Hello fellow keepers of numbers,

Big week for accounting tech. Karbon acquired Aider and continues to be an industry leader in AI. OpenAI dropped Agent Builder in the hopes of capturing some of the AI agent marketshare. Anthropic dropped Claude Sonnet 4.5, which is probably the most slept on AI model by accounting firms. And Perplexity made their AI browser free, which was immediately flagged for security issues. Let’s break down what actually matters.

THE LATEST

Karbon acquired Aider, an AI-powered bookkeeping and advisory platform, adding automated period close and client reporting capabilities to its practice management software. Aider automates reconciliations, uncategorized transaction cleanup, and generates management reports with real-time insights and AI commentary.

The acquisition deepens an existing integration between the two platforms. Karbon customers will get automated period close with task coordination, AI-generated reports with KPIs and executive commentary, an AI advisory assistant for data analytics and meeting prep, and firm-wide dashboards that surface advisory opportunities. AI agents that execute tasks are reportedly coming soon.

Why it’s important for us:

Karbon has been heavily leaning into AI and automation recently. I salute their ambition, and they are certainly one of the most intriguing software companies in the industry to watch.

Interestingly, I see most firms these days diversifying their tech stack as more tools begin to “major” in one area. Firms are often finding value in utilizing different tools that do one thing well and integrate with the rest of their tech stack.

But it’s hard to ignore this acquisition. Most of Karbon’s competitors offer either practice management or advisory and automation tools. Bundling both into a single platform could be a major differentiator. I suspect Karbon might lock some of the Aider features behind a paywall. But regardless, this is great news for the industry and, especially, for existing Karbon users.

Perplexity dropped the paywall on Comet, their AI-powered web browser, making it free for everyone after initially charging $200/month when it launched in July . The browser includes a built-in AI assistant that can summarize web pages, draft emails, manage tabs, and automate research tasks.

The AI browser space is heating up. OpenAI is reportedly planning to launch their own browser, Google just integrated Gemini deeper into Chrome, and Anthropic has been working on browser-based AI agents. The Browser Company’s Dia was recently acquired by Atlassian for enterprise use, and several other AI-first browsers are emerging.

Security researchers have flagged vulnerabilities in Comet. LayerX found a vulnerability called "CometJacking" where malicious URLs can exploit Comet through prompt injection attacks, potentially exposing sensitive data from connected apps. Perplexity has noted that this prompt injection attack does not lead to any impact.

Why it’s important for us:

Picture this: Client emails asking about a tax strategy you need to research. You tell the browser to read the email, research the question, and write a report with citations. When it's done, another agent drafts your email response. Another logs the research in your DMS. Another preps a draft invoice for the consulting time. Your actual work might be 30 minutes to review and approve everything.

That's the optimistic vision. And it's (possibly?) not that far off.

But we're not there yet. AI browsers still have trouble when you ask them to do things in actual applications. They'll click the wrong button, miss a field, or just give up halfway through. I've tested a few of these and the success rate isn't good enough for client work.

And the obvious problem is accountants guard client data like it's nuclear codes. A browser with known security vulnerabilities that can access your entire email and calendar creates a lot of concern. While Perplexity’s Comet browser is SOC 2 Type II certified, it’s still difficult to trust when dealing with tax returns, financials, and confidential client info.

Pass on it for now. But don't forget about it or the other AI browsers. These tools will get better and more secure over the next year. When they do, the individuals and firms that already understand how to use them will have a real advantage.

Anthropic announced the release of Claude Sonnet 4.5. The model has the ability to operate autonomously for over 30 hours on complex, multi-step tasks. Pricing remains unchanged at $3 per million input tokens and $15 per million output tokens.

Current Claude enterprise users are reporting good results with Sonnet 4.5. A financial firm is using it for investment-grade financial analysis, while a security firm is reporting 44% faster processing times for operational tasks. Novo Nordisk built an AI-powered documentation platform using Claude that reduces the time to write regulatory-grade documentation by 90%.

The model is available through Claude’s web interface, API, and integrations with platforms like Amazon Bedrock and Google Vertex AI.

Why it’s important for us:

Claude AI models are probably the most underestimated by accounting firms right now. ChatGPT seems to be the most widely adopted with Microsoft Copilot a close second.

It’s worth noting that Microsoft recently announced they now provide access to Claude models in Copilot Studio. Even the largest investor of OpenAI knows Claude is awesome.

