Hello fellow keepers of numbers,

The AI in accounting and finance news is heating up lately. Lots of deals between vendors and AI companies, new AI agent features, and some impressive valuations.

Ramp launched Stack as an AI system inside Ramp for accounting and finance teams to build agent skills and automate their processes. Microsoft released AI models that put them into competition with the lower-end models from Anthropic and OpenAI. The Codex app is becoming friendlier to non-developers. And Wolters Kluwer tried to make sure we don’t forget about them with a new press release.

Plus, a demo where I create a custom dashboard in Cowork to track docs received for tax returns and when returns are ready to start.

THE LATEST

Ramp published a post introducing Ramp Stack, an AI system for accounting and finance teams that it built through working sessions with more than 20 partner firms, ranging from fractional controllers and outsourced bookkeeping shops to virtual CFO practices and regional CPA firms. The company says it developed Stack against real client engagements and live reconciliations rather than in a lab.

Stack lets a firm build its own Skills, which are reusable automations, along with Coworkers, which are AI agents that run tasks, and scheduled routines. It adds formula-based workpapers that keep output traceable and auditable, firm-level permissions meant to protect a firm's process knowledge, and native access to Ramp's existing spend, AP, and banking data.

Ramp says one design partner, Specialized Accounting, now closes some client months more than 50% faster, and points to a fractional controller who ran a sales tax calculation in 20 minutes that previously took hours. Those figures come from Ramp and its design partners, not independent testing.

Ramp did not disclose pricing or general availability in the post.

Why it’s important for us:

Not part of this announcement, but Ramp also just announced a $750M raise at a $44B valuation.

Before I get to Ramp Stack, let’s take a second to appreciate the minds of those building things at Anthropic and OpenAI. Basically every software on the planet is building the same features into their products right now. Not only are they copying Anthropic and OpenAI, but those AI companies have allowed all the software companies to build on top of their infrastructure.

As for Ramp Stack itself, it’s really not that different from Claude Cowork, Claude Code or Codex under the hood. They’ve wrapped it in a clean interface inside a tool a lot of accountants already use.

You can create a set of skills specific to the accounting and finance functions. My guess is it’s just Claude Cowork behind the scenes with a little bit of Ramp “magic” built in. You’ll still have to build it around your existing processes.

Definitely not a knock. This is a great launch, and the interface could make it catch on fast. I’m very interested to see where this goes and what real users say once they’ve tested it for a few months.

Microsoft unveiled seven in-house MAI models at its Build developer conference, a family it says it trained from the ground up on its own data without using third-party models.

The lineup is led by MAI-Thinking-1, a reasoning model that Microsoft says human raters preferred over Anthropic's Claude Sonnet 4.6 in blind side-by-side tests. Microsoft framed the release as building its own frontier models rather than leaning on OpenAI, whose models have powered most of its AI products to date.

The rest of the family covers a range of tasks. One is a coding model now built into GitHub Copilot and VS Code. Others handle image generation and editing, transcription, and voice. Several also ship in cheaper, faster "Flash" versions.

The MAI models are generally available now through Microsoft Foundry, with distribution also on OpenRouter, Fireworks, and Baseten. Microsoft did not detail pricing.

Why it’s important for us:

I almost made this a trending news item, which is pretty much all you need to know about where Microsoft stands in the AI landscape right now. Also, just an awful name for their models… MAI?

Regardless, if MAI-Thinking-1 is genuinely better than Claude Sonnet 4.6, that's impressive. Sonnet 4.6 is the current Sonnet, so being preferred over it is not nothing. It wouldn't put Microsoft's best model very close to Opus or GPT-5.5, but it would mean they're finally in the ballpark.

The "preferred by human raters in blind tests" is also more useful than it sounds. Benchmarks have mostly stopped meaning anything to me, but a blind side-by-side at least tests the model in the wild, assuming the test is legit. So I'm cautiously optimistic.

