Connected reading

This guide is the practical companion to the story:

• Read the Field Note: We Tried to Let AI Handle the Reminder Emails. That Was the Problem.

The mistake hiding in plain sight

A local business owner tried to automate her client reminder emails. She picked that workflow because it looked like a perfect fit: repetitive, time-consuming, easy to template.

She was right about all three.

She was wrong about what mattered.

The emails went out. One landed on a longtime client in the middle of a health crisis — and read like a threat instead of a reminder.

The tool didn't fail. The placement did.

She hadn't asked the right question before she handed the work off. She asked: what does this look like? She should have asked: what is this carrying?

Work that looks administrative can still be trust-heavy.

That's the sorting mistake — and it's the most common way AI goes wrong in a real business.

The trap

The trap is not using AI. The trap is a sorting method that breaks down in a very specific way.

Most people scan their workflows for two things: what takes the most time, and what feels most administrative. Then they hand those off.

That method surfaces real AI candidates. It also misses something important.

Some work looks routine but isn't. The surface — repetitive, templatable, low-effort — does not always match what's underneath. Some of the fastest tasks in your business are also the most trust-sensitive. Some of the most administrative-looking communications are carrying a relationship that took years to build.

When you sort by appearance, you miss what the work is actually carrying.

That's when things go wrong.

Why it feels reasonable

The questions most people ask when thinking about AI are honest ones:

What's repetitive? What takes too long? What could be templated?

Those are good questions. They just describe the surface — how work looks, how long it takes — not what it's doing inside a relationship.

A five-minute reminder email and a five-minute urgent client update take the same amount of time. They are not the same kind of task. Time is not a reliable guide to what's at stake.

Three ways people sort work wrong

1. Sorting by repetition

If something happens often, it feels like a strong AI candidate. Sometimes it is.

But repetition tells you about frequency — not weight.

A daily client check-in is repetitive. It's also the thing that keeps trust alive. Repetition and low-stakes are not the same thing.

2. Sorting by relief

The tasks you most want off your plate are often the first to get handed off.

That instinct is understandable. But the work you find heaviest is sometimes heavy for a reason. It requires judgment, tact, or personal knowledge that isn't in the task description.

The weight is a signal. Handing off the weight doesn't make the underlying need disappear.

3. Sorting by the spreadsheet

A list of names, due dates, and payment statuses looks like data.

It's also a list of people and their circumstances.

Spreadsheets strip the context that real businesses run on — the difficult month, the health scare, the long relationship, the history that changes how you'd handle this one differently. When you let the spreadsheet define the task, you lose the information that would have told you to slow down.

Sort work by what it carries

The better question before AI touches any workflow is not what does this look like? It's what is this carrying?

Every workflow carries some mix of the following:

• Trust — Does handling this well require the client to believe you know them?

• Fear — Could the wrong phrasing create alarm, confusion, or anxiety?

• Memory — Does the right response depend on knowing something about this specific person?

• Judgment — Does this require a call that rules alone can't make?

• Reputational exposure — If this lands badly, does your name take the hit?

• Intimacy — Is this a relationship communication, not just a transactional one?

Work carrying a lot of these belongs with you. Work carrying very little can move further from your direct involvement.

This is not a complicated framework. It's a different set of questions asked before the work gets assigned.

The three zones

Once you know what a workflow is carrying, you can place it accurately.

Human-Owned

You write it or you send it. Judgment matters here, trust is fragile, or your name is fully on the line. Longtime clients. Sensitive situations. Anything where one wrong sentence causes real damage.

AI-Assisted

AI drafts, you approve. AI handles the mechanical work — templates, sorting, first drafts — and your review catches what needs catching. You're not handing off judgment. You're protecting your time while keeping your eyes on the output.

AI-Automated

AI runs it without your review on each instance. Reserved for low-context, low-risk, opt-in work with approved language. Newer clients. Standard confirmations. Situations where the cost of a slightly imperfect sentence is genuinely small.

One line worth remembering: a lot of AI mistakes happen when work that belongs in AI-Assisted gets moved to AI-Automated before the trust in the output has been earned.

Five fast questions

Before AI touches any workflow, run it through these. Sixty seconds. Catches most placement mistakes before they happen.

• If this lands wrong, do I get blamed? If the answer is you — by name, by reputation — it needs your involvement.

• Could one wrong sentence scare, confuse, or offend someone? If yes, the workflow is carrying more than it looks like.

• Does this require memory, context, or judgment AI doesn't have? If the right response depends on knowing this specific person, AI is working without the information it needs.

