An AI that Reads & Replies
I get ~100 emails a day.
AI Finance Club members with questions, brand partners, webinar replies, and a lot of people who want to sell me SEO services!
For a long time I did what you probably do. Open Gmail, scroll, decide what matters, then feel bad on Sunday evening about the 40 emails still there. It wasn’t really even about time. The problem was the sorting. Because before anyone can answer an email, somebody first has to work out what the thing even is.
For a while that somebody was a person on my team. Now an AI reads every new email arriving in my Gmail, works out who sent it and what they want, and assigns it into one of 14 labels. It never asks me anything, and for some of these labels a second workflow goes further and drafts the reply.
Annabelle and Yasser (my team) used to spend their day reading and replying to e-mails. Now the AI answers where it can, and any e-mail that needs a human gets to them instead. That's where they now spend their time. On humans adding value, not admin.
I want to open the hood and show you the inside of it.
I don’t think you should build the exact same thing (most of you should not). I'm showing you because it is much easier to aim at something once you have seen what good looks like.
Then I’ll show you the small version you can build straight away, in one of the tools you already have.
n8n
|
|
n8n is an automation tool.
Basically, it is a logic based workflow that has a series of in this -> then that workflows to carry out tasks automatically.
Super simple example: IF the e-mail body contains ‘x’ word or phrase, THEN add ‘y’ label.
On top, you can add AI into your steps:
Another example: IF the e-mail sounds like someone trying to sell me SEO, THEN draft a reply saying ‘no thanks’.
Plus, all of these tools are making building workflows much easier.
On 9 September, n8n launched n8n Assistant. All you do is describe the automation you want in plain English, it plans it, builds the nodes on your canvas, asks you for the credential (login details for a tool) when it reaches a step that needs one, runs it, and when something fails it works out why and fixes it.
Other platforms like Make, Zapier and Microsoft Power Automate now have this capability to.
But before you just randomly start asking AI to build workflows for you let me teach you how to think about this.
LEGO
In our August office hours, Bridget – one of my members – asked me to explain what n8n even is.
You start with a trigger (the starting event so the workflow knows when to start), then you click building blocks together until the thing does what you want.
Four bricks make up almost every email automation, mine included.
|
|
- A trigger, so something starts it without you.
- An AI step, where the model reads and judges.
- A database lookup, where it fetches the right information before it answers.
- And a messaging action, where a human gets told what happened.
Super important to know. Automation is rule based (If this happens then do that).
An AI workflow is an automation with an AI step in the middle producing one specific output (see whether this is an SEO pitch). Whereas an agent is given a goal instead (search all of my e-mails and reply to every SEO pitch with a no).
My inbox is an AI workflow, not an agent.
This is because we need control of the rules and the logic.
My 100+ e-mail per day workflow
|
|
Brick 1: The trigger.
Gmail is checked every 60 seconds, and only for unread mail.
Nowhere in the workflow is there a step where a human opens the inbox to start things off. It is triggered by a new email arriving.
- Tip you can use anywhere: The first question of any automation is never "what should AI do here", it is "what tells it to start". If the answer is "me, when I remember" this is not a good enough starting point.
- Tool you can use this with: Gmail filters and Outlook rules do this today for free. A step up, ChatGPT Work can start an event-triggered task when a new Gmail message arrives from a given sender or with a given subject (more on this at the end).
Brick 2, the AI step.
We have a 9,300 characters long prompt and it bans keyword matching.
The labeler is not allowed to sort by looking just for words. It has to weigh who sent it, what they want, and how they relate to my business, then pick exactly one label out of 14.
Most of the prompt it is a list of the mistakes it kept making during earlier tests. For example, A paid sponsorship is very different to an affiliate offer. A meeting invite is never a notification etc.
- Tip you can use anywhere: Your prompt improves more from the errors you name than from the instructions you add. Every time it gets something wrong, write the correction into the prompt rather than fixing the output manually.
- Tool you can use this with: Any AI tool. Paste ten emails into ChatGPT, Claude or Copilot, give it your categories, and see where it goes wrong, and keep going until it becomes super accurate.
Brick 3: The database lookup.
