Hello fellow keepers of numbers,

Today is a big one. It’s packed with some interesting news. Gemini 3 Flash has launched and is immediately a top contender for anyone looking for quick and high quality answers. Claude in Chrome launches for all paid subscribers with some very interesting features. Two studies on AI, and the PCAOB weighs in on AI for audits. Also, stick around for tips on how to use labels to tag files in Google Drive, which can be useful for filtering in Google Drive and when using AI automations.

I’ll be taking a break to celebrate the holidays with family, so there will be no newsletter next week. Let’s all hope the AI companies take a break for just a few days as well. For those who celebrate, have a great holiday with the friends and family!

THE LATEST

Google announced Gemini 3 Flash, a new model in the Gemini 3 lineup that is tuned for “near frontier” reasoning while keeping latency and costs low. It is a multimodal model that can handle text, images, video, audio, and PDFs with a context window up to 1 million tokens. Google is positioning Gemini 3 Flash as the workhorse for agentic workflows, code assistance, and complex multimodal analysis.

Gemini 3 Flash is available in preview through the Gemini API, Google AI Studio, and Vertex AI, and it is also being rolled into consumer products like the Gemini app and AI Mode in Search. The model supports tool and function calling so it can orchestrate workflows and interact with external systems.

It’s set as a cheaper alternative to Pro for high-volume use. Pricing is listed at $0.50 per 1M input tokens and $3 per 1M output tokens, with discounted rates for cached input tokens via context caching.

Why it’s important for us:

This is the fast version of Gemini 3. Previously, they released Pro, which is equivalent to “Thinking” in ChatGPT and Opus in Claude. Gemini 3 Pro has received rave reviews by nearly all who use it. After using it myself for a few weeks, I think it’s a good model for chat, but a great model for coding and agentic workflows.

After using Gemini 3 Flash for a few hours, it’s clear that it competes with the best models in the game today. It’s smart and fast. I also tested it by linking it with documents in my Google Drive, and it returned accurate information very quickly.

This is easily the cheapest of the state of the art models now, which makes it an extremely enticing model for automations and agentic workflows. ChatGPT and Claude are fantastic at tool calling. In my opinion, that’s where Gemini has lagged behind over the last year. I’ll be interested to test further to see if tool calling has been improved.

Take Google very seriously. Google products are far more automation-friendly than Microsoft, and Gemini is exponentially further along than Copilot. The gap between Copilot and Gemini is honestly not close at the moment. And there’s no reason to think the gap will close any time soon. Firms open to reviewing their platform providers, or any new firms, should strongly consider Google.

Google Research, Google DeepMind, and MIT released a study testing when multiple AI agents working together outperform a single AI agent. The team ran 180 experiments across four benchmarks, including financial analysis tasks. They tested five architectures: single agent, independent agents, centralized systems with a coordinator, decentralized peer-to-peer systems, and hybrid approaches.

For financial tasks that split into separate pieces, centralized multi-agent systems improved accuracy by 81% compared to single agents. The setup used a coordinator managing specialized workers analyzing revenue trends, cost structures, and market data simultaneously. Independent agents showed 17.2x error amplification through unchecked mistake propagation, while centralized coordination contained errors to 4.4x through validation checkpoints.

Performance varied dramatically by task type. Sequential reasoning tasks requiring step-by-step logic degraded 39-70% under all multi-agent configurations. The study found a capability threshold where single agents solving more than 45% of tasks see negative returns from adding coordination. The team built a predictive model using efficiency and overhead metrics that correctly identified optimal architectures for 87% of held-out configurations.

Why it’s important for us:

This is one of my favorite studies I’ve seen so far because I think it perfectly explains why so many companies have failed to successfully implement AI or AI agents. We should consider two examples.

1) Single agent performs a task and passes to a separate agent

This is the equivalent of prompting ChatGPT with a difficult task and getting an output, then passing that output to another chat in ChatGPT to perform a new task using the output from the first, and so on.

