What's Next For Accounting? - Issue #123

Author

Jason Staats
February 22, 2024

✅ Completing Checklists With AI

We love our checklists, but by the third time you’ve done one you’re on autopilot. Here’s a wonderfully janky AI assistant that’ll check these items off for us.

Like humans, AI assistants perform best when provided explicit instructions. We’ll instruct it on how to complete each checklist item, then create an annotated PDF workpaper to show its work:

Here’s a GPT that completes 3 tasks:

  1. Confirm the financials statements are cash basis

  2. Confirm the prior period is presented

  3. Confirm total assets equal liabilities & equity

To try it, open this GPT, and upload this example balance sheet to it.

Not rocket surgery, I know, but I want to give you an example of how building explicit instruction once can automate review tasks. We’re genuinely building software here, something that can be re-used and scaled. Here’s how it’s prompted to complete each task:

Task 1: Confirm the financials statements are cash basis

This one was pretty straightforward. I added a check in the event there were AR or AP accounts present, and told it to make the search case insensitive, since reporting systems will capitalize differently. A couple times it claimed to not know how to extract text from a PDF. I noticed on a successful attempt it used the “pymupdf”, so I added that to the prompt.

Task 2: Confirm the prior period is presented

This took some trial & error. It’s pulling all the dates, and reviewing them to see if there’s a combination of dates that look like two reporting periods. It isn’t perfect, but you don’t get many dates on a balance sheet so it’s fine for most reports. Interestingly it’s running code to extract dates, then the LLM itself looks at the dates to find two comparable dates it thinks are reasonable.

This one was the most involved. It struggled to consistently find the right figures, but when it did, I had it write concise instructions as to how to do it again.

Much of this was trial and error, and any time it threw an error I had it modify the instructions to account for handling that error. Getting it to continually improve the instruction was a big unlock, mitigating the ways it most commonly broke.

Last I told it to summarize the results of each task in a PDF, and merge a highlighted copy of the original doc with the new summary workpaper.

Is it fiddly? Yes
Will it break in ways that surprise you? Absolutely
Isn’t this better done by traditional software? If you’re a developer. But worth noting it’s running Python here, so we’re using an LLM to oversee traditional scripting. Also worth noting I’m not a developer, and most accountants aren’t.
How much does this cost? Probably $30/mo on the ChatGPT Team plan. If you’re using the API or doing it in OpenAI playground, this was coming in around 30 cents a pop.

Most importantly, this is the worst & most expensive it will ever be. Checklists of the future will be self-completing.

Related info:

💰 Niche Of The Week

Boutique Bed and Breakfast Establishments
Small to medium businesses, annual revenues $500K to $15M.
Key Pain Point: Seasonal occupancy rates and meal service cost accounting.

How to build this one from scratch:

  • Platforms - Develop partner content with the software platforms that B&Bs are engaging with.

  • Property Managers - Identify the folks being engaged to manage B&Bs, and learn how you can support their clients.

  • Communities - Add value to private communities of B&B runners. Don't sell, just be visible & helpful.

  • Newsletters - Sponsor, or develop content for newsletters about this specific type of short-term rental property.

  • Podcasts - Be a guest on relevant pods, or sponsor them to be the only accounting firm talking about B&Bs.

  • Influencers - Partner with the leading voices in the space, to add social proof to the quality of your work.

📅 This Week

😍 From The Community

My private peer networking community Realize for firm leaders, now capped at 500 members.

▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▯▯▯▯ 391/500 members