Hello fellow keepers of numbers,

Lots of interesting nuggets this week relevant to the accounting space. PwC has expanded their alliance with Anthropic, as they now see the power of Claude Cowork. Grant Thornton is the next top accounting firm to move toward automated audits. And Anthropic releases a plugin for small businesses that will likely have an impact on a lot of accountants.

Plus, OpenAI released a new feature in their mobile app that lets you connect to your computer and run tasks from your phone using your computer’s files. I’ve included instructions on how to set it up.

THE LATEST

PwC expanded its strategic alliance with Anthropic to roll out Claude, Claude Cowork, and Claude Code to U.S. teams, with the rollout extending across PwC's global workforce. The deal builds on PwC's earlier internal deployment of ChatPwC and three AI incubation pods focused on finance, supply chain, and deal-making.

Part of the expansion is a new Office of the CFO business group, a dedicated practice that uses Anthropic's full product suite to redesign client finance organizations. PwC says the group will work with regulated industries where accuracy and auditability are critical, taking on engagements that range from targeted help with specific finance tasks to "top-to-bottom redesigns of the whole function." Internally, PwC teams have been using Claude for journal entries, variance analysis, RFPs, and annual planning optimization through Claude Code.

PwC is also reinventing deal execution end-to-end, with agents working alongside deal teams across diligence, value creation, and integration.

PwC and Anthropic are launching a joint Center of Excellence to train and certify 30,000 U.S. PwC professionals on Claude. More than 5,000 PwC leaders have already gone through training at the firm's Advisory Leadership Exchange.

Why it’s important for us:

PwC's first deal with Anthropic was about using Claude inside the firm, training people on it, and selling consulting to companies interested in implementing Claude. But Claude Cowork has changed the game for a lot of businesses. Apparently, PwC is no different. I suspect that’s why they’ve made this expanded partnership.

I find the info on the Office of the CFO most interesting. Anthropic has been pushing hard on the finance space for a while now. It’s clear that the top accounting firms Anthropic has partnered with are a big reason why. Claude knows what a journal entry is and what variance analysis looks like in part because firms like PwC have been helping train it.

This is yet another data point on AI companies moving much faster than companies trying to build agents and custom apps internally. I don’t know if PwC finds ChatPwC worthwhile, but it obviously doesn’t check all the boxes since they’re pushing harder toward Cowork and Claude Code.

I’m still most bullish on vendors that integrate with Claude via a native connector or at least a public MCP. And I’m most bullish on businesses that build their agents on top of these connectors inside of agentic tools like Cowork, Claude Code, and Codex.

Grant Thornton launched gtap, short for Grant Thornton Analytics & Automation Platform, a proprietary cloud-based audit platform that embeds analytics, automation, and AI into each stage of the audit lifecycle. The firm describes it as an "AI-enabled agentic audit model" built around intelligent agents that work alongside auditors. gtap ingests data from any ERP system, runs full-population analysis, and automates routine and transactional audit activity.

The platform replaces a patchwork of legacy tools with a single environment that standardizes data ingestion, transformation, and analysis across engagements. Grant Thornton says auditors can use gtap to surface risks and anomalies in real time and produce audit-ready outputs at scale.

Grant Thornton is deploying gtap across its U.S. Audit & Assurance practice, beginning with private company audits and expanding to public company audits later in 2026. The firm did not disclose the AI vendors or models behind the platform.

Why it’s important for us:

EY rolled out agentic AI inside Canvas a few weeks ago and said they expect end-to-end audit automation by 2028. PwC has put it sooner, within calendar year 2026. Now Grant Thornton has its own platform doing essentially the same thing.

The details on gtap are still pretty minimal. But it sounds like it ingests data from an ERP and runs the audit procedures. I assume there’s still human review after the agents in gtap complete their analyses.

Audit seems to be shifting toward agentic automation a lot faster than tax. Audit has well-defined procedures, and it’s often fairly black and white. Tax often has a little more gray area and advisory, which makes it slower to automate. Tax software is also notoriously difficult to integrate, which makes even the data entry aspect harder to automate.

The direction we’re moving in audit is pretty clear now. It’s almost certainly going to lead to much more accurate financials because the agents will be able to analyze full populations instead of small samples.

The big firms have the money to invest in this custom software with built-in AI agents. But even on a smaller scale, things like Claude Cowork, Claude Code, and Codex can be used for simpler audit procedures.

Anthropic launched Claude for Small Business, a new offering aimed at small and mid-sized businesses that runs through the Claude Cowork platform. The plan ships with 15 ready-to-run agentic workflows and 15 skills covering finance, operations, sales, marketing, HR, and customer service. Specific tasks include payroll planning, monthly close, invoice chasing, margin analysis, and campaign management.

The plan integrates directly with several SMB software platforms. Anthropic named Intuit QuickBooks for payroll, monthly close, cash flow, tax prep, and reconciliation; PayPal for settlements, invoicing, disputes, and refunds; HubSpot for lead triage and campaign attribution; Canva for content generation; and Docusign for contract management. Google Workspace and Microsoft 365 are also included.

Anthropic is positioning the offering around small businesses that have lagged enterprises in AI adoption, despite accounting for roughly 44% of U.S. GDP and employing nearly half the private-sector workforce. "Small businesses make up nearly half the American economy, but they've never had the resources of bigger companies," said Daniela Amodei, Anthropic's co-founder and president.

The announcement did not include pricing details.