• Would I say this out loud to a client I've known for years? If it would feel cold or wrong spoken aloud, it will feel cold or wrong in writing.

• If it goes out under my name, do I want my eyes on it first? Your name is on it. Does it get your review?

If any of these give you pause, the workflow belongs in Human-Owned or AI-Assisted — not AI-Automated.

The Caldr method

You bring what's real. AI helps clarify and stress-test. You decide and act.

Applied to workflow placement: you know your clients, your relationships, and what's actually at stake. AI can draft, sort, template, and structure. But the placement decisions — what AI touches, what it doesn't, what gets your review — stay with you.

AI should support the work. Not blur the line between what is yours and what isn't.

Put This to Work Now (15 minutes)

Pick one workflow you're already using AI for — or thinking about using AI for.

Step 1 — List the steps

Write down each step from trigger to outcome. Not just "send the email" — the full sequence.

Step 2 — Mark what each step is carrying

Next to each step, write one or two words: trust, fear, memory, judgment, reputation, intimacy, or none. Some steps will carry nothing significant. Some will surprise you.

Step 3 — Sort each step

Label each one: Human-Owned, AI-Assisted, or AI-Automated. Sort by what it's carrying — not how long it takes.

Step 4 — Ask ChatGPT what you missed

Paste your workflow and labels into ChatGPT and say:

"Here is how I've sorted this workflow. What am I missing? What are the ways this could go wrong? Are there steps I've marked AI-Automated that should be AI-Assisted or Human-Owned?"

Use AI to pressure-test the placement — not to make the placement for you.

Over the shoulder

A financial advisor sends quarterly update emails to every client. She's been using AI to help draft template language from portfolio categories — conservative allocation, growth-oriented, recently rebalanced. It's saving her real time, and she's not pasting anything sensitive into a general tool.

The problem she hasn't noticed: she's been treating every segment the same way.

On the spreadsheet, it all looks identical — quarterly update, every client, same cadence. But some of those clients are not in the same situation they were three months ago.

Newly retired clients. Clients who called last quarter feeling anxious about the market. Clients in the middle of a major financial transition. For those segments, a polished AI-drafted update isn't wrong because it's inaccurate. It's wrong because it doesn't account for the client's context — and the right communication for them requires her judgment, not a template.

She runs the five questions against her "newly retired" segment.

If this lands wrong, who gets blamed? Her. Does it require context AI doesn't have? Yes — the relationship history, the transition, the reason this update may land differently right now. Do I want my eyes on it before it goes out? Absolutely.

She moves that segment to Human-Owned. For the rest, she keeps using AI-drafted templates with her review.

She asks ChatGPT one question — using de-identified segment descriptions, no client details:

"Here are the client segments I'm sending quarterly updates to. Which of these, based on life stage or financial transition, are most likely to require a more personalized communication rather than a standard template?"

ChatGPT flags two more segments she hadn't considered. She agrees with both.

The workflow barely changed. The placement got smarter.

A quick definition

Trust-heavy work is any task where the relationship, the context, or the stakes make a wrong move costly — and where the cost lands on a real person, not just a process.

It doesn't have to be dramatic. A simple reminder can be trust-heavy if the person receiving it is in a vulnerable moment and your name is on it.

The test is not how the task looks. It's what the task is carrying.

Closing

AI belongs in the business. It just doesn't belong everywhere.

The operators who use AI well aren't the ones who automate the most. They're the ones who sort the most carefully. They know what their work is carrying before they decide who — or what — handles it.

A good system protects judgment before it chases speed.

The companion tool for this piece helps you sort one real workflow into the right zones — and build your own version of the boundary map.

A few questions worth answering

What if I don't know where to draw the line?

Start with the five questions. Run your most sensitive workflow through them first. You don't need to sort everything at once. One workflow, sorted carefully, will teach you more than a full audit done quickly.

What if my business doesn't feel very relationship-heavy?

Every business that sends communications under a real person's name has relationships — even if they feel transactional. The question isn't whether you have them. It's which workflows touch them.

What if AI already caused a problem?

Don't swear off the tool. Debrief the placement. Ask: where did judgment go missing? What was the workflow carrying that AI wasn't equipped to handle? Redraw the line and try again with better placement. One mistake is data.

What if I'm the only one in the business?

Then the stakes for misplacement are higher, not lower. There's no one to catch the error before it goes out. The five questions matter even more when you're the only set of eyes on the work.

Read the story behind this guide

• Go to the Field Note: We Tried to Let AI Handle the Reminder Emails. That Was the Problem.