Hello fellow keepers of numbers,

When I popped open my document to start the newsletter this week, I thought to myself, “This was a pretty slow week.” And look at everything we’re covering. What has life become?

Claude can now control your computer. It’s already better at it than your grandparents. Maybe better than your parents if you were born before 1990. Still not very good though…for now. Aiwyn Tax also exposed their tax calculation engine in Claude via a connector. Likely much more to come on this as they build out more functionality with the connector. Lastly, Delve was accused of helping hundreds of companies get questionable SOC reports.

Plus, stick around for a demo that might help you use AI to write emails in your specific style and tone.

THE LATEST

Anthropic announced a new "computer use" feature for Claude Cowork and Claude Code, now in research preview for Pro and Max users. Computer use lets Claude control your Mac directly from the desktop app. Claude can see your screen, click and type, open apps, work with files, and navigate browsers to complete multi-step tasks.

Anthropic has also recently released Dispatch, which is a feature that allows you to message with your Claude desktop app directly from your phone. You text Claude a task in the mobile app, and it executes on your desktop in the same continuous conversation.

The computer use feature currently runs on macOS only and requires both the Claude desktop and mobile apps, plus a paid Pro or Max subscription. You choose which folders and connectors Claude can access, and it must ask for approval before touching new apps or taking sensitive actions like deleting files. Activity runs locally on your machine rather than in Anthropic's cloud environment.

Anthropic explicitly warns that computer use shouldn't currently be used with sensitive data or PII.

Why it’s important for us:

This is a mind-boggling feature when you see it working. It’s also a look into the future. We’re probably not that far off from a world where AI can control computers as well as an advanced human user. At a certain point, they’ll probably be far more efficient at operating it as well. Let’s face it… We’ve already been beaten at our own game by the AI add-ins in Excel that create full workbooks in seconds. Why would operating the computer be any different?

But the future isn’t now. Currently, this is extremely slow. There’s really no benefit to computer use right now unless you’re away from your computer and trigger it via your phone using Dispatch. Even then, the failure rate for tasks involving computer use is pretty high.

Then there’s the security aspect. This is certainly not ready for business use yet. The AI models are smart, but computer use is a different domain for them. We can’t yet trust them to properly protect sensitive data. If they have full access to anything on your computer, they could easily share a file with someone they shouldn’t, put sensitive information in a public form, or any other number of risky scenarios.

The other big security concern is auditing the actions of the AI. If it runs for 30 minutes on your computer, it’s incredibly difficult to audit every little thing it did. Maybe it did exactly what it was supposed to. But maybe it took a short 3-minute break, popped open eBay, and bid on that Taylor Swift keychain that you definitely didn’t want. That’s obviously what I tell people when they see mine.

Aiwyn announced that its tax engine is now available as a connector inside Claude, allowing users to perform 1040 tax calculations, run what-if scenarios, and generate completed return PDFs directly through Claude's conversational interface.

The integration is powered by the same tax engine behind Aiwyn Tax, which acquired Column Tax's technology and has processed over one million returns in production. Users can upload tax documents like W-2s, and Claude handles the rest through Aiwyn's deterministic calculation layer.

Aiwyn is positioning this as a way for firms to explore AI-assisted tax workflows without disrupting existing systems like CCH or UltraTax. It's not a replacement for production tax software and doesn't file or e-file returns. Instead, it's designed as a low-risk entry point for firms to test how conversational AI and tax calculations can work together before committing to deeper workflow changes.

Why it’s important for us:

This is basically the proprietary tax calculation engine being exposed directly in Claude. I think this is the first tax software to be integrated into AI, at least that I can remember. Intuit has a deal with Anthropic to include QBO as a connector, but obviously that’s not tax software and also still not available.

Ultimately, it sounds like this is intended to be something accounting firms can use to test how an AI tool can interact with tax software. At the very least, this is a great marketing play by Aiwyn.

The future is starting to look like AI becoming the interface and software applications becoming calculation engines and data storage. Expensive data storage…

The connector will probably be useful as-is for scenario modeling, with the caveat that they’re suggesting you first scrub your data. Some firms have already become comfortable with putting data into AI models, so I’m not sure why this connector would be different than any others. I suspect the Aiwyn warning is really just a legal liability concern because the industry has no guidance from regulatory bodies.