Why it’s important for us:

Anthropic gave this a fancy launch and a 10-city tour, but if you read the announcement closely, what they actually shipped is a plugin. The connectors already existed. The skills they publish in the plugins are free for everyone. There's no discounted software or special subscription.

Most small business owners aren’t going to figure out the right combination of connectors and skills on their own, so a pre-built set of them is actually great. And the free training is nice too. So this is still a cool announcement.

But this is less plug-and-play than it sounds. The skills were built around a specific tech stack: QuickBooks, PayPal, HubSpot, Canva, Docusign, Google or Microsoft. If you aren’t on that exact stack, there’s still custom work to get any value out of it.

Some of these skills take a direct shot at the accounting and advisory work firms are doing for clients. Payroll, monthly close, cash flow analysis, quarterly review.

But I think this is a great opportunity for firms. Small businesses will need help doing the customization and setup I mentioned. They’ll likely also need help making sense of things. Accounting firms could jump on the back of this announcement to help their clients adopt this plugin to better understand their business and operations. It’ll also likely improve the relationship with their accountant.

TRENDING NEWS

CCH Axcess launched Advisor, an AI assistant built on what Wolters Kluwer calls 'Expert AI' and trained on the firm's tax and accounting content: Wolters Kluwer hasn't done a lot well lately when it comes to integrations and AI, but this is interesting because of the number of firms using CCH. They should have a lot of high-quality content to train the AI.

Codex launched inside the ChatGPT app, bringing OpenAI's agentic AI to mobile: AI agents are upgrading and entering our pockets. Claude Code is available on mobile, but only in cloud environments or using '/remote-control'. Claude Cowork is available on mobile using Dispatch. They all have limitations right now, but it's improving quickly.

Notion launched a CLI for managing Notion content from the terminal: This is useful for the Claude Code and Notion users. The CLI should make it simpler to use Notion within Claude Code. They've also made workers available via CLI, which can be viewed as a replacement to Zapier in some cases.

Anthropic shipped Agent view in Claude Code, a UI that gives developers visibility into what agents are doing during a task: An improved UI for managing agents when using Claude Code in the CLI. This helps users avoid 5+ tabs.

Anthropic raised Claude Code usage limits by 50% through July 13th: This is another 50% on top of the doubled rate limits they provided last week. They delivered some bad news to developers using the Claude Agent SDK within their apps, so it looks like this usage was given to soften the blow. This is not applicable to Cowork.

Anthropic overtook OpenAI on Ramp's business AI adoption index, based on Ramp customer spending data: Lines up with what we've been seeing. I've seen a lot of accountants and firms gravitating towards Claude. Cowork seems to be a huge differentiator in the market.

Xero launched XeroForce, a no-code AI agent builder that lets small businesses and accountants build natural-language agents for financial workflows: I'll reserve final judgment until I see this in action. It's cool in theory, but the major accounting vendors, or really most software vendors for that matter, have yet to make custom agent builders work well.

Codex shipped a Chrome extension that runs OpenAI's coding agent directly in the browser: Codex can now use Chrome in the browser, and it also interacts with the Codex app. I'm surprised it took this long, but it's a very useful feature.

PUT IT TO WORK

If OpenAI could name everything Codex, I think they would. Codex, the model. Codex, the CLI. Codex, the application. And now Codex, inside the ChatGPT app. Let’s just assume when I say Codex here, I mean the desktop app unless I specifically say the AI model.

OpenAI released an amazing new feature in their ChatGPT mobile app this week. You can now use your ChatGPT app on your phone to work with data on your computer.

Disclaimer: This currently only works on Mac.

For those unfamiliar with Codex, it’s a desktop app that lets you do the same kind of work as Claude Cowork. You link it with folders on your computer and can run tasks completed by agents. Its complexity likely sits somewhere in between Claude Code and Claude Cowork. Not as simple as Cowork, but not as complex as Claude Code.

To get your phone connected with Codex on your computer, follow the steps below.

(1) Download Codex or update your app

Visit the OpenAI Codex site and download the Mac app. If you already have it, make sure you’ve updated and are running the most recent version.

(2) Get familiar with the Codex app

If you’ve used Claude Cowork, then this shouldn’t feel foreign.

At the bottom of the prompt box is a button to link to a specific folder on your computer.

Before we get started connecting Codex on your phone, link it to a folder on your computer, and send a test prompt (anything will do) to create an active project. You’ll see in the left sidebar of my image above that I have a Memorial Ridge project, which is the folder I’ve linked on my computer.

(3) Pair your phone

Ensure your ChatGPT app on your phone is updated as well. You should see a Codex button in your menu now. Click on the Codex button. You should see “Set up Codex mobile” with a connect button.

If you don’t see that in your ChatGPT app, open your Codex desktop app and ensure you see a menu option below “Automations” that says “Codex mobile”. Click on that. There should be a QR code to scan.

After you’ve gone through the normal login process, your Codex app should sync with your ChatGPT app on your phone. You should see something like the below with your projects from your Codex app.

You can now start new tasks from your ChatGPT app using your folders and files on your computer.

WEEKLY RANDOM

I see some complaining about Claude models not performing as well lately. Some of that can likely be related back to the increased context window.

A few months back, they increased the context window from 200k tokens to 1M tokens for Opus models. A 5x increase.

Most people thought this was amazing (myself included). But I think we’re realizing now that the models degrade much faster than the 1M token limit.

Recently, someone at Anthropic posted some practical tips about context management and your agent threads in Cowork and Claude Code.

A few highlights:

Next time Claude feels dumb, check how long your session has been running before you blame the model.