Insights · Trust

Drafts, not sends.Where AI acts alone, and where it doesn't.

The question I hear most, right before an owner trusts AI with anything real: what if it sends the wrong thing?

It's the right question. My answer is a rule I build every system around. AI drafts. A person sends.

What that looks like in practice

I built an email chief of staff for a client that reads the inbox every morning, sorts it, and drafts replies in the owner's voice. It has no send function. Its permission on the mailbox is read-only, so it can't send even if someone asked it to. Messages on a never-touch list, like anything tied to active litigation, stop it cold instead of getting a draft.

The follow-up machine works the same way. After a meeting it drafts the follow-up email and the task list from the recording, then waits. Nothing leaves until someone reads it and says yes.

A follow-up waiting for approval, with sample data

The mix I build to

Roughly two thirds of any system I build is ordinary, reliable automation: the same steps, the same way, every time. About a third is AI making judgment calls: sorting, summarizing, drafting. A person sits over the top of all of it.

That ratio is deliberate. AI is very good at the ambiguous decisions and unnecessary for the rest. Using it everywhere makes a system expensive and unpredictable at the same time.

Fixed paths or free rein

A fixed workflow does the same thing every time. An agent decides for itself what to do next. Agents are the right answer for open-ended work: research, first drafts, sorting a messy pile. They're the wrong answer anywhere accuracy is contractual.

Anything touching your customers, your money or your compliance runs on a fixed path, with a person approving. That isn't caution for its own sake. It's what makes the system usable on a Monday morning.

Three questions to tell the difference

Would a mistake reach a customer? Then it drafts and waits.

Does it move money? Then it runs on a fixed path, with a person approving.

Would you have to explain it to an auditor or a regulator? Then every step leaves a record you can show.

If the answer to all three is no, let it run.

Try this today

Pick the AI tool you use most and turn off anything it does on its own that answers yes to one of those questions. Have it draft instead. You'll lose a few seconds per item and gain the confidence to hand it more.

David

Built for regulated work

Every system I buildwaits for your yes.

Advisors, CPAs and anyone with compliance on the line: tell me what you can't afford to get wrong, and I'll show you how it would run.

Advisor AI Partners

AI operating systems and custom builds for service businesses. Jacksonville, FL. Serving nationally.

© 2026 Advisor AI PartnersJacksonville, FL. Serving nationally.