Claude continues to ship state-of-the-art models that outperform others in things that matter for work: writing, communication style, coding, and the hybrid approach where it automatically switches between quick responses and deeper thinking when needed. If you’ve been frustrated with ChatGPT’s new automatic model switching, now might be the time to try Claude Sonnet 4.5.

Ultimately, writing and communication style come down to personal preference. I find Claude more direct and conversational than other models, which tend to be extremely verbose and packed with useless jargon.

OpenAI launched AgentKit at its recent DevDay conference, a new platform that lets users build and deploy AI agents using a visual, drag-and-drop interface. CEO Sam Altman positioned it as a way to take agents "from prototype to production" without juggling multiple tools.

Agent Builder provides a visual canvas where users drag and drop nodes to create agent workflows. Users can configure agents with custom instructions, connect them to files and knowledge bases through File Search, and set up guardrails to filter sensitive information like personally identifiable data from inputs and outputs. The platform supports preview runs and versioning for testing workflows before deployment.

Model Context Protocol (MCP) servers enable agents to connect to data sources like Google Drive, SharePoint, and third-party systems. Agents can use these MCP connections to access information and perform actions across different tools.

ChatKit is an embeddable chat interface that allows developers to deploy these agents as widgets on websites or within applications. OpenAI has also launched a Widget Builder with pre-designed widgets available for users to leverage.

Agent Builder is currently in beta with gradual rollout to users. ChatKit and evaluation features are generally available. Pricing is included in standard OpenAI API rates. Early adopters including Klarna and HubSpot have deployed agents using the platform for customer support and sales automation.

Why it’s important for us:

This is big news for the automation space, but let’s temper expectations in the near-term. Social media would have you believe this just immediately killed all automation software. The reality is there are currently a few limited use cases where Agent Builder provides value.

The most obvious use case for Agent Builder comes as a simplified interface for public chatbots. You can drag an Agent tool onto the canvas, provide custom instructions, and link documentation or knowledge with ‘File search’. The agent can then be deployed to a website with a widget via ChatKit.

OpenAI’s AgentKit has potential and will likely improve over time, but there are some significant limitations when compared to other automation tools.

There are very few native software integrations currently available in Agent Builder. Despite the simple drag-and-drop interface, you’ll need to be semi-technical to build anything that connects to third-party software not currently built into the platform.

The biggest limitation is deployment. Unlike other automation tools, there are no triggers for workflows. An automation to add a new contact to a CRM might trigger when you receive an email. In AgentKit, the only way to deploy your agent without writing custom code is through a chatbot. Your agent runs when someone sends it a message.

This is still one to watch over the coming months as new features become available and more software is added as native integrations.

PUT IT TO WORK

Below is a link to a Loom where I walk through best practices for instructions in Projects in ChatGPT, the new share feature that allows you to share your projects across the company, and an example of how you can utilize a Project to create a proposal from meeting transcripts.

ChatGPT Projects: Best Practices & Meeting Transcript Example - Watch Video

WEEKLY RANDOM

Perplexity recently dropped their Search API and my mind has been racing with use cases.

Most search APIs use a single search term, scrape Google, and return a pile of links with raw HTML data.

Perplexity’s Search API understands what you’re asking for, figures out the right queries to run, scrapes a massive amount of web pages, and returns clean data with the relevant snippets already pulled out for you.

This is going to significantly change how automations, applications, and API integrations pull information from the web. For example, consider a workflow that scrapes IRS guidance and court cases to find answers on a tax-related question.

Old way: Send “IRS guidance on deductions for purchasing an EV” to a web scraper. Get back 5 web pages with hundreds of links and HTML that looks like this:

Disgusting.

Perplexity way: Send “A client wants to deduct a new Tesla Model 3 purchase this year. Their estimated taxable income is $450k. What guidance applies?” The API figures out what to search for, scrapes the relevant IRS pages, and hands you back the actual info you need in a clean format.

$5 per 1,000 requests. The pricing is refreshing in this age of AI. Half a cent per request, no matter how complex your search request.

This could be used for tax research workflows, client intel, industry research, and regulatory updates. And we’re barely scratching the surface on use cases.

I suspect we’ll even see existing software companies build this into their products and provide us new, more advanced features. Regardless, I’m excited to start building this into my own workflows.