Hello fellow keepers of numbers,

Big week for Deloitte, both good and bad. Deloitte deployed Claude across the firm alongside a certification program shortly after they were caught citing fake sources courtesy of ChatGPT and poor quality control. Baker Tilly and Moss Adams partnered to provide AI consulting services. Maybe Deloitte can hire them.

THE LATEST

Deloitte and Claude announced the rollout of Anthropic's Claude AI assistant to more than 470,000 employees across 150 countries, marking the largest enterprise AI deployment in Anthropic's history. The partnership expands on an existing alliance first struck in July 2024 and includes significant investments in both financial resources and engineering support.

The deployment includes specialized Claude "personas" tailored for different roles, including accountants who will receive AI built for financial analysis and auditing tasks. Deloitte is establishing a Claude Center of Excellence with trained specialists who will develop implementation frameworks, share best practices, and provide ongoing technical support to scale AI pilots into production.

As part of the collaboration, Deloitte and Anthropic are co-creating a formal certification program to train 15,000 Deloitte professionals on Claude. The companies are also building AI solutions with compliance features specifically designed for regulated industries like financial services, healthcare, and public services.

Why it’s important for us:

This is the most obvious signal yet that AI tools are becoming important infrastructure in major accounting and consulting firms. It’s interesting that Deloitte has chosen to make this deal with Anthropic for a few reasons.

Claude seems to be lagging behind ChatGPT and Microsoft Copilot in deployment among accounting firms. I’ll be interested to see how this impacts how other accounting firms view Claude moving forward.

This deal likely has a lot to do with another piece of news being covered here today (spoiler: Deloitte issued a report with fake AI citations).

The “personas” tailored for different roles is also very interesting. There aren’t any specifics on this yet, but it could mean anything from custom solutions Anthropic is building for them to simple chatbots that are connected to a data source with context on specific job roles.

It’s becoming clear that it’s now the responsibility of accounting firms to provide their employees with access to state-of-the-art AI tools and a proper training program to ensure they’re adopted and utilized appropriately.

Baker Tilly and Moss Adams announced Collaborative AI, a set of AI enablement services designed specifically for middle-market businesses to safely utilize AI. The services connect employees to company data and knowledge bases and provide integration insights.

The consulting managing principal Mark Steranka emphasized that Baker Tilly's professionals guide businesses through every step from discovery and use case identification to integration and ongoing data management. With specialization across more than 30 industries, the firm delivers AI solutions tailored to each client's specific challenges and goals.

Why it’s important for us:

Baker Tilly and Moss Adams announced they’re now an AI consulting firm. They’re entering a space where there’s serious need, but I’m going to be skeptical on effectiveness for now.

They’ve targeted one of the most relevant issues in adopting AI for most businesses today: connecting data and knowledge to AI tools. I’ve seen a lot of firms struggle to implement even simple AI tools and automations because the underlying data isn’t properly structured or accessible.

The difference between messaging a chatbot and implementing a useful AI agent isn’t just the ability to build the agent. The biggest factor is usually the quality and accessibility of data to provide context to the AI agent.

It seems natural for consulting and advisory practices, like Baker Tilly, to consult and advise on AI if they’re knowledgeable. I have no doubt we’ll continue to see more enter this space.

Deloitte Australia agreed to partially refund the government after delivering a $290,000 report that contained AI-generated fabricated citations and quotes. The 237-page compliance review for the Department of Employment and Workplace Relations included references to non-existent academic papers, a fake quote from a federal court judge, and multiple invented citations.

A University of Sydney researcher flagged the errors after noticing citations to academics and books that didn't exist. The original report, published in July, was replaced with a corrected version in September that disclosed the use of OpenAI’s GPT-4o in its creation. Deloitte removed over a dozen fabricated references while maintaining that the report's core findings and recommendations remained unchanged.

Why it’s important for us:

Damn. At this point, we all know that AI occasionally hallucinates. How does Deloitte let this happen?