The reply is written from our knowledge base, not from the model's imagination.
Once an email is tagged Customer Follow-up, a second workflow picks it up, goes to a database of our standard answers, and drafts from there.
Super important, AI only drafts replies for questions we know the answers to. Plus communication that needs more attention will be given to a human.
- Tip you can use anywhere: Decide up front which words AI is never allowed to generate. Anything legal, anything about price, anything with your name promising something.
- Tool you can use this with; Your "database" can be one Google Sheet or one Word document of your 20 most common answers, attached to a Claude project or a ChatGPT project. It does not need to be a database.
Brick 4, the messaging action.
When there’s an error it creates a Slack notification.
Any failure anywhere sends a Slack message with the workflow name, a link to the failed run, and the exact error text. And the moment someone asks to speak to a human, the workflow stops and posts in Slack for somebody to pick up.
- Tip you can use anywhere: An automation with no visibility of mistakes is a big risk. Decide who gets told when something doesn’t work, and be super clear the point when AI stops and a human gets involved.
- Tool you can use this with: Slack or Teams if you have them. A message to yourself works too. Even a Gmail label called "AI stopped here" that you check once a day counts.
Your starter LEGO set
|
|
The advice I give is this.
Choose your level of automation based on how much of a problem the workflow is. Volume and time are a good way to look at this.
Using e-mails as the example.
Level 1, around 10 a day. Stay in the chat and build nothing.
Connect Gmail to Claude or ChatGPT, paste your morning's emails, ask for a sorted table. If it is ten items a day, you don’t need an automation.
Level 2, the same thing but you are bored of retyping it.
Write the steps up once as a skill and put it on a Claude or ChatGPT schedule. Still no need for automation.
Level 3, dozens to 100 per day.
Now you need a real builder, so n8n, Zapier, Make, or Microsoft Power Automate.
When a member asked me about hundreds of invoices landing in a shared mailbox, this is the level I sent him to.
Level 4, 100+ per day
That is where n8n is the right tool as it handles complexity well, and where my inbox is.
Pro tip – Don't only think about your personal inbox. Think about your whole company, or one department's main mailbox that causes a lot of headache. The same rules apply here.
7am inbox brief in about 20 minutes
If you want to give this a try, you can do this is ChatGPT and Claude. It works more agentically than an IF this THEN that workflow, so be careful of variance in the output.
In Claude Cowork:
|
|
- Go to Customize, then Connectors, add Gmail and sign in.
- Open a Cowork task and test it manually first. Give it this morning's emails and ask for a table with sender, what they want, how urgent, and a suggested first line for your answer.
- Run it three or four mornings and correct it each time. The corrections are the most important bit.
- Click Scheduled in the left sidebar, then New task, then Create with Claude, and describe what you just built. Or set it up manually if you prefer.
- Scheduled tasks run remotely, so it still runs when your laptop is shut.
In ChatGPT Work:
|
|
- Go to Settings, then Apps, and connect your Gmail account.
- Test the same prompt in a normal chat until the output is right.
- Open Work and describe the event and what you want done about it.
- Check the Trigger, Condition and Prompt it proposes back to you, then authorize it.
- Manage it under Scheduled.
Both of them will draft. Neither of them should send. I once interviewed someone who ran my whole test through ChatGPT and copy-pasted the result without checking a thing. It looked smart, which is what makes it dangerous. AI prepares the work, you still decide what is sent in your name.
The One Thing to Remember
You don’t need a super complicated workflow with 62 nodes (or workflow steps)
A trigger, an AI step, a lookup and a messaging action are the same four LEGO bricks whether you build them in n8n or in an AI tool.
All I did was move a sorting job off a human, so we could spend the time on more human task and less robotic task.
And this is the way that you should be thinking too.
Best,
Your AI Finance Expert,
– Nicolas
P.S – The best way to learn from me directly is to join my weekly masterclasses. 60-mins + Q&A where I’ll get all your AI in Finance questions answered. Join me here.
P.P.S. – The inbox brief is just the start. Here's how to build the full AI clone that runs your day → How to Build Your AI CFO Clone in Claude (2026)