Hello fellow keepers of numbers,

This week, I’ve had to grapple with my aging body because I threw out my back working on my garage door. I’m going to be burnt toast when AI takes everyone’s jobs, and I have to make money doing physical labor.

Until then, I’ll just keep writing about it. Digits released their new MCP that offers write functionality, and introduced locked periods. Ramp introduced Glass, their product that helps a company share knowledge, skills, and chats across teams throughout the organization. Lastly, Anthropic released Claude Dashboards as a new type of artifact. I demoed it last week, but they deserved more coverage because of how good they are.

Plus, stick around for a demo of ChatGPT Space. Nope, not the kind with stars and planets. The kind that stores documents and even brings in your Google Drive and SharePoint docs. I also showed ChatGPT Pages, which is the equivalent of Claude Docs.

The Latest

Digits launched Write MCP, allowing connected AI assistants to edit transactions, post journal entries, manage bills and invoices, and update the chart of accounts.

The tools support work such as correcting transaction categories, fixing split lines, creating bills, drafting invoices, and adding accounts. Agents receive only the write tools allowed by the connected user’s Digits role.

Digits also introduced Locked Periods. Changes in open periods apply automatically, while edits to locked periods enter a review queue for an admin or manager to approve. The controls apply to both people and AI agents.

Write MCP is available now at no additional charge and connects to Claude, ChatGPT, Cursor, and other MCP-compatible tools.

Why it’s important for us:

Digits continues to push the envelope for AI in accounting.

The MCP now lets Claude or ChatGPT make changes directly in the general ledger. This is obviously a big step beyond pulling reports and analyzing the numbers.

Another thing I really like about its MCP, which I don't think is new, is that one connection can switch between clients. The AI can identify the business you want to work with, select it, and then switch to another one. No manually disconnecting, reconnecting, and logging back in for each client.

That alone is a huge improvement over the QBO connector experience. And QBO's connector is still very limited in the actions it can take in the platform. Digits is significantly ahead at this point. For firms already using Claude and ChatGPT heavily, I think this is becoming a major differentiator when choosing a GL.

I also really like the locked periods feature. An agent can propose a change to a closed period, but it can't apply that change without approval. This is a pretty forward-thinking feature because we all probably have some hesitancy in releasing agents to make changes throughout the GL.

Admittedly, I'm not doing the day-to-day bookkeeping in a firm. I'd be interested to hear from the people who are to see how useful this feels in practice. On the surface, I think it's a really good guardrail as we build more trust in agents doing this work.

If I were starting an accounting firm today, I think I'd have a hard time choosing anything other than Digits as the GL.

Ramp introduced Glass, an AI workspace focused on shared company knowledge. Ramp describes the product as a way for teams to learn from work done across the organization, rather than keeping each employee’s AI work separate.

Glass joins a broader expansion that includes Stack for month-end close, Accounts Receivable, AI Spend Management, and Router for choosing AI models.

Ramp also raised $1.85 billion at a $60 billion valuation. Co-founder Eric Glyman said the company now serves 85,000 businesses and passed $2 billion in annualized revenue over the summer, with positive cash flow.

Ramp is accepting early-access requests for Glass. The company has not published pricing or a general availability date.

Why it’s important for us:

Ramp is starting to feel less like a fintech company and more like an AI lab. Outside of Anthropic and OpenAI, I think they've released some of the most interesting products for accounting and finance over the last six months.

Stack has gained some steam in the last few months. It’s similar to Claude Cowork, but with accounting and finance skills and connectors included out of the box. Less setup and customization required for someone who may not be an AI nerd like some of us.

Glass seems relevant for everyone. A lot of companies are trying to figure out how to take their company knowledge and make it useful across the organization. One person builds a useful skill or teaches their AI how the company handles something, and everyone else still has to figure it out for themselves.

I like that Glass is trying to make that work reusable. The knowledge and skills can be shared across the company, and people can even work together in the same AI session. It's using models from other AI companies, with Ramp building the harness around them.

Cost seems to be another benefit. Glass routes requests between models. In a world where it’s routing correctly, it’ll cut down costs while still getting a similar quality output. You don't need the most expensive model for every task, especially when hundreds of people are using AI throughout the day.

I don’t think this is relevant for smaller firms. But for larger firms that are on the enterprise plan for Claude or ChatGPT and have trouble controlling costs, this could be appealing.

You can build a lot of this with Claude and ChatGPT already. But there's a difference between building it from scratch and having something already set up for you. Maybe you give up some flexibility by using Glass. But for firms that don't have the time or people to build and maintain it themselves, I think that could be a pretty reasonable tradeoff.

Anthropic introduced Claude Dashboards and Claude Motion, two new beta features for building live reports and animated explainers. Claude Docs, Slides, and Design are also out of beta and available on every plan, including Free.

Dashboards can use data from Claude’s connected tools, as well as data platforms such as Snowflake, BigQuery, and Databricks. Users can ask questions in plain language, and Claude builds a dashboard that updates as the underlying data changes. Each chart shows its query and last refresh time.

Motion turns reports, charts, and walkthroughs into editable animations that can be downloaded as MP4 files. Users can change the wording, numbers, and timing directly or ask Claude to revise them.

Dashboards is in beta on paid plans, while Motion is limited to Team and Enterprise. Both are off by default for Enterprise organizations and require an administrator to enable them.

Why it’s important for us:

I showed Claude Dashboards in last week's Put it to Work, but I don't think I gave this update enough attention.

