Hello fellow keepers of numbers,

It’s conference season, and I write this as I’m waiting on my delayed flight. So that’s fun. Wasn’t Elon supposed to build underground tubes for bullet trains? What happened to that?

Unlike Elon, Black Ore follows through on their promises. They’ve just released the tool for broader availability. The waitlist is huge, and I imagine it’ll continue to grow. Y Combinator, the prestigious startup accelerator, has come to the conclusion that accounting firms are valuable. And Karbon debuted their Aider integration.

Plus, stick around for a demo of the new workspace agents in ChatGPT.

THE LATEST

Black Ore launched Tax Autopilot for broad availability, opening access beyond the 75 firms it selectively onboarded during its two-year early access program. CPA Practice Advisor reported that the waitlist included more than 3,700 firms, including 40% of the Top 20 CPA firms.

Tax Autopilot automates the full lifecycle of complex return preparation, including document ingestion, data extraction, federal and state tax logic, workpaper generation, and delivery into major tax software platforms. The system accepts W-2s, 1099s, K-1s, brokerage statements, bank records, and other source documents in multiple formats.

Black Ore says participating firms saw more than 99% accuracy across tens of thousands of returns, more than 98% autonomy, up to 98% time savings per preparation, and up to 80% lower costs per return. The platform shows its work, links each data point back to source documents, and flags items that require review.

Tax Autopilot is SOC 2 Type II certified and operates in encrypted environments. Firm professionals remain the final authority for sign-off.

Why it’s important for us:

The biggest question with AI tax prep is quickly becoming less about whether the software can extract W-2s and 1099s, and more about whether it can actually become the place a tax team wants to work.

From the outside, the “AI prepares your tax return” idea sounds exciting, but I think the market is trending elsewhere. A lot of tools are going to be able to read documents, classify them, extract data, and push numbers into tax software. That’s still valuable because not many people want to type numbers into a software built in 1987. But the real differentiator will be the workflow tools built around the prep.

Can a firm see what’s moving, what’s stuck, what needs review, what the AI touched, and where the support came from without bouncing between a portal, Excel tracker, review note, email thread, and the tax software? Because if not, then we’re just adding another piece of software to a stack that already has too many pieces of software.

The next question is review. If the AI prepares the return and then punts a pile of weird issues to a human, we’ve made progress, but we haven’t changed the job as much as the marketing implies. If it can do a real first review, spot problems, explain what happened, and help correct issues inside the same workflow, that’s way more useful.

I’m skeptical on the messy stuff. 1065s can be extremely complex, and there’s often large variability between one 1065 and another. Complicated K-1s and K-3s are tripping up most of the AI tools right now, and they’re all actively building and adjusting their software to get them right with the accuracy needed for a firm to adopt the tool. Black Ore is likely no different.

Y Combinator (YC) published its Summer 2026 Requests for Startups, including a call for AI-native service companies in accounting, tax, audit, compliance, insurance brokerage, and healthcare administration.

The thesis is that services spending is much larger than software spending, and many service categories are already outsourced, which makes them easier to replace with AI-native products. YC framed the opportunity as companies that sell the completed service rather than software that helps humans do the work.

The accounting angle sits inside YC's broader view that AI has moved from a product feature to a foundation for rebuilding entire categories.

Why it’s important for us:

This was an unexpected announcement. YC is one of the most prestigious investor groups. Historically, they’ve invested almost exclusively in software companies (SaaS). It’s not often you think of a traditional accounting firm as a “startup.”

They’ve specifically called out accounting, tax, and audit as interesting AI-native service companies. There are a few important signals here.

(1) The “AI will replace all accountants” crowd on Twitter is pretty dumb. But we already knew that. The accounting firm model will definitely change. The workflows and services will look quite different five years from now. But service-based businesses obviously have a lot of value in the future, as told by YC.

(2) This obviously begs the question: what does an AI-native accounting firm look like? Not every firm needs to become some venture-backed, billion-dollar empire. Most shouldn’t, and most of us don’t want that. But every firm should be thinking about the path from “some of us use Copilot” to “AI is embedded into how work moves through the firm.”

The path probably won’t be a straight line because a lot of things will change over the next several years. But regardless of what the Twitter goblins say, we’re in a pretty good industry.

Karbon released its rebuilt Aider period close experience for all U.S. customers, turning Aider's close workflow into a more direct part of Karbon's practice management system. The update gives firms a portfolio-level view of client close status and connects Aider close checklists to Karbon work items.

