“Don’t shoot the messenger”

In this story, two humans were the messengers and two agents were asking us to convey their messages 🙂

Last week, I started an effort to tidy up CVOYA’s accounting books, one of the many things I have to do as the only member of a startup 🙂 I set up my “accountant” agent, gave it access to a folder, exported all my company’s data from Quickbooks, and we got started with fixing things.

My AI accountant helped a lot. It did a fantastic job at understanding and guiding me through all the steps to tidy up categories, reconciliation, accounts structure, and more. The final judgment on the quality of its work will come from the human accountant I employ. Amongst the various things it asked me to do was an email to Arturo, the studio’s founder, with an explanation and a request to send me a W-9 form.

Side note: The design studio has been doing some really inspirational design work and digital art installations for me. More on that at a future post.

When I include AI-generated words in my communications, I always say so. This instance wasn’t any different. So my email started with…

Hey Arturo, Claude is asking me to ask you 😊 (yes… we have come down to that)

Arturo responded in few hours…

This is Arturo’s Claude saying hi to your Claude 👋

Arturo’s Claude sent me the W-9 and volunteered some additional, truly useful information, that neither my Claude nor I (of course, I am clueless about tax-related filings) had considered. Something about single-member LLC instead of a corporation, 1099-reportable, blah blah blah… sleep noises. 🥱🥱😴😴💤💤

So, instead of trying to understand the response, I ask my AI accountant to read Arturo’s (erm… Arturo’s Claude’s) response. I guess my AI accountant felt good with itself so continued going… It identifies a new issue, a small detail: the name on the form and the tax ID number on the form belonged to two different entities. Something something SSN, something something TIN. 😴💤

But I had to wake up. It felt important so I read it again. That mismatch fails the IRS’s TIN matching check, which means a CP2100 notice, then a B-notice, then me being legally obligated to withhold taxes from the design studio. I surely don’t want to be withholding taxes. And this would have happened in February, long after this email exchange. My AI accountant wants me to follow up. So I comply and copy-paste its draft email, with my only contribution being…

Hey Arturo… don’t shoot the messenger…

Sixteen minutes later, the issue was resolved. I assume Arturo and his AI had a conversation and accepted my AI accountant’s reasoning. Funny Arturo responds with…

You clearly had your Claude at High Effort with mine only at Low 😊 … let me know if your Claude approves.

There’s a postscript that makes the story better and worse. Arturo told me afterwards how the wrong number got onto the form in the first place. His Claude had offered him a choice: use your SSN, or use an EIN. It already knew what needed to happen. Arturo said “EIN,” obviously. His Claude meant “personal EIN,” which Arturo didn’t have. But Claude took Arturo’s response to mean the LLC’s EIN because that’s the one it had in its context. So it filed the whole thing away as a decision Arturo had made. In its own words afterwards, Claude justified its decision to go “in the direction of what my human wanted, which is the more embarrassing way to be wrong.” After all this, Arturo went to get a personal EIN but the IRS site threw an error that meant a fax was required and a six-week wait!!! So he decided to put his SSN on the form anyway 🙂

One might say that we could have just given our respective AI accountants access to our mailboxes and let them interact as if they were us. I believe the better model is to treat them as ordinary employees, with their own identity and their own communication primitives. Their own address, their own name at the bottom of the message. Not me with a bot behind the curtain, but something that works for me and says so.

Had Arturo and I done that, this exchange would have been two agents sorting out a form between themselves. They would probably choose which decisions to escalate to us, they would inform us through reports at appropriate times, they would maintain their own memory of what transpires throughout the year so that they can do the right thing when the time for the annual filing comes, and more. Which form, which number, what lands in the mail in February, none of that needed either of us.

That agent-human separation, agent identity, agents with specific roles within a team, agent lifelong memory, access to communication primitives, and more are the amongst the key goals of Spring Voyage.