The Permission Slip

The most expensive sentence in financial services is six words long. Are we allowed to do that?

Nobody in that meeting is wrong for asking. That's what makes it so hard to argue with, and it's why everything stops the second somebody says it out loud.

The fix is a lot smaller than the pause it causes, and most of it comes down to which Claude plan your firm is actually on. That takes about thirty seconds to check.

One more thing before you scroll. Bridge AI has a newsletter for wealth management and finance now. It's called AI Under Management, it runs every other week, and the first edition goes out this Thursday. What's below is the teaser. Subscribe to AI Under Management and you'll get the whole thing on Thursday.

This week:

The six words that stall AI adoption at every finance firm, and how to go get the answer

Why I stopped opening my own spam folder

A frontier AI model with nobody's name on it, Claude designing drug candidates that beat the published record, and a $966 million fine for letting software deactivate people

The Deep Dive

The Permission Slip

It goes the same way every time.

A few people at a firm start using AI on their own. Nobody planned it. Within a few weeks they're getting hours back, and the work coming out is better than what it replaced.

So they want to go further. Stop pasting things in by hand and let it reach the systems where the work actually lives. The CRM. The portfolio system. The accounting file.

That's when somebody asks the question in a meeting. Are we allowed to do that? Nobody knows. And once it's asked out loud, you can't un-ask it.

Everything stops. Not because compliance said no, but because nobody had asked yet, and now they have to.

It isn't a technology problem. It's a plumbing problem, and it's a smaller job than it feels like.

The thirty second version of the fix: check which Claude plan your firm is actually on, because most firms get this part wrong.

Free, Pro, or Max. That's a personal account. Your firm has no administrator, no access, and no contract with the vendor. Since June 3 that also puts you out of step with Regulation S-P, which now requires you to oversee the vendors holding client data. You don't have a vendor to oversee. You have an employee with a subscription.

Team. Better, and it fools people. An owner can export conversations, but it's a manual export run by one person, and any employee can delete their own chat first. No audit log.

Enterprise. Where it starts working. Audit logs, plus a compliance API an archiving vendor can pull from continuously without anyone remembering to.

And while you sort that out, don't stop building. A specific client's account, holdings, or AUM stays out until this is settled. Market research, investment process, marketing, and prospecting do not.

The compliance work isn't the project. It's the permission slip. Go get it, and build while you wait.

THE FULL VERSION LANDS THURSDAY

That was the teaser. The whole thing, including what Enterprise actually costs, the two settings that keep the bill from running away from you, and the archiving vendor you still need sitting behind the compliance API, goes out this Thursday in the first edition of AI Under Management. Subscribe to it here.

From the Field

Nobody Wants To Go Dumpster Diving In Their Spam Folder

Everybody has the story, though. The wedding invitation that landed in spam. The offer letter nobody saw for two weeks. So you open the folder once, scroll past ninety pieces of cold outbound selling you SEO, find nothing, and never open it again.

I stopped opening mine. Every Friday, Claude opens it for me on a schedule.

It pulls the entire Junk Email folder out of Outlook and sorts every message into three piles. Rescue, worth a look, and correctly filtered. The rules are specific to me. A real person replying to my Substack gets rescued, because readers are part of how this business finds clients. Anyone at a current client gets rescued. Anything related to my personal life gets rescued. A beautifully written cold pitch with a fake "Re:" glued to the front of it stays exactly where it is.

Most Fridays the answer is one line. Ninety two messages, all correctly filtered, nothing for you. Boring is the correct output.

Twice in two months it has not been boring. Both times it was a real person I actually knew, writing to me directly, buried under the cold outbound and sitting there unanswered. I would never have found either one.

Claude does not move anything on its own. It hands me a short list with a link on each item, and I decide what comes out.

Why this travels: every business has a folder nobody opens. Junk mail, the shared inbox, the form submissions, the voicemail box. The value is not in reading all of it. It is in having something read all of it and tell you the four things that matter.

AI News Worth Your Time

Three Stories From This Week

A frontier AI model showed up with nobody's name on it

On August 20, a model called Ox Alpha appeared on OpenRouter. Free to use, a one million token context window, text and image and video input, and a listing that described the developer only as a third party who had chosen to stay anonymous. Developers immediately started benchmarking it and found coding performance sitting at or above the paid frontier models. Stripe CEO Patrick Collison called it very impressive. Speculation has landed mostly on Chinese labs, with Zhipu's GLM family the leading suspect, though nobody has confirmed anything and the trail keeps moving.

My take: The interesting part is not who built it. It is that a model this capable can be handed out for free by an anonymous stranger and the industry just absorbs it. Two years ago a frontier release came with a keynote. Now one shows up unannounced on a Thursday. If you are choosing AI tools for your business, the model underneath is the least durable thing you are picking, because it will be replaced within months no matter what you choose. Pick for how the tool handles your data, plugs into your systems, and gets used by your team. READ THE TECHCRUNCH REPORT →

Claude designed drug candidates that beat the published record

Anthropic handed Claude fifteen protein targets and asked it to design binders, the small proteins that latch onto a target the way a large share of modern medicine works. Two outside labs, Adaptyv Bio and Twist Bioscience, physically built and tested the designs. Claude produced working binders against fourteen of the fifteen, with a 35% hit rate when focused on one target at a time, against the 10 to 15% that is normal for the field today. Some designs bound more tightly than the best result ever published. In a second test, Claude Opus 5 was given a raw file from a lab instrument, in a proprietary format only the manufacturer's software is meant to open, and returned a finished purity analysis in 19 minutes at 96.4% against the lab's own 96.33%.

My take: The chemistry half is the part business owners should read twice. Claude was handed a file it had no software to open, worked out the encoding itself, checked its own math against the instrument's totals, and produced the report that normally takes a chemist an hour by hand. Every industry has a file like that. The export from the old system, the report only Susan knows how to read, the format your vendor will not document. That work was specialized last year. Now it is just work nobody has gotten around to. READ THE ANTHROPIC REPORT →

Uber was fined $966 million for letting software deactivate drivers

The Dutch Data Protection Authority fined Uber 825 million euros, roughly $966 million, the second largest penalty ever issued under GDPR. The finding was that driver accounts were deactivated through an automated process without sufficient warning or human oversight. Deputy chair Monique Verdier said a computer should not make decisions on its own that carry consequences that large. Uber disputes it, says no permanent deactivation happens without human review, and is appealing. The whole case traces back to one French driver who lost his account in 2019 and then collected testimony from 170 others.

My take: You do not need to be Uber for this to reach you. If AI is screening your applicants, scoring your customers, or flagging accounts for suspension, the question you will get asked is who reviewed the decision, not whether the model was correct. This is why every workflow we build at Bridge AI ends with a person. The AI can do the analysis. Somebody still has to own the call. READ THE TECHCRUNCH REPORT →

That's it for this week.

If this was useful, forward it to the person at your firm who is about to ask whether you're allowed to do that. That's who this is for.

And if you want to talk about what AI could do for your business, not in theory, but in your workflows, with your team, on your timeline, reach out at [email protected]. That's what we're here for.