How do you inventory scheduled AI tasks across every assistant?

To inventory scheduled AI tasks, list each assistant your people use, open its scheduled or routines view, export what is there into one register, then record owner, schedule, access and last run. Repeat monthly. Each product keeps tasks in a different place, so a single search will not find them all.

What counts as a scheduled task

For this exercise, a scheduled task is any saved prompt that starts without a person pressing send. That includes time-based schedules, API triggered runs, and event triggers such as a new pull request. A one-off reminder counts too until it has fired.

Where each product keeps them

Based on vendor pages read on 19 October 2026:

Routines belong to individual accounts and are not shared with teammates, so an admin cannot list them from one place. Collection depends on people reporting what they have, or on admin tooling your vendor provides. Ask your vendor what an owner or admin can see.

Five steps

1. Declare the scope. Pick the assistants in use, including ones teams adopted without asking.
2. Ask owners to list their tasks. Give a short form: name, purpose, where it runs, schedule, systems touched.
3. Verify against the product. Open the scheduled view and compare. Anything in the product but not on the form is a finding.
4. Search repositories. Scan workflow files for schedule triggers and agent invocations, and look for committed agent configuration.
5. Record the register fields below.

Register fields

• Task name and plain language purpose
• Owner and sponsor
• Product, and cloud or local
• Trigger type and schedule
• Identity it runs as
• Connectors, repositories, folders and network access
• Permission mode or sandbox level
• Writes or sends anything? Yes or no
• Last run, last review date, next review date

The tool permission manifest post expands the access fields.

Findings to expect

Most first passes turn up three things: tasks with owners who have left, tasks holding connectors they do not use, and tasks nobody can explain. Treat each as a normal access review result, not a blame exercise. Ask whether the task is still needed. If not, delete it and confirm it is gone.

Worked example: a one hour discovery sprint

A security manager at a mid sized software company wants to know what scheduled AI tasks exist. She books one hour with four team leads and runs this sequence.

In the first ten minutes, each lead opens the scheduled or routines view in every assistant they use and reads the names out loud while she types them into a shared sheet. In the next fifteen, each lead adds the columns: owner, trigger, access, what it writes, last review. In ten more, she searches the organization's repositories for workflow files with a schedule trigger and for agent configuration files, and adds each hit. The remaining twenty minutes go to sorting. Tasks that write or send outside the company go to the top. Tasks with no owner go next. Everything else is read only and can wait a month.

The output is a single sheet with a risk ordered list, three tasks to retire on the spot, and a monthly review date. It will not be complete, and it does not need to be. It is a baseline, and each later pass finds less.

Sample register rows

The rows are examples of the shape of the register, not real tasks.

Close the blind spots

Four gaps are common.

• Personal accounts. People may use assistants under personal logins. Policy can require work tasks to live in managed accounts, and an amnesty period encourages honesty.
• Local only tasks. Desktop tasks live on a laptop and in a folder under the user's home directory. Endpoint inventory tools can list installed apps but usually not tasks. Ask owners.
• Event triggered tasks. They do not appear on a calendar. Search for GitHub triggers on routines and for Slack, Gmail and GitHub triggers on tasks in apps that offer them.
• Tasks created by agents. An assistant can create tasks in conversation. Review what your agents are permitted to schedule.

Keep it alive

An inventory goes stale in weeks. Tie updates to events people already do: task creation requires a register entry, offboarding includes a task check, and quarterly access reviews include the register. Automate collection where the vendor exposes an API or admin report, and be honest about the parts that remain manual. See shadow automation for discovery ideas that go beyond asking.

Try it yourself

Click through it at your own pace, or open it full screen.

An inventory of dependencies and an inventory of automations share one rule: you cannot assess what you have not listed. Notice what fields the demo records for each item.

Where Safeguard fits

Safeguard's asset discovery and agent identity capabilities are aimed at this kind of inventory problem. You can also start with the free scan at https://gold.safeguard.sh/scan on repositories where scheduled agents work.

Quick check

Sources

• Automate work with routines (Claude Code docs)
• Schedule recurring tasks in Claude Code Desktop
• Schedule recurring tasks in Claude Cowork (Help Center)
• Scheduled actions in Gemini Apps (Google Help)
• Scheduled tasks and automations (OpenAI Codex docs)

Last reviewed: 19 October 2026.