Think about the value you get from something like Power BI or Tableau. You connect your data, build dashboards, and use them to understand what's happening in the business. But building and maintaining those dashboards takes some expertise. With Claude, you can describe what you want in regular human language and get a dashboard within minutes, connected to the same underlying data via a connector.

In both public accounting and private industry, I've worked with teams that maintained our Power BI and Tableau dashboards. You'd ask for a change, and it might take two weeks or more to get back to you. Even for something relatively simple. With Claude, you can ask for the change yourself and have an updated dashboard in 2 minutes.

And you're already in Claude. You can ask questions about the numbers, have it explain a chart, or ask it to look into something that seems off. I really like being able to go from seeing something interesting in a dashboard to working through it with AI in the same place.

You also don't necessarily care about the same metrics as the person sitting next to you. Maybe you each have your own dashboard connected to the same data source, with the views and details that matter to your work. That becomes a lot more practical when you don't need someone else to build and maintain every variation.

Motion is probably less relevant to accountants, but I can see uses for animated charts and explanations in presentations or marketing materials. Dashboards is the part I'm most excited about though.

I’ve said it too many times to count, but Claude Artifacts continue to be one of the most exciting things happening in AI right now.

Trending News

Gusto introduced AI payroll workflows and a multi-client MCP that connects selected clients through one conversation: Obligatory round of applause for another MCP, but the multi-client part deserves extra credit. This is what we need for all connectors so nobody has to disconnect and reconnect each client account or build custom connectors as a workaround.

Google released Gemini 4 Argon to AI Pro and Ultra subscribers, with background tasks and more app connections: This was announced a few weeks ago and is now available. Early reports are that it’s great at math and science. Not sure that helps accountants very much. Google is still significantly behind in a lot of areas, especially their Gemini app.

Meta launched Muse Wonder, a new model family for personal and professional agents: Meta have made a lot of progress recently. They hired Alexandr Wang from Scale AI a while back, and it seems to be paying off with the progress they’ve made in new models and their Muse agent. They’re still way behind the leaders in the category for now though.

Meta released Muse Spark Open with downloadable weights and support for running locally on two Mac Studios: This probably isn’t that relevant to most accountants, but it’s a good reminder that local models are still making progress. It’s becoming a hotter topic as the models become easier to use and more capable. Running it on Meta’s suggested setup is pretty expensive to get up and running.

Microsoft introduced Microsoft-Decision-1, a model built for fast classification, routing, and other structured decisions: Jev, OpenAI's Decisions API, Cloudflare, and now Microsoft. This category is moving quickly, and I think it's one of the more useful developments for accounting automations. We don't need the smartest model spending a minute deciding which folder a bank statement belongs in or which account to assign a transaction.

TypeSafe AI added experimental image support to Jev 1.14: This could make those fast, cheap decisions useful for scanned documents too. I'd be interested in testing document classification or locating information on a form, but TypeSafe is offering image support for testing and prototypes for now. It’s not ready for live client work yet.

Microsoft previewed Copilot Studio agents with their own accounts, shared process memory, and scheduled or event-triggered work: I like this direction much more than another chatbot bolted onto Teams. An agent that remembers the prior budget review and follows up on unresolved questions sounds useful. It's still rolling out through Frontier, so I'll reserve judgment until we see how much of it actually works.

OpenAI introduced a Meetings plugin that takes notes without a bot joining the call: This is currently Mac-only and in beta. But it seems very similar to what Granola does now. No video included. I don’t think I’ll use it personally because I want my meeting notes available to use in Claude as well, and I doubt there’ll be a connector to get the ChatGPT meeting notes when using Claude.

OpenAI added support for creating dots in the ChatGPT mobile app, letting your dot start or continue Codex tasks, and asking it to update scheduled tasks: As they continue to update Dots, they become more attractive to use. However, I still do most of these things already with the ChatGPT app, so I’m not sure having a Dot will save me much time just yet. I’m going to start testing this now though.

Grok Bot added its own email address for contacting businesses, signing up for services, and scheduling meetings: The updates keep coming. I can envision a lot of great use cases for this, but I’m personally still a little nervous to have it take action on its own with its own email.

Anthropic updated its usage policy to prohibit sustained, needless cruelty toward Claude: People who like to yell at Claude have been put on notice. I will neither confirm nor deny if I’m one of those.

Put It to Work

A few weeks ago, I made a video showing how to use Claude Docs and Claude Slides. OpenAI has been cooking up some cool stuff that’s very similar, so I covered that this week. They have a much better way to organize your info, called ChatGPT Space. Inside your spaces, you can have ChatGPT Pages and Slides.

Weekly Random

Sen. Mark Kelly proposed three new taxes to fund support for workers as AI changes the economy. The proposal caught my attention because of how differently each tax would work.

The money would go into an AI Horizon Fund for job training, entrepreneurship programs, and stronger unemployment benefits. It’s still a bill, not law.

The three proposed taxes are:

The usage tax could show up on our own AI bills. The user could owe it, and the AI provider would collect and remit it. Providers could choose to pay it themselves instead. This sounds like a tax on using the software.

The excess profits tax doesn’t require figuring out which profits came from replacing employees with AI, like we’ve seen from a few other proposals. This seems like an increase in business taxes as part of an AI bill, with the revenue going toward support for workers displaced by AI. As with most bills, there are pros and cons.

I think tax professionals should understand proposals like this and ideally help shape them before they become law. There’s a lot to work through in how AI usage gets measured, what providers would need to report, and how the taxes would interact. The bill is worth a look if that interests you.