
Put the job on a schedule. Any job that happens daily or weekly with the same inputs, like sorting your email, prepping for meetings or checking your numbers, can be written down once as a set of instructions (Claude calls these skills) and set to run at a fixed time. If the job only needs apps Claude connects to, like Gmail and your calendar, it can run in Claude’s cloud with your computer off. If it needs files on your computer, it runs there. You still check the result, but you stop being the one who has to start it.
Why does your AI only work when you ask it to?
Last week, a new client of mine told me she wants a different AI bot for every job in her business. One for her newsletter, one for social media, one for a podcast, basically an AI CMO and an AI CSO, each one doing its own work.
If you’ve spent any time online lately, you’ve definitely seen a whole bunch of talk about AI agents doing work automatically. If you use AI in your own business, there’s a good chance it looks a lot less like that. You open a chat, you type what you need, it does the job, and then it sits there until you come back and ask again.
That’s fine for one-off work. But a lot of what fills an owner’s week is the same jobs on the same days with the same inputs, like the inbox every morning and the prep before every call. If AI only does those when you remember to ask, you’re still the one running them, just faster.
PYMNTS surveyed 533 small and medium-sized businesses for a report it published last month, and 72% of the ones using AI said it helps them finish tasks faster. Each of those tasks still starts with somebody opening a chat.
On November 11th, Matt Wolfe put out an episode with Peter Yang, who runs a one-person business, walking through his setup. Twice a day, at 7am and 4pm, an AI he calls his chief of staff reads his email, his calendar and his Slack, and sends him a numbered list of what needs him. He replies with what to do on each one, and it drafts the replies. It never sends anything on its own, and he never has to open a chat to start it.
So before the setup, it helps to see what a job like this looks like once it’s working.
Sources: 60% of small business AI users do more without adding staff (PYMNTS) · How To Build An AI Operating System For Your Life (Matt Wolfe, YouTube)
What does it look like when AI does a job without being asked?
When a job like this works, you’re simply receiving a notification that the work is done and it’s time for you to check it.
A very common one going around right now is the daily brief. The daily brief can do a lot of different things, and it varies person to person. It can include email triage, prep for the meetings on your calendar, a look at your business numbers and so much more. Previously, you would hire somebody to do some of those things, or you would end up doing them yourself.
All those tasks take real time. Some people take an hour or two to get through their emails. Prep before a meeting might get done in the 10 to 15 minutes before it starts, but if you have back-to-back meetings, you just can’t get that done at all. And some people don’t even track their numbers at all, and you can’t improve what you don’t track.
So something like this boosts your business, because it’s getting something done that never happened before.
I have a few jobs that run on my computer every morning, done before I sit down. Every day, it pulls in meeting transcripts from the previous day, collects my YouTube and email list numbers, checks on work that I did in my workspace, and takes all of that and builds a brief sent to my phone as a single to-do list for today.
I also have things that run less often. Every Friday morning, AI helps me look at what content to produce next by looking at outliers on different social media platforms.
When a job starts working like this, the first thing you get is relief, because that’s one job you don’t have to do anymore. Then there’s the time you used to spend doing it manually, which is now handled. And I’m able to do things I previously didn’t, because AI handles them for me, so parts of my business are actually getting done now.
But not every job should be handed over like this.
Measure yours in 3 minutes: the free Task Audit scores how much of your recurring work could run without you.
Which jobs should AI do without you asking?
The best jobs to give AI are the ones with the least input from you, that run the same way every time.
All of the previously mentioned workflows are great examples. Email triage always has to filter out spam and sort the important emails into the right categories. Weekly content research always looks over the past week for viral outliers. Checking my numbers, prepping for meetings, drafting follow-ups after calls, they’re all fairly straightforward, fairly standardized processes. Any job that happens on a daily or weekly basis, with the inputs being the same every time, can be set up to just happen at a certain time.
A job with a lot of input from you, that’s nuanced and based on your judgment, might not be one for AI to handle on repeat by itself. While we are getting closer and closer to it being solved, things like YouTube video editing for long form content still need you to put a lot of your judgment into it.
At most, you want to be doing a single check over the job once it’s done each day. If you find yourself rewriting what comes back every morning, either the job needs your judgment too often to run on a schedule, or the instructions need more detail.