The image and voice models are interesting as well, but unless either is best-in-class (which they aren’t), then it’s hard to imagine ever using them.

OpenAI announced a set of Codex features aimed at non-developers, including role-specific plugins, a new Sites tool for building shareable interactive apps, and annotations for editing generated work in place. The features extend Codex, OpenAI's coding agent, beyond software development.

Six role-specific plugins each bundle the business applications and automated skills tied to a particular type of work, so they come configured for that role out of the box. OpenAI says the plugins cover 62 business applications and 110 prebuilt skills in total.

Sites lets users build interactive websites and apps, such as dashboards or scenario planners generated from a spreadsheet, and share them across a workspace through a URL. It is launching in preview for business and enterprise customers. Annotations extend Codex's in-place editing, already used for code, to documents, spreadsheets, and slides, letting a user point to a specific part and tell Codex what to change.

Why it’s important for us:

Codex is picking up some steam lately. The feature set in the Codex app is ahead of Claude Code at the moment. The user interface feels nice, files open within the app cleanly, and the in-app browser that the AI model can control is a very nice feature. Turning the agent loose on the browser is still dangerous, so proceed with caution.

The push towards non-developers is worthy of attention. Claude Cowork has dominated this market for months, and now Codex is going after the same users with its own plugins and features similar to Anthropic’s. Each plugin bundles the apps and skills for a role.

Sites are similar to Claude Cowork’s live artifacts, but seem a little more advanced. They can be as simple as a dashboard or as complicated as a full web app. You can also share them across the team without being fully published and live for the world to see (and steal your data). Tons of potential here.

I’ve really been enjoying the Codex app, but the minor downside is the fact that it runs GPT-5.5, not Opus 4.8. Plenty of people prefer GPT-5.5, but I still find Claude much better at normal tasks like reports, writing, Excel files, and slide decks.

Wolters Kluwer announced an expanded enterprise collaboration with OpenAI to build agentic, AI-native features across its tax and accounting, healthcare, legal, and compliance lines. In tax and accounting, OpenAI models will power agentic workflows inside CCH Axcess that gather client data, classify documents, and complete preparation steps that used to require manual effort.

Wolters Kluwer says the automation has already cut manual tasks by 20% to 30%. The work runs through the company's model-agnostic GenAI platform, so OpenAI becomes a key collaborator rather than an exclusive provider, and Wolters Kluwer says it keeps control of product, data, governance, and the customer experience while pairing OpenAI's models with its own curated content and domain workflows.

Alex Tyrrell, SVP and head of advanced technology at Wolters Kluwer, said the collaboration "accelerates our Expert AI vision, bringing trusted AI to the professional workflows of clinicians, lawyers, accountants, and other experts." Wolters Kluwer did not give a specific rollout date for the new CCH Axcess features, describing them as accelerating capabilities already in place.

Why it’s important for us:

Initial reaction, is this even news? I would’ve assumed Wolters Kluwer already had a deal with OpenAI. So it’s surprising that they had to announce an “expanded collaboration.” Maybe they were using mostly Anthropic previously.

In the grand scheme of things, this is good. More capable models behind CCH Axcess (and other WK products) should make the AI inside of them better.

But I’ve seen enough of these announcements to know this feels like lip service. Agents that gather client data, classify documents, and run the prep sound great on paper. Will it actually be useful for CCH?

Seems like anyone can sign a deal with OpenAI or Anthropic these days. Whether they actually do something useful is another story. And that’s not just directed at WK. They have a lot of work to do to beat out the third-party tools focused directly on these pain points. Some of those third-party tools are already doing great things.

TRENDING NEWS

Sage added AI automation across its Intacct 2026 R2 release, including a new AI Gateway that connects financial data to outside AI tools over MCP: There’s a theme for the first several pieces of trending news. MCPs. So instead of writing the same take over and over, let’s just say that MCPs are great, all vendors should make them, and we’ll applaud the other vendors below who have announced one.