As a side note, hopefully this integration will finally open the eyes of the regulatory bodies in the accounting space to the urgency we need here, particularly for tax. The industry needs some type of official guidance.

An anonymous investigator published a long-form Substack report alleging that Delve, a compliance automation startup, mass-generated nearly identical SOC 2, ISO 27001, HIPAA, and GDPR reports from templates rather than real audits.

The investigation is based on a leaked Google Sheet containing hundreds of draft reports allegedly showing Delve pre-writing auditor conclusions before any evidence was reviewed. TechCrunch and other outlets have since covered the story, and at least one named customer has publicly confirmed it left the platform.

The leaked dataset reportedly includes 400+ SOC 2 reports where company descriptions, test procedures, and auditor conclusions are almost word-for-word identical, including the same grammatical errors and "could not be tested" language across hundreds of clients. Evidence, such as board minutes and risk assessments, was allegedly offered as one-click "adopt this template" artifacts, creating the appearance of governance that never actually happened.

Delve has published a formal response disputing the claims, saying it's a software platform rather than an audit firm and that independent, accredited auditors issue all final reports. Critics point out the rebuttal doesn't engage with the core allegation: that hundreds of reports are statistically near-identical across multiple auditors, and that the named auditing firms appear to be Indian certification mills operating through US-facing structures.

Why it’s important for us:

File this one under “maybe not all things should be automated.”

In all seriousness, I contemplated skipping this news because it’s still somewhat speculative at the moment. There’s a mountain of concerning evidence though, and the response was a bit lackluster because it didn’t really address the biggest concerns. So we should at least be skeptical of Delve.

I think the most important takeaway from this story is that we should probably be a little more attentive to SOC reports. Especially in the new age of AI where people want to automate every little thing at nearly any cost. We’re entering a new world of IT and data security that we’ll all have to figure out how to navigate.

Obviously, we can’t do our own compliance audits, so it’s sort of a problem without a solution. But maybe we can open our eyes to some red flags that might exist, and at least verify the independent auditor involved in compliance audits.

PUT IT TO WORK

This week, I decided to test Claude Cowork’s ability to create a skill to draft emails in my very eloquent, unique voice… Okay fine, I talk not much good and help needed is from AI bot.

Grammarly just deleted my account on my behalf for writing that last sentence.

Anyway, here’s a video where I create this skill. Cowork unexpectedly created a custom web app to A/B test the draft emails with the skill, so it was a fun one.

Using Claude Cowork to create personalized emails - Watch Video

TRENDING NEWS

Claude mobile app now supports interactive apps in conversations: Connectors that are integrated apps within Claude can now render information in the chat on your mobile device. This is very cool and could be useful for doing things while on the go.

Google launches Gemini 3.1 Flash Live to power Search Live: You can have real-time voice and camera conversations with Google Search. Point your phone at something, ask what it is, and get spoken answers with web links.

Apple plans to open Siri to all AI models: You’ll be able to use ChatGPT, Claude, or Gemini as the brain behind Siri. This could be a major shift in how we interact with our phones if integrated well.

Claude Code added channels to route conversations through your phone: Thus far, native channels include Telegram, Discord, and iMessage (if you’re a macOS user). You can now do the same things you do in Claude Code by messaging the respective app on your phone. I’ve been testing this for the last week, and it’s quite impressive. Likely more on this from me in the near future.

WEEKLY RANDOM

Amazon launched a Health AI agent on its website and app that lets you ask health questions, interpret medical records, manage prescriptions, and connect with licensed providers through One Medical. If you’re a Prime member, you also get 5 free virtual consultations.

We’ve seen a lot of AI companies focus on personal health recently, most notably ChatGPT. Love or hate Amazon, this is a pretty cool offering. I’m a Prime subscriber because my family averages 2.7 packages per day, which seems impossible, but I promise it’s not. This might be something I try.

As much as I use AI, I have yet to adopt it for my personal health. I’m sure a lot of people have concerns about putting medical records into an AI, but I think the positives for me outweigh the negatives. In some exceptional cases, it could literally save lives.

My wife also says I’m a hypochondriac (no comment), so it’ll at least improve her life since now all my complaints will be made to AI.