Picking the job is also where you write it down. In my issue on handing a task to AI like a new hire, I went through how to write a job down so somebody else can do it, and it’s the same process here. You write down what it needs to look at, what a good result looks like and what it shouldn’t touch.
That brings me back to my client, who wants a bot for every job.
Do you need a different AI bot for every job in your business?
You can create the AI team. You can create a bot for each role, you could have your AI CMO or your AI CSO, and these are all great and interesting things to help you in your business. But AI is only as good as its instructions.
A CMO agent that’s just broadly told to be the chief marketing officer will still be kind of general in its output. You need either specific agents or specific skills to get outputs that are more in line with what you’re actually looking for. A CMO agent by itself might be more tailored towards marketing, but if you give it a dedicated SEO skill to use or agent that it’s in charge of, it will be much stronger in SEO.
This applies to basically every part of your business that you use AI for. If you ask two AIs to do the same job, the one with the better skills, or the agent with instructions dedicated to one task, will produce the better result. I went deeper into why in my issue on why AI gives generic answers.
My client has tried building one of these before. Back in the spring, she built a market research bot in Claude. It took her about a week, and it ran so much that she ended up having to upgrade her AI plan.
And this matters even more once a job runs on a schedule. When you ask AI for something in a chat and the answer comes back a little off, you fix the prompt and ask again. At 7am, nobody’s there to fix the prompt. The job does exactly what its instructions say, every time, whether those instructions are good or not.
That’s why the written skill is the part that makes a scheduled job work. You can ask Claude with a basic prompt to set up a scheduled job, but a skill or agent file that you know works every time will save you a whole bunch of troubleshooting later. If it doesn’t work when you ask for it, it won’t work at 7am either.
How do you set up AI to run a job on a schedule?
Before you schedule anything, call the skill yourself and check that it works. Run it a few times on a normal day and read what comes back. If you’d be happy getting that result every morning without changing it, it’s ready.
Then you have to decide where the job runs, in the cloud or on your computer. A job is local if it needs any files on that device. So if AI has to reference, update or change files on your computer, that’s a local job. A cloud job doesn’t need local files, so you can just ask Claude to set the skill up on a schedule and put it in the cloud.
The first way is the cloud. If you’re on any paid Claude plan, you can describe the job and the schedule in a conversation, and Claude asks a few follow-up questions and proposes a name, a schedule and the instructions for you to confirm. You can set it to run hourly, daily, on weekdays or weekly, and it can use the apps you’ve connected and the skills you’ve made. Because it runs on Claude’s side, it keeps running even when your computer is asleep or the app is closed.
A daily brief that only needs your email and your calendar is a good fit here. Give it Gmail and Calendar and nothing else, set it for weekdays at 7am, and click Run now once so you can watch it work. On that first run, check that it could actually see your email and calendar, because if it can’t, the brief comes back empty.
The second way is on your computer, for anything that lives in files. If your numbers are in spreadsheets in a folder, the cloud can’t see them, so the job has to run where the files are.
The Claude app can do this too. The setup is the same as the cloud one, with one switch to run it only on your computer and a folder for it to work in. The catch is that it only runs while the Claude app is open and your computer is awake, so if you forget to open it one day, it won’t run. Adding Claude to your startup apps and turning on Keep computer awake in its settings fixes most of that, and if it does miss a run, it does one catch-up run the next time the app opens.
If you don’t want to depend on the app at all, your computer has its own scheduler. Windows has Task Scheduler, and on a Mac it’s called launchd. You don’t have to learn either one, because you can ask Claude Code to set it up for you. This still needs the computer on, but not the Claude app, and it’s how my own scheduled jobs run every morning.
For a five to ten person firm, the same split works for every seat. Anything that comes out of your email, calendar or the apps your team already uses can go in the cloud. Anything that comes out of the spreadsheets and folders on someone’s computer runs on that computer. Either way, the job lands with the person who used to do it, and their part becomes checking it.
Most of the problems start after setup, once the job has been running for a while.
Sources: Schedule recurring tasks in Claude Cowork (Claude Help Center) · Desktop scheduled tasks (Claude Code docs)
What can go wrong when AI runs a job on its own?
My own daily brief failed six mornings in a row at the end of October. The account paying for the AI behind it had run out of credits, so every morning the job started, hit the same error and stopped. Nothing came to my phone, and it was almost a week before I noticed it wasn’t there.
