Hey friends,
I was in a client's office a few weeks ago and they pulled up their CRM to show me an account. The last note said "called, no answer." It was dated eleven months earlier. Two people in that room had been working with that client the whole time.
Nobody was embarrassed, because everyone's CRM looks like that. What struck me was the explanation. Their sales lead said, quite reasonably, that people were too busy to write notes.
That isn't the reason. The reason is that we ask people to do admin at six in the evening after a full day of meetings, and then act surprised when it doesn't happen.
The Upgrade Isn’t A Better Prompt
Most people use AI by opening it and asking it something. That works, but it means the AI only knows what you remembered to tell it, and it only works while you're sitting there.
There's a feature almost nobody uses called scheduled tasks. You write the instruction once, pick how often it runs, and it happens without you. Daily, weekly, monthly.
Three of them below. All three read the same two things, your email and your calendar. Only the question changes. The whole setup takes about half an hour, and the licence is around thirty dollars a month.
How To Set It Up
Use the Claude desktop app, not the browser. Download it from claude.com/download if you haven't already, and sign in on a paid plan. Scheduled tasks work on Pro, which is the entry tier.
Then:
Open the desktop app and go to Cowork.
Connect your email, your calendar and your CRM under connectors. This is the part that takes the time, and it's the part that matters. A scheduled task can only reach the tools you've connected to your Claude account, so if your CRM isn't connected, none of this works.
Click "Scheduled" in the left sidebar.
Click "New task" in the top right, then "Set up manually."
Give it a name, paste in one of the prompts below, choose how often it runs, and save.
That's it. Repeat twice more for the other two.
The tasks run remotely, so your laptop can be shut and they still fire. Start each one on manual for the first week so you can watch what it produces before you trust it.
Give The Boring Work To The Robots
Writing CRM notes is a ROBOT. Repeated, obligatory, boring, operational. Regular readers know where I'm going. If a task repeats and needs no judgement, it should not be sitting on a person's list at the end of a long day.
Run this one daily, early evening.
Look at my calendar and my email from the past 24 hours. For each external person I met with or exchanged email with, find the matching contact in my CRM and add a note. Keep every note to these four things and nothing else: what was discussed, in no more than two lines; what was committed and by whom; what has changed since our last contact; what is still outstanding. If a contact does not exist in the CRM, list them at the end for me to add manually rather than creating anything new. Do not write a summary. Do not include internal meetings.
The cap is the important bit. Ask for a summary and you'll get four paragraphs against every contact, nobody will read a word of it, and you will have automated the production of noise. That is worse than an empty CRM, because at least an empty CRM is honest about what it doesn't know.
Six months of this, and someone picking up the account can read the whole relationship in five minutes.
The Unhappy Email Arrives Too Late
Everyone wants AI to tell them when a customer is unhappy. The trouble is that by the time somebody writes you an unhappy email, you already know. You aren't detecting a problem. You're filing a report on one you've already lost.
The signals that actually help show up earlier, and they have nothing to do with tone. Run this weekly, Monday morning.
Review my email with external clients and customers over the last 60 days. I am looking for early warning signs, not complaints. For each client, check four things. One, has their average reply time to me got noticeably slower compared to earlier in the relationship. Two, have new people been added to threads who weren't there before, particularly anyone senior. Three, has the nature of their questions shifted from asking how to do something to asking why something happened. Four, has a client who used to email regularly gone quiet. List every client showing at least one of these, say which signal fired and give me the evidence with dates. Order by how concerned I should be. Do not comment on tone or mood.
Run that and it'll surface three or four names. You'll recognise every one the moment you see it. You just hadn't put it together.
Your Marketing Team Isn’t Short Of Ideas
Ask any marketing person what they need and they'll say content. They're wrong. They can generate ideas all day.
What they cannot generate is proof. The client's actual sentence in an actual email. The number that actually moved. The date it happened. That material sits inside your business, buried in threads nobody outside the delivery team ever reads.
So don't ask for ideas. Ask for evidence. Run this monthly.
Go through my email and calendar from the last month and find the three things that happened which would be worth talking about publicly. I am looking for evidence, not topics. For each one give me: what actually happened and when; any specific number that changed, with the before and after; any sentence a client wrote that is genuinely complimentary, quoted exactly, with the date and who wrote it; the name and email of the person I would need to ask for permission before publishing anything; and one line on what is still missing before this could be used. Ignore anything internal. If there were only one or two worth having this month, tell me that rather than padding the list.
That permission line matters more than it sounds. You can't publish a client's email. But a win, the proof, and the name of who to ask is something your marketing team can act on the same day.
The Rule That Makes All Of This Safe
Here's where people brace themselves. Is it dangerous to let AI loose on your email and your customers?
Only if you let it send things. So don't.
The rule we run on everything is simple. Drafting is always permitted. Sending is never automatic.
Notice that none of the three prompts above sends anything to anybody. The CRM note writes itself, because a wrong note is embarrassing rather than expensive. Everything that touches another human comes back to you as a list or a draft. You keep every decision that matters. You just stop doing the typing.
Claude Now Signs Everything It Writes
One more thing worth knowing before you start handing work to a machine.
Since the start of August, every new Claude model has been putting an invisible watermark into what it writes, and Anthropic is working backwards to add it to the older ones. It isn't a label you can see. It isn't hidden characters you could find and strip out. It works by nudging the low-stakes word choices the model makes as it writes, thousands of small decisions between "overcast" and "grey", into a statistical pattern that can be read with a key. It applies everywhere Claude does, worldwide, and there is no opt out.
The internet did not take this well. The complaint, roughly, is that people are being marked as AI users for work they think of as theirs.
I think the panic is aimed at the wrong thing. Anthropic has been careful about what the mark actually proves, which is that Claude touched the text, not that Claude wrote it. If you write something and Claude tidies it up, nearly all the words are still yours, so there's barely anything to detect. The mark carries no identifying information either. It can't say it was you, or your company, or which conversation it came from.
More to the point for everything above: none of those three tasks produce anything you publish. CRM notes. A list of clients going quiet. A page of evidence with names attached. All of it internal, all of it for your eyes. The third prompt hands you raw material and a permission question rather than finished copy, which was already the right design and is now also the safer one.
One practical note. The detection tool isn't public yet. Anthropic says it's coming. So anyone currently selling you a Claude watermark checker is selling you nothing.
A Calm Takeaway
Thirty dollars a month is less than one hour of the admin time any one of these replaces. That's the actual sum. Not cheap software, but an hour against work that is currently eating five.
The setup is the real cost, and not because it's difficult. It's because connecting your systems forces you to look at where your information actually lives, and most businesses have never done that on purpose.
You already know which tasks in your week repeat. You've known for years. This week, pick one and give it away.
See you next week,
— Aamir
📲 Resources & Links
🎧 Listen to the Podcast Episode on: Spotify | Apple Podcasts | YouTube
📘 Book: The CEO Who Mocked AI (Until It Made Him Millions) by Aamir Qutub