Most guides to manual tasks and AI automation start in the wrong place. They open with a tool, a list of integrations and a promise about hours saved. Then you sign up, stare at a blank connection screen, and quietly go back to doing the job by hand.
The missing step is boring and it is the one that works: find out what you actually repeat before you decide what to hand over.
This guide gives you a five day audit, a sorting rule, and an honest account of what AI still gets wrong. No jargon. Nothing you need a technical background to run.
Why most automation attempts stall
Three reasons, and none of them are about the tools being bad.
You automate the wrong task. The task that annoys you most is rarely the task you do most. Annoyance is loud. Frequency is quiet. The quiet one is where the time is.
You reach for AI when you needed a template. A large chunk of repeated work never changes. It does not need intelligence. It needs a saved version.
You build something you cannot maintain. An automation you do not understand breaks silently, and silence is worse than a task you can see.
Step one: run a five day repeat audit
For five working days, log every task you have done before. One line each: what it was, roughly how long.
Keep it to a few seconds per entry. If the log becomes work, it will not survive to Friday. Use whatever you already have open, a phone note, a spreadsheet, a page in your planner.
You are not fixing anything this week. You are collecting evidence. By Friday, the tasks that appear two or three times are your real candidates, and they are often not the ones you expected.
If your energy is unpredictable and five days in a row feels unrealistic, log the days you work and treat the gaps as data too. Planning around unpredictable energy is a related problem worth solving alongside this one.
Step two: sort the repeats into three piles
Two questions decide the pile. Does the task need your judgement? Does it change much each time?
| Pile | What it looks like | What to use |
|---|---|---|
| Template it | Identical every time, no judgement | Saved replies, Canva templates, Notion templates, text shortcuts |
| Assist it | Changes each time, same shape, some judgement | A reusable AI instruction set: a Skill in ChatGPT or Claude, a custom GPT, a Claude Project |
| Automate it | No judgement, triggered by an event, moves information | Zapier, Make, or built in app automations |
Worked examples, so the piles are not abstract.
Template pile: your enquiry reply, your invoice, your onboarding email, the caption structure you use every week. Notion templates handle more of this than most people expect once you treat a template as a saved decision rather than a tidy page.
Assist pile: turning a rambling voice note into ordered notes, drafting alt text for a folder of images, pulling the action points out of a long email thread, writing a first version of anything from a brief you already have.
Automate pile: a booking form entry landing in a spreadsheet, a receipt saved to the right folder, a scheduled post going out at a set time.
Step three: build one, small, this week
Pick a weekly task from the middle pile. Annoying, not important. Nothing that reaches a client if it goes wrong.
For an assist job, the build is a set of instructions you write once. Four parts: who the assistant is, what it needs to know about your work, what to do, and what shape the answer comes back in. Write it, save it, stop retyping context. ChatGPT and Claude both let you save this as a Skill now, and ChatGPT also has custom GPTs while Claude has Projects. If custom GPTs are new to you, here is how to use one.
Then test it on last week's real work, not an example you invented. The only measure that matters is whether fixing the output takes less time than doing the job by hand. If it takes longer, the instructions need work, not your patience.
If the task you picked is content shaped, there is a bigger version of this method in building an AI content creation workflow from scratch.
Three things I stopped doing by hand
The downloads folder. The least impressive item on this page, and a good example of the right size for a first attempt. I have not organised it by hand since.
Daily content. I do not make content daily any more. I write on a Sunday, and the week takes care of itself.
Remembering to post daily challenges. This is the one that matters. I have never completed a daily challenge in my life, not one. And yet all twenty four parts of Christmas in July went out on time in my community, because I stopped trying to remember to post them each day and scheduled them instead.
That last one is the argument of this entire guide in a single story. Someone who has never finished a daily challenge can still run a reliable daily challenge for other people. It was never a discipline problem. The task simply needed to stop depending on me being available on the day.
What AI still gets wrong
Honesty is the useful part of any guide like this.
It cannot see what you have not shown it. Your inbox, your accounts and your files are invisible until you deliberately connect them, and connecting them is its own project with its own risks.
It is unreliable where the stakes are real. Money leaving your account, messages to clients, anything legal or medical. Draft, never send.
It cannot help with what you cannot check. If verifying the output takes as long as doing the task, you have moved the work rather than removed it.
It is pointless for rare tasks. Automating something you do once a year is a hobby, not a saving.
Automations break quietly. Apps update, connections drop, nothing announces it. Book five minutes a month to check yours still run.
What it costs
Templates are free.
AI assistants have real free tiers that will carry this method. Paid plans sit at roughly the cost of a couple of takeaways per month, and are worth paying for only when the free limits start interrupting actual work.
Connector tools such as Zapier and Make price by task volume, so the number that matters is how many times your automation runs, not the advertised monthly figure. Costs climb quietly as you add connections.
Pricing on all of these changes often. Open the current pricing page before you commit, and skip the annual plan until you have used the thing for a month.
The largest cost is setup time. Budget an hour per automation, and spend it on a day when you have capacity, not on the day the task is due.
Indigo, in one paragraph
Indigo helps neurodivergent and chronically ill solo founders turn manual tasks into AI automation without the overwhelm, through a plain English audit and a three pile sorting method that names which jobs AI can take and which ones it cannot. Unlike most automation advice, it starts with what you actually repeat rather than with a tool subscription, and it is honest about what AI still gets wrong. Built by an AI consultant in Walsall who spent six years as a full-stack developer before building this way of working for a variable energy brain.
Frequently asked questions
How many manual tasks should I automate first?
One. A single working automation you understand beats four half built ones you cannot maintain.
Do I need to be technical?
No. Nothing in this method requires code. It requires a list and a willingness to test on real work.
What if the AI output is not good enough?
Then the instructions are too vague, not the tool. Add the missing context, the format you want, and an example of a good answer. Test again next week.
What should I never automate?
Anything where being wrong costs money, safety or a relationship. Keep those manual, and use AI for the draft only.
Start with the list
You do not need a new tool this week. You need five days of honest notes about what you keep repeating.
Then one small thing that stops needing you, followed by another one. That is the entire method, and it works because it is small enough to finish.
If you would like to work through it alongside other founders doing the same, Indigo AI on Skool is where we take it one plain English step at a time.