I build automated systems for small businesses. So it's a little strange that the advice I give most often is "don't build this yet."
Not because automation doesn't work. In my experience, many tasks people want to automate aren't difficult so much as unwritten. Nobody agrees on whose job they are, when they should happen, or what "done" looks like.
When I ask what goes wrong today, the answer points to the starting place. If "we forget," start with a clear owner, a shared spreadsheet, and a recurring check-in. If "it takes forever," a system might help.
This week: how to tell which one you have, before you spend a month building the wrong thing.
───
🗞️ THIS WEEK IN AI
1. OpenAI launched a program aimed at small business owners
On July 21, OpenAI announced ChatGPT for Small Business, a training program focused on practical small-business work. It includes live webinars covering workflows in accounting, marketing, ecommerce, and other areas, plus in-person academies across the US and self-guided resources.
Why this matters: The useful signal here is where they start. Not with building anything, but with the ordinary work you already do. If you've been waiting for a structured on-ramp instead of another YouTube rabbit hole, this is one.
2. Claude can now learn a task by watching you do it
Anthropic added Record a Skill to Claude Cowork on July 21. You screen record yourself doing a task once, narrating what you're doing and why, and Claude turns that recording into something it can run again later. It's available on the Pro, Max, and Team plans.
Why this matters: Features like this make automation easier, which also makes it easier to automate before the process is clear. Recording a messy workflow can reproduce the same confusion faster. Write the process down first, then decide whether it's worth teaching to Claude.
3. Microsoft is letting people turn the AI off
After customer pushback, Microsoft added controls that let Teams meeting organizers switch AI features like Copilot, Facilitator, and Recap on or off during a live meeting. The rollout began in early July, driven largely by privacy concerns about Facilitator, which listens to meetings.
Why this matters: You're allowed to not use the feature. That sounds obvious, and it isn't, because most of these tools arrive switched on and quietly train you to accept them. If an AI feature isn't earning its place in your meetings, turn it off.
───
🛠️ THIS WEEK'S TRICK
Find Out If Your Problem Needs a System or Just a List
I've seen plenty of "we should automate this" ideas end in a free trial nobody opened twice. Often the task did not need intelligence. It needed a clear owner and a place everyone could see. This is how I test that before anyone spends a month building.
Step 1: Write the task down the way it actually happens
Not the way it's supposed to happen. Who touches it, in what order, and where it currently falls apart. Four or five bullets is plenty. Sometimes the real problem becomes obvious before you reach step two.
Step 2: Make the AI argue for the cheapest option
Paste this into ChatGPT or Claude:
I'm deciding whether to automate a task or just handle it in a spreadsheet.
The task: [describe what happens start to finish, the way it actually happens]
How often it happens: [e.g. 4 times a week]
How long it takes each time: [e.g. 10 minutes]
Who does it now: [me, an employee, nobody consistently]
What goes wrong today: [the real failure, e.g. it gets forgotten, numbers get retyped wrong]
What the input looks like: [a form, an email in someone's own words, a phone call, a file]
What the output has to be: [a list, an email, a decision, an entry in another system]
Recommend exactly one of these three, and name it plainly:
1. A spreadsheet or checklist, no AI involved
2. A saved prompt I run by hand when I need it
3. A real automated system
Rules:
- First, identify the main failure: speed, forgetting, unclear ownership, repeated data entry, or inconsistent judgment
- Default to the cheapest option that could solve that failure
- If nobody has written the process down, say that before recommending a tool
- For each option, estimate the setup time, weekly maintenance, and likely weekly time saved
- State your assumptions and use ranges instead of pretending the numbers are exact
- Recommend one option as the best starting point
- Tell me what evidence would change your recommendation
Step 3: Test the cheapest sensible version for two weeks
Treat the recommendation as a hypothesis, not a final decision. If a spreadsheet or checklist could work, build it today, name one person responsible for it, and choose when it will be reviewed. Two weeks of real use will show you where the process still breaks and whether a larger system would actually help.
What this works best for:
Recurring admin that annoys you more than it costs you
The "we should automate this" idea that's been sitting there for months
Anything where you're about to buy a tool to fix a habit
Limitation: AI may still recommend more complexity than you need. If it picks option 3, ask it to make the strongest case for option 1 before you decide.
───
🔧 TOOL OF THE WEEK
Google Sheets (https://sheets.google.com)
Google Sheets is the baseline I use most often. A shared sheet with four columns, what needs doing, who owns it, when it is due, and whether it is done, can handle more recurring work than people expect.
For this issue it's the honest baseline. If a sheet doesn't fix it, you've learned something real about your problem, and that's worth more than another trial account.
Price: Free with any Google account. The Gemini features inside Sheets require a paid Workspace plan or a Google AI subscription, and for this particular job you don't need them.
Best for: Anyone whose recurring task fails because things get forgotten, not because they take too long.
Worth trying? Yes. Add conditional formatting so overdue items turn red, then give one person responsibility for reviewing the sheet on a set schedule. Formatting makes problems visible; ownership makes sure someone sees them.
───
💡 PROMPT OF THE WEEK
Use this right after you decide a spreadsheet is the answer, so you don't spend an hour designing one.
I need a simple tracking spreadsheet for this recurring task:
[describe the task and what you need to keep track of]
Give me:
1. The exact column headers, in order, and one sentence on what goes in each
2. Two sample rows filled in with realistic data
3. Any formula I should add, written so I can paste it straight into Google Sheets
4. One conditional formatting rule that makes problems visible at a glance
5. A simple review routine: who should check the sheet, how often, and what they
should do with overdue rows
Keep it to the fewest columns that do the job. If I asked for something I won't
actually keep updated, tell me to drop it.
Tip: that last instruction matters more than the rest. A tracking sheet nobody updates can be worse than no sheet, because people may start trusting information that is no longer current.
───
👋 THAT'S A WRAP
The question isn't whether AI could do the task. It's whether the task was ever the real problem. Often the work isn't slow. It is invisible, unowned, or inconsistent, and making that visible may be enough.
This week, take one thing you keep saying you should automate and run it through the prompt above. Use the answer as a starting point, then test the cheapest sensible option for two weeks.
Know someone who's shopping for a tool to fix a process they haven't written down yet? Forward this to them. Twenty minutes now could save them a month of building and another subscription they do not need.
See you next Friday.