breaks, you find the cause faster.

Disable routines you are not using. An inactive routine still appears in the list and clutters the view. Disable or delete it.

Audit your routines quarterly. Some will have outlived their usefulness. Others will need refining as your workflow shifts.

What This Creates

Hours back each week from tasks that ran in the background.

Consistency. The routine runs the same way every time.

Nothing forgotten. Scheduled tasks do not slip through the cracks.

A growing automation layer that compounds. Each routine you build makes the next one easier because you understand the pattern.

Common Mistakes

Automating before understanding. If you have not done the task manually, you will not know if the automation is wrong.

No review period. Trusting the output immediately without checking the first few runs.

Over-automating. Trying to remove yourself from tasks that still need human judgement.

Vague instructions. The routine does exactly what you ask. If your description is unclear, the output will be unclear.

Connector overreach. Granting more tool access than the routine needs. Keep permissions narrow.

Start Simple

Pick one task you do every week that takes 15 minutes or less. Automate that first. Get it working. Then add another.

Do not try to automate your entire workflow in one session. Build the layer over weeks, not days.