We don’t have a lot of details on exactly how this happened. I suspect that the sources cited weren’t make or break for the conclusion drawn in the report, so it just slipped through the cracks in review. In fact, it’s entirely possible the reviewers didn’t even know AI was used to research or create the report.

Take a stroll down theoretical lane with me: You’re a 57 year old Australian partner at Deloitte, mate. You don’t use AI. You don’t even really know what it can do. You barely know how to log into the DMS to sign off on deliverables. A 237-page paper comes across your desk (quite literally, you still work with pen and paper). It seems plausible to skim a few citations that aren’t critical to the report, so you do that. You’re completely oblivious to the fact that people on your team even used AI at all in this report. The possibility of fabricated citations didn’t even cross your mind.

I kind of feel for the partner in my theoretical scenario. But maybe my scenario is wrong and Deloitte just sucks at quality control.

As someone with direct experience creating reports like this, I have a sobering truth for some of you… The research rarely comes first in these reports. In reality, a conclusion is drawn to meet the expectations of the client, and then they find research to support that conclusion. Maybe that’s not shocking. But my guess is in their search for sources to support their conclusion, AI led them astray.

This obviously undermines trust in Deloitte’s reporting and deliverables, but it likely adds skepticism to other firms’ reporting as well. I can sympathize with a client who might receive an important report or opinion they expect to rely on and now have anxiety about the fact that it might contain fabricated information.

You think there’s correlation between this news and the partnership between Deloitte and Anthropic?

Props to Deloitte for correcting the issue and taking swift action. Training staff and partners on best practices for AI is a good step in the right direction to minimize this moving forward.

This isn’t the first time something like this has happened, and it definitely won’t be the last. It’s an important reminder why everyone within accounting firms needs to be educated on AI - both the benefits and the downfalls.

PUT IT TO WORK

You should stop nesting folders in SharePoint.

When we utilize AI, automations, and agents, AI needs easy access to your data. If your SharePoint looks like a Russian nesting doll of folders, AI is going to have a hard time finding things. As are you, probably.

I keep running into firms who want to connect AI with their SharePoint data but have very complex folder structures. When we try to connect a custom GPT, automation, or AI agent to the data (SOPs, business context, project information, etc.), we have to specify the folder. This makes jumping between folders to grab multiple files extremely difficult.

Instead, you should consider using metadata. In SharePoint, metadata just means adding columns to tag your files.

Instead of nested folders, use one client folder with all files sitting flat inside. No subfolders. Just files and metadata. Metadata fields could include things like Document Type, Year, Status, Service Line, Project Name, and more. When someone uploads a file, they tag the metadata fields.

If you’re concerned about an overwhelming amount of data in one folder, consider using views. The views can group files by metadata fields. For example, you can have a view specific to Year and Project (e.g., “2025 1040 Return”).

Under this approach, AI can navigate to your client folder and easily access any file needed. If needed, filter by certain metadata fields.

Custom GPTs, automations, and AI agents will thank you later.

WEEKLY RANDOM

If you haven’t seen yet, OpenAI announced Sora 2. This is OpenAI’s video generation model, which also outputs audio to match the video. In a blog post from CEO Sam Altman, he discusses the launch of the Sora app and notes that this is a huge moment for creativity.

OpenAI has released a new social media app (currently invite-only). The app features AI generated videos and offers the ability to “remix” videos with your own twist. Interestingly, OpenAI has also introduced a cameo feature where users can create their digital avatar and choose to allow others to use their likeness in AI generated videos.

We are firmly in the I don’t know how to feel about this territory…

It’s scary to think there are videos out there where it’s nearly impossible to distinguish between what’s real and what’s AI. Sora 2 uses watermarks to notify you that it’s AI generated. And seemingly within seconds of it’s release, there are now tons of apps that can remove that watermark…

That being said, AI generated videos can be impressive and creative. They still require a ton of planning, creativity, and time to create something truly great. Take a look at Dave’s video below.

On the bright side, more people will get to flex their creative muscles and create awesome things they might not have otherwise been able to make. On the cynical side, we end up with an AI TikTok with 10x as many dumbass dancing videos.