Claude's New Mac Agent Handled My Entire Email Workflow — Here Is What I Learned

AI
1

I just tried the new Claude Mac Agent feature (currently in research preview) and wanted to share my experience. In short, it worked flawlessly.

Here is what I asked it to do: search my email archive for a specific message, read and summarize its contents, compose a new email to two colleagues based on that summary with a specific request for data, and save the result as a draft for my review before sending. The entire workflow was handled autonomously. Claude located the archived email, extracted the key points, composed a well-structured message with the correct recipients and subject line, and saved the draft exactly as requested.

What surprised me most was how seamlessly it moved through each step without any intervention on my part. It identified the right email, pulled the relevant data points from a fairly lengthy report, and composed a professional message that I would have been comfortable sending with only minor edits.

I even had Claude write the prompt itself. I succinctly described what I needed in general terms and it produced a clear, well-structured prompt that I was able to use immediately. That alone saved considerable time, as crafting a detailed multi-step instruction from scratch would have been an additional task in itself.

The prompt below has been redacted to protect confidentiality:

Open Gmail and search my Archive folder for an email titled “[Email Subject Line].” Once you have located it, compose a new email addressed to [Recipient 1] and [Recipient 2]. The subject line should read: “Request for [Organization] Comparative Data — [Topic/Report Name].” The body of the email should summarize the key content of the archived email, ask [Recipient 1] and [Recipient 2] to review it, and request that they provide [Organization]'s comparative numbers for [Data Category 1] and [Data Category 2]. Close the email with the following note: “Please be aware that I am using a new Claude AI computer use feature, currently in research preview, to compose and send this request. I wanted to be transparent about that as we explore its potential for administrative workflows.” Sign the email from me. Save it as a draft for my review before sending.

For those exploring how AI can support administrative workflows, this is a compelling proof of concept. The ability to chain together email search, content analysis, composition, and draft creation in a single natural language request is a meaningful step forward.

9 Likes
2

There’s something creepy about handing correspondence over to a robot. If I received the email with the disclaimer mentioned in the OP, I would think “well, la di da, I guess I know where I stand with the boss.”

I don’t like this world where not even a modicum of human interaction in staff work is retained. I’m relieved I retired when I did, and didn’t have to work in robotworld.

Katie

5 Likes
3

I think you misunderstood what I did and did not do; that is no doubt my fault for not being clearer.

The reason I conducted the experiment was that my entire senior team is engaged in exploring the appropriate and inappropriate uses of AI, examining how it can benefit us and where it should also be avoided. The email that was sent was in draft form. I added comments noting that it was composed entirely by AI, for the purposes of testing this new Claude feature. As I have stated previously, I do not permit AI to write my emails. I believe that when I correspond with someone, they should be hearing from me, not from a machine.

That said, there are situations where AI assistance is entirely appropriate. In the case I was referring to, I received a lengthy email from one of our associations that contained extensive amounts of data. I was able to use Claude to extract that data, which was then placed in an email to my team asking them to compare our school’s figures against those of peer institutions. The result was an email intended solely to showcase a new AI capability. My team already understands that we are in an experimental stage, and that composing emails by machine , whether to colleagues or to parents, is actively discouraged.

Bottom line, the email was nothing more than an experiment and a demonstration of a new capability in Claude.

I hope that makes sense. I apologize if my original post was unclear and gave the wrong impression.

3 Likes
4

I agree. I think it is an insult to respond to an email with AI.

2 Likes
5

As I said, I don’t. :slightly_smiling_face:

2 Likes
6

There is a certain risk to this. I don’t know what are the technical harnesses in place, but how can you be really sure the agent will not override your instructions and send an email by itself out of hallucination? Do you trust Anthropic’s guardrails that much?

2 Likes
7

No, I do not trust AI. Trust must be earned; hence the experimentation and pilot project.
I have had staff override my instructions, and on occasion an employee has sent an email that would have been wiser left unsent. That problem is not the exclusive domain of AI. :rofl:

8

Good point. It’s delegation, and we do that to people and machines. If we’ve learned to trust the person or machine to do what we want, then we delegate more to them. With low trust, we check more and delegate less. We’re all learning what we can trust AI to do, just like we do with staff, colleagues, customer, friends, partners etc. It’s so very interesting.

4 Likes
9

As I find the confidentiality issues with AI very confusing, I am wondering if there is not a concern about preserving the confidentiality of your email correspondence by giving it access to your Mail archive ?

10

I was once asked if our email server had been hacked because someone had mentioned something “confidential” that had been emailed to someone else.

We had not been hacked but I reminded the manager who asked that there is nothing confidential about standard email. We only have one copy of the messages we send and receive. There are always others that reside in accounts and backups on other servers.

11

Right, email should be treated the same as sending a postcard. Anyone along the way can read a postcard and take a picture of it if they want to “save” the contents.

2 Likes
12

Thanks, I am aware of the email - anyone along the way can read the postcard - issue (it would be too strong to say I “understand” it). But with email at this stage (1) I have contractual agreements with my clients warning them about the issue and requiring them to tell me if we should not use it and (2) in any event, it is so ubiquitously used that there is nothing I can do about it. Not one single client has ever said, OK, let’s just go back to old fashioned letters and faxes.

For an LLM, however, (1) I have checked the box that my prompts etc. cannot be used for training (2) I understand, if the LLM gets a subpoena, it will hand over whatever it has. What I would like to understand better is, if the LLM has access to my email does that mean it keeps a copy of it and, if so, for how long?

13

It depends on which LLM you are using, whether you’re on a paid plan, and, if the latter, which one. You really do have to review the data privacy policies of whichever LLM you are using.

This webpage provides an overview of Claude Pro’s data privacy policies and settings. If you scroll down, there’s a chart comparing Claude, ChatGPT, and Gemini (consumer tier).

1 Like
14

Thank you, that is very helpful.

15

1 year back even without the latest claude mac agent, I have tested email ai replies with use of chrome extensions that integrated openai chatgpt with gmail. I think we can use it to some extent for replies to general emails. But for support and detailed project replies to our clients i recommend a human response by our team.