Hello fellow keepers of numbers,

This week was actually a relatively slow one for the AI companies for once. Maybe they’re focusing on damage control after all the chatter about how afraid of AI everyone is (if your algo hasn’t fed you this yet, congratulations on having a better algo than me).

Deloitte announced a new practice that will help companies build and deploy open models so they’re more in control of their own AI infrastructure and costs. We also got two more AI-native accounting firms backed by venture capital.

Plus, stick around for a demo of using GPT-6 Astra controlling ChatGPT’s in-app browser to find and evaluate competitors.

And happy football season to those who celebrate!

The Latest

Deloitte launched a global Open Model Engineering practice to help companies build AI systems using a mix of open and proprietary models.

Deloitte’s engineers will help clients choose the right model for each job, customize it, connect it to the company’s software, and operate the finished system. The work is intended to give clients more control over where their AI runs, what data it can access, and how much it costs to operate.

For projects involving AI agents, Deloitte can use Zora AI, its platform of prebuilt agents. Zora turns the selected AI models into agents that can perform specific business tasks and connects those agents to the client’s existing systems.

The practice will initially focus on NVIDIA’s Nemotron family of open models and serve clients in North America, Europe, and Asia Pacific. Deloitte plans to hire, train, and certify specialized engineers to work directly with clients through fiscal 2027.

Why it’s important for us:

This is the first time I can remember seeing an accounting firm build an entire service line around open models.

I think the biggest advantages come down to cost and control. Today, if you use ChatGPT or Claude, you’re basically renting the model. You pay whatever they charge, work within the limits they set, and hope neither changes too drastically. If a company runs an open model itself, it has much more control over those decisions.

With ChatGPT or Claude, we usually write instructions that say, “You’re an expert in [X] industry,” or “Here’s how our firm handles [Y] work.” Those instructions sit on top of the model. With an open model, a company can go a level deeper and better customize or “train” the model around its work, terminology, and company knowledge.

Local and open models have been around for years. Until recently, they weren’t smart enough or fast enough to be a serious option for most businesses. We’re getting to a point where that tradeoff makes a lot more sense.

And accounting is a pretty good fit. A lot of our work doesn’t require the smartest AI model in the world. It needs a model that’s reliably good at a specific type of accounting work and understands the firm’s rules. An open model might actually be able to do nearly everything the firm needs.

The Zora AI part is also interesting. From what I understand, Deloitte can help a company customize the model and then use it to power Zora’s prebuilt agents. I think of it a little like building a skill for a specific task and then letting different AI models use it.

I suspect local and open models will get a lot more attention over the next few years. Deloitte building an entire global practice around them is a pretty strong signal that companies are starting to take that option seriously.

Rational launched what it calls a “zero-human” accounting firm, while Minerva announced that it acquired an accounting firm and rebuilt its operations around AI agents. Both companies sell accounting services directly, with AI performing much of the underlying work.

Rational describes itself as an accounting service rather than software that customers operate. Businesses hire Rational for the completed work, while its agents gather source documents and complete the books behind the scenes.

Minerva is keeping accountants in the firm but using agents to handle routine accounting work. It says this allows employees to spend more time on client relationships, professional judgment, and advisory services. Minerva claims the acquired firm’s operating profit margin increased from 5% to 70% after the rebuild, though it has not published financial details supporting that figure.

Minerva is now acquiring additional accounting firms and partnering with existing owners to make their operations AI-native. Y Combinator lists Rational as an active Summer 2026 company, but Rational has not explained in detail how its human-free signoff and tax filing process will work in practice.

Why it’s important for us:

If you’re looking for a good reason to be pissed off, I’d recommend clicking the link for the Rational announcement and watching their launch video.

This is yet another week where we’re talking about accounting firms backed by startup accelerators, private equity, or other outside investors. If nothing else, it proves a point we’ve discussed a lot lately: there is significant value in this industry, but many firms and their processes are handcuffed by the outdated software they use today.

Rational and Minerva are approaching the problem differently, but they’re starting from the same idea. They want to free the firm from its old technology and rebuild the entire operation with AI native to everything it does.