Kick launched a read-and-write MCP that lets an AI agent run bookkeeping directly inside its AI-native ledger: Yay (*applause*)

Firm360 launched a Claude connector that answers plain-language questions about firm health, like WIP, realization, and client profitability: Yay (*applause*)

Abacor launched an MCP connector that lets firm users query their Abacor data through Claude: Yay (*applause*)

HubSync launched Halo, an AI operating system that uses specialized agents to automate the full tax and accounting engagement, from intake to delivery: Well, we’re seeing this everywhere now. Ramp just announced it. Wolters Kluwer just announced it. Tons of others previously announced it. Agents are being widely deployed in nearly every software.

Anthropic launched Dynamic Workflows in Claude Code, a research preview that runs hundreds of subagents in parallel on one job: This one is for the Claude Coders. I’ve been using workflows for the last week, and it’s kind of incredible. It seems to be a much more natural way to have Claude spin up subagents to do specific tasks, review work, take actions, etc. I ran a workflow yesterday for something a normal Claude Code session would’ve done in about 5-10 minutes, and it ran for 97 minutes, and the output was fantastic. Might cover this more in a future demo.

OpenAI rolled out a more capable ChatGPT memory that uses a "dreaming" process to refresh itself in the background: Memory is one of the most interesting features in the AI space right now. It sounds useful on the surface, but there are actually a lot of downsides. Do you want your AI to automatically save a memory that was actually for one specific client/project? There’s also some discussion around how much usage memories eat up. With memories on, both ChatGPT and Claude will consistently use agents to decide when a memory should be saved, which actually chews up a lot of usage just to make those decisions on your behalf.

Google released Gemma 4 12B, a free model that runs entirely on a laptop, even offline: Feels obligatory to include an open-source model in the news once a month as a reminder that AI can also run locally. As these models become more capable, this is a serious option for firms that want to control AI costs.

Microsoft introduced Scout, an always-on agent inside Teams that takes actions on its own, like scheduling meetings: About a month ago, Microsoft told us they were trying to integrate OpenClaw features into the M365 stack. This is them making good on that promise. I’ll have to test this, but I like the idea a lot. Microsoft is doing some interesting things lately.

The White House issued an executive order asking AI labs to voluntarily submit new models for a 30-day security review before release: It’s good that regulatory bodies are finally paying attention and taking action. But I posed this same question a few weeks ago…do we really trust the government to be the experts on whether an AI model is dangerous?

PUT IT TO WORK

Tax season can make us lose our minds, turn friends into enemies, and lose all faith in humanity. How hard is it to respond to an email or send that one document we’ve been waiting on for 3 weeks?

Apparently, one of the hardest things of all time. Which is why accountants need software just to track what docs have been received and when to follow up. In this demo, I build a simple solution that tracks the status of tax return docs connected directly to a DMS.

Tax Intake Dashboard with Google Drive - Watch Video

WEEKLY RANDOM

Anthropic published a report this week saying it wants the ability to slow down, or even pause, frontier AI development.

They’re suggesting the world could benefit from a worldwide pause on AI until we’ve put the proper controls and guardrails in place. The piece outlines why we’re at a pivotal point in the development of AI right now.

Anthropic is seeing signs of AI models, particularly Mythos, performing better than humans at important tasks. Which suggests that we could be close to the point where models are good enough to train the next generation of AI. Even if that’s not the case, Anthropic argues that the models’ capabilities are good enough now that the pace of improvement will still far exceed the reaction speed of regulatory bodies and the economy.

The piece highlights some proof within Anthropic to support their claims.

It’d be easy to write this off as a marketing ploy. They just filed for IPO, they’re arguably the leaders in the clubhouse right now for AI, and they’re about to drop Mythos. Yet this still strikes me as a genuine concern.

A worldwide pause seems extremely unrealistic. Especially when there are so many geopolitical factors involved. But arguably the most capable AI lab on the planet is warning us about the consequences of continuing the pace of AI development. This feels like it should be the biggest story of the year thus far.