Let’s assume ChatGPT compiles information that’s 95% accurate in the first task. It passes the output to another chat that then completes its task with 95% accuracy. Now the output of the second chat is 90.25% accurate (95% x 95%). This compounds with each task.

2) Multi-agent system performs a subset of tasks

This is a system where one agent orchestrates the completion of tasks by other agents. You can think of this like a manager who delegates workpaper prep to 3 different staff. Each staff is in charge of one workpaper. The manager will review each workpaper. If there are issues, the manager will ask the staff to correct the workpaper. Once the 3 workpapers are accurate, the manager may then use them to prepare a deliverable.

In this system, the staff are reviewed based on a specific set of criteria. Before proceeding, the output from each staff is reviewed. This means the output of the staff is near 100% accurate since there was expert-level review.

Ultimately, the manager’s involvement leads to a more accurate end product. This is no different than how a multi-agent system functions. The manager is the orchestrator agent who reviews products of sub-agents and ensures the output is accurate before proceeding.

Now, what’s extremely interesting is financial tasks appear to be far more successful than other tasks conducted in a multi-agent system. I suspect this is because financial tasks are very procedural. First step one, then step two, next step three. Whereas, other tasks might be much more subjective. This results in an orchestration agent being much more effective than other tasks since it has a specific set of rules and requirements for sub-agents before it proceeds.

This is a perfect example of why I think agentic systems are the next major step for AI. Claude Code and AI IDEs will provide firms with agentic systems capable of orchestrating sub-agents and completing very valuable tasks and projects.

Anthropic opened Claude in Chrome to all paid plan subscribers on December 18, ending months of limited testing. The browser extension is now available to Pro, Team, and Enterprise users after gradually rolling out to Max plan subscribers in November. The expansion includes integration with Claude Code and new admin controls for organizations.

Claude in Chrome sits in a side panel and handles browser tasks like filling forms, clicking buttons, and copying data between systems. The extension can manage multiple tabs, work in the background, and send notifications when it needs input or finishes a job. A new workflow recording feature lets users demonstrate steps once and Claude learns to repeat them. Scheduled tasks run automatically on daily, weekly, or monthly schedules.

The December update added Claude Code integration, connecting Anthropic's command-line coding tool with the browser extension. Claude Code builds code in the terminal while the Chrome extension tests it in the browser and reads console errors for debugging.

Why it’s important for us:

Anthropic is the best in the game right now, and I really don’t think it’s that close. The only area they’re lacking is images. Claude in Chrome is another huge announcement. The best AI tools right now are the ones we can layer on top of the things we’re already doing.

That’s exactly what Claude in Chrome is. You’re probably already working in Chrome, or another browser built on Chromium (where you can use Chrome extensions like Claude in Chrome).

The use cases are incredibly valuable, especially for accounting firms. This feels like it’s going to replace a lot of RPA use cases. Because so much existing accounting software has limited API access, poorly built APIs, or annoyingly expensive APIs, a tool like Claude in Chrome that can take instructions and go perform browser-based actions is going to be incredibly useful.

The biggest question is security. While it’s not exactly the same as the AI browsers like Comet and Atlas, there are still similar questions. Personally, I don’t want to be the first to adopt this. But I’ll definitely be the second once I’ve determined it went well for the first.

Side note: I know there are other accountants who are big supporters of Claude, but the large majority use Copilot or ChatGPT, understandably. But it’s becoming criminal how underrated Anthropic’s full suite of products is right now in the accounting industry. Between Claude’s chatbot, Claude Code, Claude for Excel, and Claude in Chrome, there’s pretty much no use case that could go uncovered.

Gallup’s latest workforce survey found that 45% of U.S. employees used AI at work at least a few times in the past year in Q3 2025, up from 40% in Q2. Frequent use, defined as a few times a week or more, rose from 19% to 23%, and daily use increased from 8% to 10% over the same period.

