✋Welcome to The CAS Cache, a newsletter designed to help accounting firms grow their CAS offerings in five minutes or less.

Disclaimer: Some links below support my writing of this newsletter, and some give you a deal.

Issue Sponsor 😎

In a world where AI can do the generic stuff, the firms that win are the ones with a clear niche and a real voice. Full Stadium Marketing creates focused, polished campaigns that get your firm in front of the right clients. As a CAS Cache subscriber, you'll get a personalized action plan showing exactly how to stand out: get yours here.

Want to sponsor The CAS Cache? Reply to this email to get more details.

The books only appeared to be overstated by $527k

That was the first real result I have seen from an AI agent being turned loose in a QuickBooks Online file. Multiple bank accounts where the posted transactions in QBO did not come close to matching the bank account.

One colleague pointed out in the last issue that this kind of thing could create profitable clean-up work for accountants. He is probably right. But in this case, the entrepreneur cannot afford the clean-up that is now required. And, to be honest, a lot of CAS people do not exactly get excited about cleanup work anyway.

Where the agent went wrong

I went through a few months of bank reconciliations with my friend to figure out where the agent went wrong.

The biggest issue was invoice payments. It appears the agent could not properly match them, so instead of linking the payment to the existing transaction, it created new transactions coded to “Ask My Accountant”.

The next issue was even harder to pin down. It looked like some transactions had been deleted, but I could not find evidence of that in the audit log. So my guess is the agent may not have deleted anything at all. It may have simply failed to add transactions that should have been there in the first place.

Either way, the result was the same: books that looked touched by automation, but not actually understood by it.

The bigger lesson for CAS firms

That feels like the bigger lesson for CAS firms.

In the age of AI, discovery and some form of upfront assessment will matter even more. Before you price the work, before you promise a timeline, and before you assume the books are a little messy, firms are going to need a way to uncover whether they are walking into a normal cleanup or an AI-generated train wreck.

We are still a long way from full AI takeover

This also reassured me that we are still a long way from full AI takeover.

I ran into something similar in a completely different context. I tried using an AI agent to create a photorealistic layout for a glamping ground I am building. I gave it a site plan, aerial photos of the land, and images of the domes and amenities. It kept getting close, but it could not get the layout right, even with a schematic site plan.

At this point, I am probably going to pay someone on Fiverr to do something I assumed AI would be very capable of doing.

My takeaway

We need to keep experimenting with AI. We need to keep testing it in the real world. That is how we learn what it can do, where it helps, and where it still breaks down.

But if this experience taught me anything, it is that we are still far from AI replacing CAS work. For now, it is much better at creating new edge cases than eliminating the need for human judgment.

Thanks for reading, Luke Templin!