The checklists are built around the usual close review problems: unreconciled transactions, data quality issues, outstanding invoices and bills, and P&L and balance sheet variances. Karbon says teams can create templates by client type or service type, add custom checks, and push certain QuickBooks changes from Aider.

When a period is marked as closed in Aider, Karbon can update the related work item status and trigger configured automations. The rebuilt period close experience is available now for U.S. Karbon customers.

Why it’s important for us:

Karbon acquired Aider back in September 2025, and this is the culmination of the work they’ve put in integrating the two.

The interesting part is that it now connects more directly into one of the major practice management systems accountants are already using. Close status, checklist progress, outstanding alerts, and Karbon work items can all talk to each other.

I do think we should leave room for early hiccups. This is newly integrated, and software has a funny way of showing unexpected issues no matter how much you test it ahead of time. Aider was one of the leaders in this category before Karbon acquired them, so the expectations are high.

Aider still appears to be a separate paid product, starting at $10 per client per month with a $300 monthly minimum and a 12-month commitment. So this probably isn't just "Karbon versus everyone else." It's "Karbon plus Aider versus Double and the other close automation tools."

That still puts Karbon at the top of the AI close automation conversation. It just means the buying decision is a little more complicated than the announcement might make it sound.

TRENDING NEWS

Grant Thornton tied US partner bonuses to AI adoption goals: This comes shortly after the PwC ultimatum for their partners. It’s unclear to me how they’re going to determine what that looks like and if it’s the same for all partners. 

OpenAI renegotiated its Microsoft partnership: The revenue share now has a hard cap on the amount and timeline. It’s also opened the door to other cloud providers, such as AWS. We’re just watching a divorce in slow-mo.

Gemini added file generation inside chat, including PDFs, Word documents, Excel files, Google Docs, Sheets, and more: Google has been incredibly slow to get here given how strong Gemini has been for the last few years. But this is a major step in the right direction, and opens the door for their version of Cowork.

Anthropic lined up as much as $40 billion in Google investment and compute support, including $10 billion now and another $30 billion tied to performance targets: This is a massive announcement because Anthropic was really becoming strained for compute with their exponential growth in recent months. They now have the capacity to grow into their projected growth.

OpenAI missed internal revenue and ChatGPT user-growth targets: This is almost certainly because Claude ate up so many of their potential customers in the last 4-6 months. There’s some risk here because they’ve already committed around half a trillion dollars (!) to data centers and compute over the next several years. But they’re probably not all that concerned behind the scenes. We’ve seen that they can raise hundreds of billions at will…

Meta was blocked by China on their $2B acquisition of Manus, the AI agent startup with Chinese roots: The geopolitical tensions are rising, and AI agents appear to be major assets. The blows just keep coming for Zuck.

Anthropic may regain federal access as the White House drafts guidance that would let agencies work around the company's supply chain risk designation and use Claude models again: Turns out the government liked Claude a little more than Trump and his cabinet anticipated.

DeepSeek launched V4 Preview with open weights, a 1M-token context window, and Pro and Flash models available through chat and API: This was a bit of a lackluster launch for something that was really highly anticipated. It seems they’ve dropped back a bit in the race, but they’re still among the best open-source options.

PUT IT TO WORK

I do a lot of demos of Claude, so I figured it was time I went back to a demo for ChatGPT. They released workspace agents last week, and I really like how simple it is to build one.

Below is a demo of an agent that runs for month-end close, using the same demo data as last week.

ChatGPT Workspace Agent: Month-End Close - Watch Video

WEEKLY RANDOM

Mayo Clinic published a study on an AI model that can help detect pancreatic cancer on routine abdominal CT scans up to three years before diagnosis. The model identified 73% of prediagnostic cancers at a median of about 16 months before diagnosis, nearly double the detection rate of specialists reviewing the same scans without AI.

Pancreatic cancer is one of the deadliest. It turns out even the most experienced humans read scans as normal when AI can detect early signs of the disease.

We need more of this news in the world.

So many people right now are anti-AI. I understand it to some extent. But we all just hear the stories about the bleak future of a jobless economy run by a bunch of robots.

AI companies need better PR teams. This is the type of news that could change the sentiment.

I’m seeing more of these stories pop up over the last few months. Maybe we’re only a year or two out from AI that can positively impact society at large. Hopefully there’ll soon be a world where diseases are caught long before they’re fatal, and we end up with new cures for things that were previously untreatable.