AI use is concentrated in knowledge-based roles. In technology and information systems, 76% of employees said they used AI at least a few times a year, compared with 58% in finance and 57% in professional services, putting those sectors among the heaviest users.

Most employees who use AI at work lean on it for basic knowledge tasks rather than deep automation. Among AI users, more than 40% said they use it to consolidate information or generate ideas, and 36% use it to learn new things.

Organizational AI adoption and communication are lagging behind employee behavior. In Q3 2025, only 37% of employees said their organization had implemented AI tools, 40% said it had not, and 23% said they did not know, which suggests many people are experimenting with AI on their own without clear direction.

Why it’s important for us:

Frequent use of AI (a few times a week) is still below 25% of U.S. employees. This somehow surprises me but also seems plausible. It’s surprising because so many firms now seem to be providing basic licenses to ChatGPT, Copilot, or Gemini. And there are tons of employees who use AI tools that aren’t provided by their employer. Maybe the employees are lying in some cases?

On that note, 40% of employees said their employer has not implemented AI tools and 23% said they weren’t sure. Which means up to 63% of employees don’t have access to AI tools. This is wild to me. Please provide at least one licensed AI model to your employees. It’s important, no matter whether it’s ChatGPT, Copilot, Claude, or Gemini.

Of those who are actually using AI, it seems like a large majority are using it for only very basic use cases. 2025 was supposed to be the year of AI agents. But ~60% of employers still aren’t providing access to AI models. And for the remainder, employees aren’t even utilizing it nearly to its full potential.

Train your employees on AI tools. Basic training on using chatbots, how to create custom GPTs, custom projects, and how to create agents. All are important, but you must start at the foundations. I’ve personally seen how successful training can be across teams, even if for nothing other than brainstorming use cases.

PCAOB leaders signaled at the AICPA Conference on Current SEC and PCAOB Developments that inspectors will scrutinize AI adoption and private equity ownership during 2026 quality control inspections. Acting Chair George Botic acknowledged AI could transform audits by enhancing risk assessment and evidence gathering, but cited research showing AI dependence could erode critical thinking, skepticism, and professional judgment. A recent MIT study found AI usage poses risks to critical thinking and learning.

On private equity, Botic said PE capital can fund succession planning, technology investments, and growth, but warned that firms' focus on returns and eventual exit could shift incentives toward profitability over audit quality. He cited an Accountancy Europe study that warned increased profitability pressure could lead to cost-cutting, aggressive fee negotiations, and rapid expansion that strains staff. Botic also noted that PE-driven consolidation could reduce the number of firms performing public company audits, concentrating market power and leaving smaller companies with fewer auditing options.

Christine Gunia, director of the PCAOB's Division of Registration and Inspections, said PE has poured billions into accounting firms, with over 90 significant transactions since 2020 and more than half occurring in 2025. She warned that PE threatens to increase pressure on profitability, potentially leading to reduced staffing, fewer specialists, and threats to auditor independence. Inspectors will focus on AI governance and PE ownership structures during quality control inspections, she said, emphasizing that while AI has potential, the human element in auditing cannot be removed.

Why it’s important for us:

What a time to be alive in our industry, huh? There are a few interesting nuggets to break down here.

1) AI in audits

Just recently, PwC announced their belief that audits will be entirely conducted by AI and automation within the next year. I don’t even remotely agree with that view. But it’s clear that this is something important to discuss within the industry right now.

AI can be extremely useful for sampling, testing, reviewing workpapers, and tons of other great use cases for auditors. But at the same time, it can make future auditors dumber by overreliance on AI, right? Ultimately, it’s still the responsibility of the firm to properly train their staff on audit methodologies, critical thinking, and professional judgment. And it’ll be even more important for firms to train staff how to properly use AI that intersects with each of those fundamentals.

I suspect over the next 3-5 years (maybe sooner), we’ll see firms pop up offering an entirely AI-driven audit. But how much trust will we (accountants) and the markets place on a fully AI-driven audit? Time will tell I suppose.

