Hello fellow keepers of numbers,

Last weekend my son turned two. We had a dino party with family and friends. He ran around playing with his new toys, reading his new books, and we even had an appearance from one of those hilarious dino outfits. Only a few years until AI takes my son’s job and I just have a tiny little robot kiddo running around.

Just kidding, chill… please don’t call CPS on me. I love my human children very much.

This week, we got a few really interesting announcements. First, Mercury launched Mercury Books, which is a tool that categorizes and reconciles transactions as they happen. The QBO is dead crowd is back at it again.

We also got a few major announcements from Anthropic, none of which have hit my account yet because they’re probably mad that I’ve been talking up ChatGPT lately. They announced that chat and Cowork are going to merge. Most tasks will probably just go through Cowork now, which they should be anyway, but Claude will decide for us. They also announced Claude Docs and Slides as they move into direct competition with Google and Microsoft Office apps. Lastly, Anthropic is rolling out Claude Code Projects, which manages all the context, memories, and threads for each of your separate projects.

Plus, stick around for a demo showing you how to create a plugin in ChatGPT that you can use from anywhere (even your phone).

The Latest

Mercury launched Mercury Books, full double entry accounting software built directly into Mercury accounts. It automatically pulls activity from Mercury banking, IO Cards, Bill Pay, and invoicing, then categorizes and reconciles transactions as they happen.

Businesses can also connect external bank accounts, cards, payroll platforms, and other financial tools. Mercury Books supports cash and accrual accounting and generates profit and loss, cash flow, and balance sheet reports using current account data.

Mercury’s Command AI can answer questions about spending, trends, and general ledger codes. It can also handle work such as writing journal entries, organizing the chart of accounts, and categorizing transactions in bulk, though Mercury says its responses and suggested actions should be reviewed before use.

Businesses can invite their existing bookkeeper or accountant into Mercury Books or request an introduction to one of Mercury’s accounting partners. The product requires a Mercury business account and is free through December 31 for customers who subscribe by December 1. After the promotional period, it costs $35 per month and does not include bookkeeping services.

Why it’s important for us:

I've already seen plenty of accountants shitting on this announcement. I understand where they're coming from. Maybe it makes me a bad accountant...but I kind of like this.

A huge part of what QBO does is pull together data from banks, credit cards, payroll platforms, and other systems, then apply accounting logic. None of that is magic.

QBO's biggest advantage is that businesses and accountants have used it forever. It's embedded in their workflows, and changing accounting systems is a massive pain.

But what would stop banking platforms from building more of that accounting logic themselves? They'd be pretty well-equipped to categorize transactions based on a chart of accounts. They'd be well-equipped to match receipts, bills, and invoice payments to the underlying transactions. They'd be well-equipped to reconcile that activity automatically.

Mercury is especially well-positioned for this. It's already a software layer built on top of partner banks, and it already connects with external banks, credit cards, Stripe, Gusto, and more. I'm sure they have a similar long-term vision to what I'm laying out here.

I use Mercury myself, and I set up Mercury Books yesterday. My bank, Stripe, and Gusto were connected in about 30 seconds. Mercury is also a really great product that a lot of accountants already recommend to clients. I actually have a decent level of confidence they can pull this off.

There's still a lot to improve. Out of the box, Mercury Books will probably be most useful for straightforward cash basis businesses. It'll need a lot of work before it can handle more complicated accrual basis businesses. And there's a large gap between supporting accrual accounting and reliably handling complex GAAP accounting.

It's not going to kill QBO. It's hard to imagine anything will. Those are ridiculous claims. But there are really viable alternatives popping up now. Mercury Books could be a really interesting one, especially for accounting firms who already recommend Mercury to clients. And for $35 per month, I like the idea.

Anthropic began rolling out a single Claude interface that merges Claude Cowork and chat. As the update reaches accounts, Claude will decide whether a task needs a quick response or a longer Cowork workflow while using the same conversation history, skills, and connectors.

The company is also rolling out Claude Docs and Claude Slides in beta. Users who receive access will be able to create and directly edit documents and presentations, collaborate through comments, share the finished work by link, present slides inside Claude, or export them to PowerPoint and PDF.

Claude Design is also moving into regular conversations alongside Cowork. Documents, presentations, and designs created in Claude will remain connected to the conversation that produced them and can be opened and edited from a phone or computer.

The changes are rolling out first to Pro and Max subscribers across web, desktop, and mobile over the coming weeks and are not yet available to every eligible account. Team and Free plans will follow, while Enterprise customers will receive at least 30 days’ notice before the change.

