Happy Friday,

AI isn’t slowing down—and neither should we. This newsletter is your shortcut to practical ideas and smarter strategies to grow, lead, and adapt as tech reshapes our industry.

To take it one step further, I’ve launched a new YouTube series. Each month, I’m sharing quick, tactical videos showing how advisors can use tools like ChatGPT to save time, think better, and scale more efficiently—starting with the work you're already doing.

Watch my latest episode here.

The AI Edge

Helping advisors scale from 1 person to many people.

AI Application:

Claude Connectors (Gmail)

Overview

Connectors are the plumbing between Claude and the software you already run your business on. You find them under Customize in the desktop or web app, and there's a directory of them covering email, calendars, file storage, CRMs, and a growing list of partner tools. The Gmail connector is the one I want to talk about, because email is where most of our unfinished work actually lives.

What makes this different from a typical integration is that it isn't a single on/off switch. When you open the connector, you get a permissions screen listing every individual action, and each one can be set to always allow, needs approval, or blocked. Reading messages and threads can run freely. Drafting, labeling, and deleting sit behind an approval gate by default, and you can block them permanently if you'd rather they never happen at all.

That granularity is the whole story. You are not handing Claude your inbox. You are deciding, action by action, what it's allowed to touch. And if you'd rather not connect your production mailbox at all, you can point it at a dedicated account instead. Same idea works in Microsoft.

How to apply Claude Connectors to the business today:

• Set up a triage label before you set up anything else. Create a label (or an Outlook folder) called Research, and drag only the emails you want Claude working on into it. Everything else in the inbox stays untouched because you're never querying it.

• Run the permissions screen first, not last. Leave read access on, leave draft and label creation on "needs approval," and block delete entirely. Five minutes here is what makes the rest of this defensible.

• Ask for rationale, not just a reply. The prompt that worked was: read these, research them, tell me what they're actually asking, back it with data, then draft a response in a conversational tone. The reasoning is the part you review. The draft is the part you edit.

• Consider a dedicated AI email account. Forward research questions into an account that exists only for this purpose. It keeps client PII out of the connected environment while you're still figuring out what your policy should be.

• Let it write to drafts, once. Ask Claude to put the reply in your drafts folder and approve that single action. Now you've got the response sitting in Gmail where you'd expect it, and the reasoning sitting in Claude where you can go back and check it.

Advanced

• Move from chat to Cowork once you've validated the output. After a manual run proves the prompt works, the same instruction can go into Claude Cowork: review whatever lands in Research each day, provide rationale for each, draft the shorter version of the response, and put it in drafts.

• Use a Project as the container. Set up a project called Emails, put your standing instructions in the project (how you want responses framed, what sources you want cited, how long the reply should be), and have each email spawn its own thread. You end up with a browsable record instead of one endless conversation.

• Schedule the run. An hourly or daily job checking the Research label turns this from a tool you remember to open into a process that's already done by the time you sit down.

• Encode your house style in the instructions. The first output came back thorough but long. Telling the system upfront that you want bullets, a source link for every claim, and a two-paragraph client-facing version fixes that permanently rather than one email at a time.

• Pair the rationale thread with your review workflow. When a draft looks off, you don't guess at why. You open the thread, read what it based the answer on, and either correct the instruction or correct the assumption. That's how the thing gets better instead of staying the same.

What could this application mean for the future of our business?

The interesting part isn't speed. Running that first query manually took about as long as answering the emails would have. I say so out loud in the demo, and it's worth repeating: the manual run saved zero time. What it did was prove the process worked, which is the prerequisite for automating it.

I think that's the pattern we're going to keep seeing. The value isn't in the single query. It's in doing something by hand until you trust the output, then handing the repeatable version to something that runs while you're in a client meeting.

The competitive question this raises is uncomfortable. Two of us get the same email from a client asking about Social Security at 62. One of us answers the question that was asked. The other gets a draft that catches what the CPA missed, flags a survivor benefit nobody raised, and cites the source. Neither of us worked harder. One of us just built the workflow first.

Client expectations move accordingly. Not toward faster replies, but toward replies that are visibly more complete than the ones they're getting elsewhere.

Every draft still needs your review before it goes out, that doesn’t change. And the human in the loop is more important than ever.

See Claude Connectors in action:

Two client emails were sitting in the Research label in my Gmail. One from a woman planning to claim Social Security at 62, with a CPA thread already attached. One from a client working through a divorce, weighing whether to keep the house or sell it.

Claude read both, research both and drafted multiple versions of responses.

The beauty was that Claude showed it’s work, tied it’s insights to source data and then gave insights that were deeper than one may have given just with a quick response.

Seeing how to set this up is powerful. But it is not the end state, it’s the beginning state to get to the end state of automation with a human in the loop.

We talk through the initial test and then discuss ways to automate this going forward.

Watch the full demo to see the permissions screen I set before any of this ran, the one line in the output I'd edit before sending, and how the whole thing gets handed to Cowork so it runs without me.

-Matt