2) PE in accounting

This one is a bit of a slippery slope. Personally, I think the PE investments in the accounting industry are going to do a lot of good in driving new tech innovations and adopting better processes and technologies within firms.

Maybe I’m jaded by my slightly younger age, but it feels like the accounting industry, and specifically the tech within it, needed a shot of adrenaline to get us moving forward more quickly. So many accountants still use software that was made in the 1980s, and it still looks the same too.

On the other hand, PE investments are going to create independence headaches. If PE firms own a material portion of the industry, there’s going to be a lot of overlap between audits done by accounting firms backed by a PE company that has equity in the company being audited. That was a mess of a sentence so here’s an example:

PE Firm holds a 70% stake in Accounting Firm.

Accounting Firm audits Company A.

PE Firm holds 10% equity in Company A.

Accounting Firm is auditing a company in which its investor (PE Firm) holds a significant interest.

Is this bad for the industry if there’s too much PE investment? I don’t know the answer to that, and I won’t pretend to be an expert on the subject. What I do know is that I’m excited to see how this pushes forward the investment in technology across the accounting industry.

PUT IT TO WORK

In a previous newsletter, I covered how to add metadata to files in SharePoint. I want to cover the equivalent in Google Drive.

As I mentioned for SharePoint, it’s important not to have a complex folder structure. If your SharePoint or Google Drive requires you to click through 7 folders to get to a file, it’s not set up well for AI or automation. We’ll try to resolve this now. The image below is what your Google Drive might look like after adding labels.

Google Drive allows you to add labels to your files. To create labels that are available in your Google Drive, you have to go into your Admin Console. Once logged in, you’ll click on Security > Access and data control > Label manager.

In this example, I have two labels. One is for document type and the other is to mark a file as shared externally.

To create the Doc Type label, click “+ New label” to create a label. Name the label “Doc Type” and check the box next to “Drive and Docs” under Applications. Next, click Add Fields and select the plus icon next to the “Badge list” label. Click the “Add fields” button. Add all your relevant document type options and color code them however your heart desires. Once you’re done, click the “Publish” button.

The badge list label allows you to select from a number of options (e.g., document types) and shows that specific document type next to your file name in your Google Drive. These are the colored labels in my example above.

Next, we’ll add a Shared label to mark when we’ve shared a file with a client, which will also enable us to filter for these files. Again, click “+ New label”, name it “Shared”, and check the box next to “Drive and Docs” under Applications. We’re now done with this label since we don’t need dropdown options for this type of label. Click the “Publish” button.

Now in your Google Drive, you can right click a file and you should see a Labels field. You can apply your labels from here. It also might pop open a right side panel that allows you to select your labels for that specific file.

We can now filter folders by the labels. To do this, click on your folder name at the top of the Google Drive screen, hover over “Folder information”, and then click on “Search within [your folder name]”. You should now see a Labels option that allows you to filter using your new labels.

You can continue to add new labels or update the options for your document types, as needed.

WEEKLY RANDOM

Cursor announced a new Visual Editor inside the Cursor Browser that lets users change a live web app’s interface and have AI write the underlying code. The tool brings your running web app, codebase, and visual editing tools into one window so you can adjust layouts and styles without constantly switching between an IDE, browser, and dev tools.

This is probably going to go under the radar, but it provides the ability to easily customize any website, portal, or dashboard. Instead of coding a website, or using a no code builder like Wordpress, firms can now use something like Cursor to update the website, portal, or dashboard using AI and by dragging and dropping elements on the page.

This might provide a lot of customization for anything a firm wants to build if they’re willing to invest the time and effort. Dashboards that report time and expense by project. Portals that allow firms to collect documents from clients and prospects. Websites that include lead magnets. All of these just got a lot more possible.

Until someone packages it in a simpler interface, the future seems to be in AI IDEs. Whether its agentic workflows, custom applications, or custom websites, the barrier has been removed.