Automate your day with AI

Start with a workflow inventory

Daily work feels fragmented because the same small actions appear in different places: copying an update into a tracker, preparing a meeting brief, routing a request, checking a record, or following up after a deadline. Before automating anything, make those actions visible.

For three working days, record tasks that meet at least one of these conditions:

• The task follows the same sequence most of the time.
• The inputs already exist in a structured system.
• A missed step creates delay or rework.
• The result can be checked against a clear standard.
• The task interrupts focused work without requiring much judgment.

Do not start with the task you dislike most. Start with the task that has the clearest boundary. A bounded workflow is easier to test, observe, and reverse.

Choose one outcome, not a collection of features

Write the desired result as an operating outcome. "Use AI for meeting prep" is too broad. "Prepare a source-linked brief from the agenda, the last meeting record, and open actions by 8 a.m." is testable.

Define the baseline before building:

• How often does the workflow run?
• How long does the current process take?
• Which errors or omissions occur?
• Who performs the work and who relies on the result?
• What is the cost of a late, wrong, or missing output?

The baseline lets you compare the automated routine with the current one. Without it, a faster-looking demo can hide extra review work or new failure modes.

Map the trigger, inputs, actions, and finish state

Every reliable routine needs a clear start and finish. Map the workflow in this order:

1. Trigger: the event that starts the workflow, such as a new form response or scheduled time.
2. Inputs: the approved records, files, or messages the workflow may use.
3. Rules: the conditions that determine the next step.
4. AI assistance: the narrow task that benefits from classification, extraction, drafting, or summarization.
5. Review: the condition that requires a person to approve or correct the result.
6. Action: the update, notification, or handoff the workflow performs.
7. Finish state: the evidence that confirms completion.

This structure separates deterministic automation from model-assisted judgment. Use ordinary rules for stable conditions. Use a model only where the input varies enough to need interpretation.

Add controls before convenience

The NIST AI RMF Playbook provides suggested actions for governing, mapping, measuring, and managing AI risk. For a personal or team routine, translate that guidance into a small control set:

• Limit the workflow to the minimum data and permissions it needs.
• Keep a run log with inputs, outputs, actions, and errors.
• Route uncertain or incomplete cases to a named reviewer.
• Prevent the workflow from sending, deleting, purchasing, or approving without explicit authority.
• Preserve a manual path when the automation is unavailable.
• Add a stop control that the workflow owner can use immediately.

These controls matter even for a simple routine. A meeting brief can expose confidential notes. An automated follow-up can contact the wrong person. A status update can overwrite a more current record. The control should match the consequence.

Test with normal cases and awkward cases

Run the workflow in observation mode before letting it take action. Compare what it would do with what a person actually decides.

Include cases with missing fields, duplicate records, conflicting dates, unusual wording, and unavailable dependencies. Write an acceptance rule for each failure type. For example, a missing project owner might stop the workflow, while an ambiguous priority might create a review task.

The guidelines for human-AI interaction from Microsoft Research emphasize clear capability, understandable feedback, efficient correction, and support when a system is wrong. Apply those principles by showing the source used, making corrections easy, and recording the reviewer decision.

Review the routine on a fixed cadence

Automation is not finished when it runs successfully once. Review it after the first week, then on a cadence that matches its risk and frequency.

Check:

• completion rate and failure rate;
• time spent on review and correction;
• recurring exception categories;
• changes in source systems or permissions;
• whether the workflow still solves the original problem;
• whether the owner and recovery path are still current.

Expand only after the first routine produces a stable result. Reuse the method, not a copy of the workflow. Each new routine needs its own outcome, inputs, controls, and acceptance criteria.

A practical first routine

A good first candidate is a daily action digest. It can collect open tasks from approved systems, group them by owner and due date, flag missing ownership, and prepare a draft summary. A person reviews the digest before any follow-up is sent.

This routine is useful because its source records are visible, its result is easy to compare, and a reviewer can correct it without changing the underlying data. It teaches the operating discipline that makes larger automations reliable.