Rational’s version is a “zero-human” accounting firm where agents handle everything. This seems pretty irrational to me (hehe). Agents can create a lot of value across an accounting firm, especially when the tech stack has been built properly. But they aren’t good enough to handle everything on their own. Nor should they. So the team building Rational and I fundamentally differ there.

Minerva’s model is much more interesting. They’re acquiring existing firms and rebuilding their technology and processes. They're using custom tools and agents to move the manual work away from employees so those people can spend more time working with clients and doing high-value work. I have far more confidence in that version of an AI-native firm.

Minerva also claims it increased the acquired firm’s operating profit margin from 5% to 70%. If they started at 5%, there was probably quite a bit to fix before AI even became a major profit margin contributor. But I also don't find that jump entirely impossible if they truly rebuilt the firm. The uncomfortable question is what happened to the staff. Minerva says humans aren’t going anywhere, but they haven’t shared the numbers, and I suspect layoffs played some part in that increase.

Regardless, this is what the industry is trying to figure out right now. How do we rebuild an accounting firm around AI and move people away from data entry and toward clients, judgment, and advisory work? I don’t believe Rational has the answer, and Minerva may not either. But that’s where people are focused, and it’s where they’re spending a lot of money trying to find out.

Trending News

Caseware launched Verity in Excel and a client response validation agent, claiming nine hours saved per audit engagement: I like agents that live in Excel. I think they feel more natural to use for a lot of people since they're in the tool they're already using. Surfacing data for an engagement inside of a workpaper is really useful. I'm just a bit skeptical of how well Caseware's AI works in practice. But I like where they're focusing.

OpenAI launched ChatGPT for Financial Services with GPT-6 Astra, built-in premium data, and firm-managed templates for bankers: I don't care much about the investment banking part. The important thing for us is the product shape: trusted industry data, firm templates, and citations built into ChatGPT. Accounting feels like another good vertical.

Cursor launched Projects, a persistent coordinator that maintains context, delegates to subagents, and acts on schedules, pull requests, and Slack signals: This is a project manager for your agents. No need to worry about new chats, maintaining memories, or improving your instructions. This can be managed by your Cursor Project. This is really cool, and it's partially why I was bullish on the Grok Bot release. We'll see more Grok Bot-esque features making its way into Cursor's app.

OpenAI added writing style learning to ChatGPT Work using connected Gmail, Drive, Slack, and SharePoint content: I love this. Getting AI to sound like you usually means building and maintaining another set of instructions. Learning from the places you already write is a much cleaner setup.

Grok Bot added sales connectors and installable agents, plus a Microsoft Teams integration that can search, read, and send messages: The updates are flying for Grok Bot. I'll never be mad at more connectors for AI tools, and Teams is a particularly useful one since so many firms already live there.

Meta launched Muse, a personal agent that keeps working across apps and asks for approval before sensitive actions: I haven't used Meta AI for anything serious, so I'm skeptical. But this is another sign that background agents are becoming normal consumer products.

OpenAI released ChatGPT Images 2.5 with faster generation, better reference image fidelity, and more reliable editing: ChatGPT images are my favorite already, so this is presumably just making the best even better.

Put It to Work

GPT-6 Astra is really awesome, and it’s great at using ChatGPT’s in-app browser. It also chews through usage faster than anything you’ve ever seen with your human eyes.

And you might know me at this point… Nothing gets me more excited than blowing my allotted usage on a task I didn’t even need to run. So, here you go!

I’ve demoed how to use ChatGPT to navigate to websites and grab competitors’ information to inform your own business decisions and business strategy.

Weekly Random

Ever worry about the economy? Ever wonder how you could make your anxiety about the future even worse? Well, look no further!

Anthropic built an interactive model showing how AI could affect the U.S. economy through 2030. You choose how capable AI becomes and how quickly businesses adopt it. Then it shows you the possible impact on economic growth, unemployment, wages, and the split between workers and owners.

I'd encourage you to try it whether you're an AI skeptic or wildly optimistic. Anthropic also includes three starting scenarios if you don't want the pressure of predicting an entire economy.

This is particularly interesting though because it's focused on knowledge workers. Which means this is directly related to accountants.

I'm not smart enough to know if this is even remotely close to being a well-built model. But a lot of what they wrote made sense to me as I read it.

If you already have anxiety about the future, maybe skip this one or keep that Xanax handy.