Why it’s important for us:

There are two big updates here.

Merging chat and Cowork:

This makes a ton of sense. I use Cowork for probably 99% of what I do in Claude now. If the question turns into actual work, I'm already in the place with the tools, skills, and project context I need. I suspect most Claude users haven't made that switch yet and are still using chat more than they should.

Claude can decide whether it should give you a quick answer (chat) or do the work (Cowork). Much simpler. Especially for people who don't spend an unhealthy percentage of their day testing every new button Anthropic releases.

Claude Docs and Slides:

This puts Anthropic in direct competition with Google and Microsoft. Claude Docs and Claude Slides are essentially the equivalent to Google Docs and Slides. I'd imagine Claude Sheets are coming soon too.

But my mind immediately went to document storage.

One of the biggest pain points for accounting firms trying to use AI is getting their files into the AI tool. If your firm uses something like CCH Document or GoFileRoom, there's no connection to Claude. You have to find the file, pull it out of the DMS, upload it to Claude, and give Claude enough context to understand what it's looking at. Then you do it all over again the next time.

With Claude Docs, the document is created inside Claude and stored in your Claude account.

So let's assume this rollout goes well. A year from now, if you're creating work inside Claude Docs, Claude Slides, and whatever else follows, do you still need the same type of DMS?

If you're working on something for a client, you could just ask Claude to surface the document. Does it matter which folder it's stored in? Does the folder structure even need to make sense to a human if Claude can consistently find the correct file and understand what it's for? Probably not.

Some of this obviously still matters. You need version history. You need retention policies. If there's a legal battle or the IRS is performing an audit, you have to surface the necessary information. There are also questions about permissions, security, and whether businesses want all of their work stored inside an AI tool.

But it makes me wonder what a DMS looks like when the documents are already being created and stored inside the AI tool you're using. There are still some obvious improvements that I'm sure Anthropic will roll out soon.

For example, a natural next step would be connecting these files to Claude Projects. Open a project for a client or engagement and see every Claude Doc and Slide being used for that work, along with the conversations and context behind them.

I don't have access to it in my Claude Team plan yet, so I haven't been able to test it. Still, I think this was an obvious next step for Anthropic. OpenAI is surely working on this too. I'll be very interested in the direction of this project over the next year or two.

If this works, the long-term story isn't that Claude connects to your DMS. Claude becomes the DMS.

Anthropic began rolling out Projects, a new Claude Code feature that coordinates related work through one ongoing conversation. Users can submit several tasks at once, and Claude decides which tasks need separate threads, which belong together, and which can be handled directly in the main conversation.

Each thread is a separate Claude Code cloud session with its own context window. The threads can run simultaneously, continue after the user closes their laptop, and be opened individually for review or additional instructions. New threads start with the Project’s shared repositories, files, instructions, and memory.

An Overview panel organizes the threads by status, including those still working, waiting for input, or ready for review. Projects can also work with uploaded documents and Google Drive folders, with completed files collected in a shared Library.

Projects are in beta for select Claude Pro and Max subscribers who use Claude Code cloud sessions and don’t already have projects on the web or desktop. Anthropic says access will expand to more Claude Code users on those plans over the coming week, with Team and Enterprise support coming later.

Why it’s important for us:

Last week, I said I loved the direction Cursor was going with Projects. Anthropic has now rolled out essentially the same idea for Claude Code, and I love it here too.

Once you see it, the concept feels obvious. You go to a project and tell Claude what you need. Claude decides whether that's one task or three, sends the work into separate threads when needed, manages the context and memory across them, and comes back with the result.

It's basically like talking to a project manager. I don't need to understand all of the complexity happening behind the scenes. I just tell the project manager what I need, and it manages the people, conversations, and information required to get it done. Except in this case, all of those people are Claude.

Projects haven't reached Claude Team plans yet, so I haven't been able to test this myself. But I'm really excited to use it. I also suspect we'll see some version of this make its way into Cowork since it already has projects. This just makes a ton of sense to me.

Trending News

NeoCognition introduced ApprenticeBench, a simulated seven-month accounts payable apprenticeship where Fable 5.1 and GPT-6 Astra reportedly beat human professionals: This is a massive test where they put AI to work and let it learn over time. The results were interesting because Fable and GPT-6 Astra scored better than the human tester. But the agents actually cost more per task and got slower over time as context and memory piled up. This was autonomous, so I'd expect those shortcomings would be less relevant when you put a human in the loop with AI. But it shows there's still a lot of solutions required for autonomous accounting.

