Automation & Agents

People automate the wrong things. The rule that prevents most wasted weekends: never automate a process you haven't run manually enough times to describe in writing. If you can't write down the steps, you don't understand it well enough to hand it off, to software or to a person.

Three rungs:

• Manual: you do it. Fine forever for anything rare, high-stakes or still changing shape.
• Assisted: you run it, AI does the heavy part, you approve. This is where most of the value is, and where most people stop too early.
• Automated: it runs without you, on a schedule or a trigger. Reserve this for tasks that are frequent, well-defined and low-blast-radius when they go wrong.

Test for the top rung: frequency x time saved beats setup cost, and a silent failure costs you little. A weekly report qualifies. Sending messages to prospects under your name does not. That one stays assisted, forever.

Help me triage my repeating tasks.

MY WEEK: [list everything you do repeatedly. Be specific about frequency and how long each takes]

For each, classify:
- MANUAL: keep doing it yourself, and why
- ASSISTED: AI drafts, I approve (highest value, do these first)
- AUTOMATED: can run without me

For every AUTOMATED you propose, tell me: what breaks if it fails silently, and whether that's acceptable.

Then rank them by (frequency x time saved) / setup effort and tell me the ONE to build first.

Be conservative about the automated category. Anything that sends something under my name to a real person stays assisted.

"AI agent" is the most abused term in this industry right now. Stripped of marketing: an agent is a model given a goal, a set of tools it can use, and permission to take multiple steps on its own. That's it. The difference between an assistant and an agent is that the agent acts more than once without checking in.

That's genuinely powerful and genuinely where things break. Every step is an opportunity to compound a small misunderstanding into a confidently wrong outcome, and unlike a bad paragraph, a bad action can't always be deleted.

Practically, most people's first two "agents" are workflows with an AI step in the middle, and that's the right place to start. It's more reliable, easier to debug, and delivers most of the value.

The question to ask about any agent someone's selling you: what tools does it have, and what's the worst thing it could do with them? If the vendor can't answer clearly, don't connect it to anything that matters.

I want to understand whether I need an agent or just a workflow.

THE TASK: [describe what you want handled]

Answer:
1. Stated plainly, what would an "agent" do here that a simple AI-in-the-middle workflow wouldn't?
2. What tools or access would it need?
3. What's the worst realistic outcome if it misunderstands?
4. Where should a human approve before it proceeds?
5. Honestly: is this a workflow, an agent, or a task I should keep doing myself?

Don't sell me on agents. If a workflow is the right answer, say so.

Your first build. The shape: a trigger (weekly, or a new row in a sheet), a step that pulls your source material, an AI step that drafts using your voice profile and constraints, and an output that lands somewhere you'll see it: a doc, a sheet, an email to yourself.

The critical design choice: it drafts, it does not publish. The output goes to your desk, not to your audience. Everything else about this build is negotiable; that isn't.

Where people get stuck is the input. An automation that generates content from nothing produces filler. Feed it something real: a running note of things you observed this week, questions customers asked, a voice memo transcript. The workflow's job is to turn raw material into drafts, not to invent material.

Build it in whatever automation platform you're comfortable with, or, honestly, as a saved prompt you run every Monday. The saved prompt version works, costs nothing, and most people should start there.

Design a weekly content workflow for me.

MY INPUTS: [where your raw material comes from: notes, customer questions, voice memos]
MY OUTPUT: [what you need each week, e.g. 2 posts, 3 video hooks, 1 email]
MY TOOLS: [what you actually have]

Give me:
1. The workflow, step by step, plain language
2. The exact AI prompt that sits in the middle of it
3. Where the output lands so I'll actually see it
4. The manual version I should run for two weeks first

CONSTRAINT: this workflow drafts. It never publishes. Design the approval step in, don't bolt it on.

If the honest answer is "just run a saved prompt every Monday," say that instead of designing something impressive.

The second build, and the one most likely to make you money. The shape: a new lead arrives (form, DM, event, sheet row), it's logged with a timestamp and your notes, a scheduled check finds anyone who's due, and an AI step drafts a personalized follow-up into your drafts folder, where you read it, edit it and send it.

That last part is the whole design. Automated messages sent under your name to real people is the single fastest way to destroy the trust your business runs on. The automation solves the actual problem, which is remembering, not typing.

Two details that make it work: log the notes at the moment of the conversation, not later, because the draft is only as good as what you captured. And build the exit: after the fourth touch with no reply, it moves to a quarterly list and stops.

The simplest working version is a spreadsheet with a "next touch" date, a calendar reminder and a saved prompt. Don't let tooling ambition stop you from having the system.

Design a follow-up system that stops me losing leads.

WHERE LEADS COME FROM: [sources]
MY CADENCE: [e.g. day 2, day 7, day 21, then quarterly]
MY TOOLS: [what you have]

Give me:
1. The tracking sheet: exact column headers, including what I must capture during the conversation
2. The workflow: how a lead enters, how I'm reminded, how a draft gets written
3. The AI prompt that drafts each follow-up from my notes
4. The exit rule: when someone moves to quarterly and how I close the loop politely

HARD CONSTRAINT: drafts land in my drafts folder for review. Nothing sends automatically under my name. Design around that, don't argue with it.

Start with the version I can build in an afternoon with a spreadsheet.

Automations fail quietly. A workflow that has stopped running looks exactly like a workflow with nothing to do, and people discover the breakage weeks later, usually because a customer mentions it.

Three habits prevent most of the damage:

• Checkpoint anywhere money, promises or a person's name goes out. Nothing that a customer will read as coming from you personally leaves without your eyes on it. This is not a phase you graduate from.
• Build a heartbeat. Have the workflow report even when there's nothing to do. "0 follow-ups due this week" is information. Silence isn't.
• Audit monthly. Fifteen minutes: is it still running, is the output still good, has the model or the tool changed under you? They do change.

And keep a note of what each automation does and why, in plain language. Six months from now you will not remember, and an automation you don't understand is one you can't fix or safely turn off.

Stress-test the automation I just designed.

THE AUTOMATION: [describe it]

Tell me:
1. Every way this fails SILENTLY: where I'd never notice it stopped
2. Every place it could produce something embarrassing or damaging
3. Where a human approval gate belongs, and where one would just be friction
4. What heartbeat or health signal I should build so silence isn't ambiguous
5. What I should check in a 15-minute monthly audit

Assume I'm going to forget this exists in three months. Design for that person, not for me today.