How to find your first AI automation: start with your week

Pick your first AI automation from your own week. Log your tasks, run each one through a five-step ladder, and leave with one thing worth building.

Series: Where to start with AI: from your week to your org (3 articles)

1
How to find your first AI automation: start with your week This article
2
How to choose your team's first agentic AI workload
3
Taking your first AI agent from pilot to production

Erbe, Brandon

Enterprise Solutions Architect

Published Nov 2, 2026

Last modified Nov 2, 2026

Your best first AI automation is already on your calendar. Log one week of your own work, then run each task through five questions that go from "stop doing it" to "hand it to an agent." Most tasks drop out before they reach AI. The few that are left are worth building, and you'll have picked one by the end of this article.
What you need: a week of normal work, a place to take notes, and access to the tools you already use (email, calendar, CRM). No AWS account is needed until you build.
Why I start with my own workI'm a solutions architect at AWS, and the question I hear most from customers is "where do we start with AI?" I used to answer with a catalog of use cases. Now I tell people to start with their own week, because that's where I started.
Automating your own work has advantages a team project doesn't. You're the user, so you know what "good" looks like. You already have access to the data. You find out within a day whether it helped. If it goes wrong, the damage stays in your own inbox.
Step 1: Log one weekFor five working days, write down every task you do more than once. For each one, capture:
How often it happens (daily, weekly, monthly)
Roughly how long it takes you
Where the inputs come from (email, calendar, a CRM, a spreadsheet, a dashboard)
What you produce and who sees it
How bad a mistake would be, and whether you'd notice before anyone else
Include the recurring rituals, not only the meetings. When I did this, the list that mattered wasn't my meetings. It was the work around them:
A Friday check of which customer meetings I'd had against what I'd actually recorded in our CRM
A weekly write-up of wins, blockers, and observations for leadership
A Monday review of what changed across the customer accounts I support
Keeping track of who's who at each customer and which project they're involved in
Daily practice questions while studying for a certification
A monthly look at my own growth against what my role expects
Every one of these is repetitive, pulls from systems I can already reach, and produces something I review before anyone else sees it. That combination is what you're looking for.
Step 2: Run each task down the ladderTake each task and ask these five questions in order. Stop at the first one that fits.
1. Eliminate. If this task stopped tomorrow, who would notice in a month? If the answer is nobody, stop doing it. Automating a task nobody needs is the most common way to waste a first project.
2. Simplify. Would a fixed template, a checklist, or a better format make it fast enough? Before I automated my customer contact tracking, I settled on one structure: a section per customer, a row per person, fixed columns, and two outputs (an editable CSV and a shareable PDF). Deciding the format did a lot of the work before any automation existed.
3. Automate without AI. Are the inputs structured and is the logic fixed? Then a script or a scheduled query is the right tool. Spotting changes in usage or spend across accounts means comparing this week's numbers with last week's. A model adds nothing to subtraction. I bring a model in only at the end, to write a short summary of what changed and why it might matter.
4. Assist. Is the output a draft that a person will judge before it's used? This is where generative AI earns its place. Each Friday an assistant reads my week and proposes the biggest win, blocker, and observation for my leadership write-up. I edit and approve every one. For certification study, a model writes ten practice questions a day with explanations. I check any answer I doubt against the official exam guide and the service docs, because generated questions are sometimes confidently wrong.
5. Delegate. Does the task span several systems and several steps, and can it run without you, with a checkpoint before anything changes? Then it's a job for an agent. My Friday reconciliation is the clearest example. An agent reads my week's calendar and email, pulls the activities I've already recorded in the CRM, compares the two, and writes a gap report to a working folder. It messages me when the report is ready. It reads from three systems and writes to none of them.
Step 3: Score what's leftWhatever reached rung 4 or 5 is your shortlist. Score each item with four questions:
Is it frequent? Weekly or more often pays back faster than monthly.
Is it mostly rules, or mostly judgment? Start with tasks where the rules are clear and you supply the judgment at the end.
Can you reach the inputs today? If you need to request access to three systems first, pick something else for now.
Would you catch a mistake before it mattered? Start where errors are cheap and visible to you first.
Pick the one that scores best. Build only that one.
What the build looks likeAssist and delegate jobs share the same few parts:
A trigger, usually a schedule (4 p.m. Friday, 10 a.m. Monday).
A model that reads the inputs and drafts the output.
Tool access to the systems involved, scoped to read-only wherever possible.
A place for the output, such as a working folder, so every run leaves a record you can review later.
A notification so you don't have to remember to check.
A human checkpoint before anything is sent, filed, or written back.
On AWS,  Amazon Bedrock  gives you a choice of foundation models. When an agent needs to run on its own schedule and connect to several tools,  Amazon Bedrock AgentCore  handles the runtime, tool connections, memory, and observability, and it works with open-source agent frameworks. The  Model Context Protocol  has become a common way to expose tools such as email, calendar, or a CRM to an agent. If your inputs include personal information,  Amazon Bedrock Guardrails  can detect and mask it in prompts and responses. If you'd rather not build at all,  Amazon Quick  offers research and automation features aimed at business users.
You don't need all of this on day one. My first versions ran from my laptop on a schedule. I'd move to a managed runtime once a job matters to other people.
What I learned the hard wayStart read-only. Every one of my automations started by reading systems, never writing to them. The ones that write to an official record, like a CRM, still stop at a draft that I approve. I'll keep that final step manual until I've reviewed weeks of output and trust it.
Design for reruns. Jobs fail, get rerun, and run twice by accident. My contact tracker only appends new people and removes duplicates by email address. It never rebuilds a file I may have edited by hand.
Personal data is still personal data. Pulling names and job titles out of email and meeting invites is useful, and it's also sensitive. Keep that kind of output local, limit it to people you actually work with, and think before you share it.
Some tasks should stay with you. I let an agent gather evidence for my monthly self-review, but the conclusions about my own growth are mine. Anything that goes to a customer gets my own read before it leaves.
Measure in your own minutes. Note how long the task took before, and check again after a few weeks. If the automation isn't saving you time, delete it. That's rung 1 applied to your own automation.
Where to start this weekLog every repeated task for five working days
Note frequency, time, inputs, outputs, and the cost of a mistake for each
Run each task down the ladder: eliminate, simplify, automate, assist, delegate
Score whatever reaches assist or delegate with the four questions
Build one, read-only, with output to a folder and a notification
Review its output every time for the first month
Check your minutes saved after four weeks, and keep it or delete it
Next in this seriesNext, I'll take the same ladder from your own week to your team's: how to choose a first agentic workload that several people depend on, and the patterns that make poor first picks.
For more on this topic, see the  Agentic AI topic page  and the  Productivity topic page  here on Builder Center.