OpenAI added collaboration, private sharing, and custom domains to ChatGPT Sites: I really love ChatGPT Sites, and I love how much the OpenAI team is focused on improving them. The major shortcoming right now is that the connectors in ChatGPT can't hook into a Site yet. Claude Artifacts offer this feature, so I suspect it'll be one of the next updates to Sites.

Dario Amodei called for slower frontier AI development, and Altman, Musk, Hassabis, and Nadella quickly backed him: Despite the sentiment of those holding office in the United States, it seems like all of the leaders of AI companies are on the same page about slowing down and wanting more regulation. It's a little bit scary at the moment, and I hope these leaders can put competition aside and come to a reasonable agreement. Even if you think the claims of danger are overblown, doesn't it seem logical to make sure we build things correctly, regardless of the technology or industry?

Ramp published Accounting Bench, where the best frontier models averaged below 50% across 137 difficult accounting tasks: Finally, we get a reputable benchmark for accounting. And it also proves the point that those marketing videos for companies saying they're automating accounting are full of shit.

Anthropic launched Claude for Financial Advisors with industry connectors and skills for portfolio, compliance, estate, and tax work: OpenAI just launched their plugin for financial services last week. Anthropic launched this a few days later. We also got ChatGPT for law just the other day. I suspect accounting is coming soon...maybe? We'll see.

TypeSafe AI released Jev, a structured-decision model it says runs 40 to 200 times faster than frontier LLMs: This is a pretty technically nerdy one. But it seems like a fairly big breakthrough for AI. This model is built to make decisions. And fast. Oh, and cheap. So why would this be useful to accountants? Well, think about where accountants make decisions. Categorizing expenses. Making decisions on research that's applicable to a specific client's situation. Determining where something goes on a return. There are a long list of these. If vendors switch from classic AI like Claude and ChatGPT to Jev, these decisions happen exponentially faster, exponentially cheaper, and with more reliable confidence scores. This could be a huge breakthrough for software using AI.

Notion launched a collaborative skills library and API for reusing agent instructions across ChatGPT, Claude, Gemini, and other tools: I need to investigate this more and maybe even do some testing. But on the surface I absolutely love this idea. GitHub feels like overkill for a lot of firms, but it's pretty important right now for skills and plugins. If Notion can become a simpler interface for non-technical users and accomplish the same thing, I think this is an enormous win.

OpenAI launched a ChatGPT add-in for drafting, editing, proofreading, and formatting documents inside Microsoft Word: They're so late to the game here, but I guess the saying better late than never applies. I haven't yet tested if ChatGPT can talk between Office add-ins like Claude. Passing information from the Excel add-in to the Word or PowerPoint add-in is a really awesome feature of the Claude add-ins.

OpenAI released GPT-Live-1, a voice model that listens and speaks simultaneously for 5 cents per minute: AI voice agents are becoming a much more realistic option for business use cases. There have been some major leaps in this category even in just the last few months.

Google released Gemini 3.8 Live with real-time voice, background tasks, and Workspace integrations: See the comment directly above. AI voice agents might be a big part of 2027.

Put It to Work

One of the biggest benefits of ChatGPT over Claude right now is how OpenAI has implemented plugins. In this demo, I walk through every step of the process from creating a skill to packaging it as a plugin to deploying it to your workspace.

You’ll learn how to turn the skills you use now into something that you can use in your ChatGPT account on the web and your mobile device.

Weekly Random

Would you let an AI robot clean your house?

Before you say no, it can fold towels and make your bed. I rest my case.

Figure tested Helix 2.5 in 30 homes it’d never seen before. Without being retrained for any of them, it tidied rooms, folded towels, and made beds.

That’s pretty cool. It’s also a little terrifying.

Over the last few weeks, I’ve written about AI agents escaping controlled tests, finding security holes nobody knew existed, and breaking into real systems because that happened to be the easiest way to finish their assignment.

Now we’re giving them arms, legs, and access to knives.

To be fair, those are different AI systems, and there’s no evidence Helix is plotting to hack my fridge. But AI safety feels a bit more personal when the model lives with me.

I think I’ll pass on them for now. Not sure I want to be the first person to be murdered by an autonomous robot. That’d be embarrassing.

But I do really like the idea of never having to do my laundry again. Or clean my house. Or even cook my dinner. I’d have so much more time to sit on the couch and do nothing but watch sports.

I have a feeling this robot wouldn’t last more than 2 days in my house. It’d see the walking tornadoes that my wife and two kids are, immediately go into panic mode, and end up hurling itself out the second-story window to end the madness.