That’s the biggest thing to plan for. When you ask AI for something and it fails, you see it fail. When it fails on a schedule, you could just not hear from it. So you want a check that the job still works. Every scheduled task in the Claude app keeps a history, so you can see each run, and each one it skipped and why. Or you can have one more job whose only work is checking that the others ran.
A few other things can go wrong, and the skill or agent file gets you past most of them because you already know it works. The first time you click Run now, it’ll ask permission for things like reading your email. Pick always allow, or the scheduled run will sit there waiting for you. And if your computer sleeps through 7am, the run gets skipped, and the catch-up might happen at 11 at night. So put a time rule in the instructions, like only look at today’s emails.
The most important part is guardrails in general. AI needs to know the things it shouldn’t do, and that goes in your skill or agent file. My email agent drafts replies, and the rules at the bottom of the skill say it never sends them. Peter Yang’s chief of staff works the same way. Anything that goes out to a client should still go through a person first.
Only connect the apps the job actually needs. A brief that reads your email doesn’t need access to your bank account or your files.
And watch how often it runs and which model it uses. My client’s research bot ran so often she had to upgrade plans. For a routine job, a cheaper model like Sonnet does fine.
If you want to start this week, here’s how.
Firstly, write down every job you do on a daily or weekly schedule, and roughly how long each one takes.
Next, pick the one with the least judgment in it, the one that runs the same way every time.
Then, ask Claude to turn it into a skill and run it yourself until you’d be happy getting the result every morning without changing it.
Finally, put it on a schedule, in the cloud if it only needs your apps or on your computer if it needs your files, and check the history after the first few runs to make sure it’s actually running.
On Tuesday I’m putting out a video on my YouTube channel where I set up both ways on screen, a daily brief in the cloud and a numbers check on my computer, if you want to watch it get built.
Or, if you’d rather have some help setting this up, I do a free discovery call.
Frequently asked questions
Why does your AI only work when you ask it to? Because a chat window waits for you. You type what you need, it does the job, and then it sits there until you ask again, so for the jobs that repeat every day or week, you’re still the one running them. In a PYMNTS survey of 533 small and medium-sized businesses, 72% of the ones using AI said it helps them finish tasks faster, but they’re still starting every task themselves.
What does it look like when AI does a job without being asked? You get a notification that the work is done and it’s time to check it. A common example is a daily brief that triages your email, preps you for the meetings on your calendar and looks at your business numbers, which used to take an hour or two, or a hire, or never got done at all.
Which jobs should AI do without you asking? The ones with the least input from you, that run the same way every time, like email triage, weekly content research, checking your numbers, meeting prep and follow-up drafts. Jobs that are nuanced and based on your judgment, like editing a long-form video, should stay with you. At most, you want to be doing a single check once the job is done each day.
Do you need a different AI bot for every job in your business? You can build one, but AI is only as good as its instructions, and a bot broadly told to be your CMO will still be general in its output. A dedicated skill for each job produces a better result. That matters more on a schedule, because nobody’s there at 7am to fix the prompt, so write the skill and make sure it works before you schedule it.
How do you set up AI to run a job on a schedule? First make sure the skill works when you ask for it. Then pick where it runs: if it only needs apps Claude connects to, like Gmail and your calendar, set it up as a scheduled task in Claude’s cloud, which keeps running with your computer asleep. If it needs files on your computer, run it there, either through the Claude app or through Task Scheduler on Windows or launchd on a Mac, which Claude Code can set up for you.
What can go wrong when AI runs a job on its own? It can fail quietly, so check the task history or have a job that checks the others ran. It can stall on a permission or get skipped while your computer sleeps, so click Run now once, pick always allow and put a time rule in the instructions. Keep it to drafts that a person approves, only connect the apps it needs, and use a cheaper model for routine jobs so it doesn’t burn through tokens.
Photo by ubeyonroad on Unsplash
Want this running in your business?
Book a free 30-minute AI discovery call. We'll look at where your time is going and show you exactly what an AI Operating System could give you back.
Book a free discovery call Free 30 minutes · No pitch · You'll know where AI fits either wayNot there yet? Find your Task Automation % with the free Task Audit, or see what a managed AIOS includes.