How AI Will Impact Public Accounting in 2024
My predictions for the year - I'll embed citations for a deeper dive on any of the below topics. This is a summary of today's podcast, where I dig further into some additional predictions .
1. Better Generative Email
This was my top prediction of last year, and we now have a basic version of it in many practice management systems. But a more intelligent email assistant is on the way.
Here's a discussion of how this works from December of 2022:
"Can you send me a copy of my 2022 tax return"
?? A draft reply is generated, with the prior year tax return attached.
"What's the status of our December financials? The banker wants a copy"
?? A draft reply is generated, using the status from project in the PM, matching the tone of voice you use with this client.
"Why was profit down last month?"
?? A draft reply is generated, pulling in data from the PM's API connection to the accounting ledger, and a transcript from a past meeting with the client discussing the same subject.
The technical:
Generative email solutions today use basic prompt engineering to generate replies. Future versions will leverage retrieval augmented generation to fetch additional context from file system, projects, transcripts, even third party APIs.
Important to perceived quality is whether the email sounds true to the user. A better version of this can be achieved by caching a tone of voice prompt for each user, generated according to past emails, and a second tone of voice prompt for each user, by client. Layering these two instructions together will generate a tone true to interactions with that specific client.
How it impacts the tech landscape:
The quality of generated emails will only be as good as the information the system has access to. This will favor all-in-one platforms, and penalize systems that don't incorporate email. A user's familiarity and comfort level with a PM system's email assistant will become sticky aspects of the product.
2. Better Research Tools
The product category most ripe for disruption, research solutions, will vastly improve in the near term. The ability of LLMs to draw relevant excerpts from authoritative literature, and present it in a structured workpaper in seconds will eliminate the bulk of research tasks currently assigned to junior staff. We aren't using LLMs as a knowledgebase here - we're layering LLMs over authoritative sources.
Here's a discussion of the right & wrong ways to use AI for research. The underlying tech will kick off a new category of research app:
The technical:
Where last year we may have thought context limits were the blocker, we now know retrievals, and a model's ability to stick to specific context drives quality responses. We're beginning to see purpose-built, retrieval-optimized models, and could ever get a future retrieval-optimized model from OpenAI, akin to chat & completion models. Conditioning the underlying research materials for semantic significance, ie tables & flowcharts, has been a challenge historically but is now simplified with vision models.
领英推荐
How it impacts the tech landscape:
This is the product category that will look most different in 24 months. Chief Justice John Roberts last week outlined how AI will fundamentally change legal research, and the judicial process. Getting to answers will be quicker & easier than ever before. Incumbent research providers have the most rich data set to draw from, but may not be the first to move on a meaningful implementation.
3. The Dawn of Agents
Agents, or autonomous AI assistants that can navigate the programs we use could make the AI developments to-date look like child's play. Agents follow a prompt from the user, and navigate software the same way a human user does to complete tasks. They're fairly slow & resource intensive today, but can complete basic tasks, like printing a cash basis prior year balance sheet and P&L from QBO.
Progress in this domain, for example, an agent that can complete a QBO reconciliation & markup the PDF statement, will bring about significant tipping points in automating large chunks of work.
The technical:
Models purpose-built for agentic interactions will be a large unlock. Some believe the next family of models from OpenAI will include an agent model, akin to chat and completion models, better optimized for agent-specific tasks. Augmenting agents with additional context, for example the help center for an app, may enable special purpose agents.
How it impacts the tech landscape:
The big question is who will we buy agents from? General purpose agents could run on the desktop or as a Chrome extension. But the day an agent is able to complete a reconciliation in QBO, who will I buy that agent from? My PM? Intuit? Someone else? Will a platform obfuscate the user interface into something more simple?
Honorable mentions:
Massively improved OCR quality thanks to vision models:
Vector search eliminating rules-based bookkeeping classification:
Better integration of meeting transcripts into PM systems
The age of avatar advisors, and creating your very own:
Microsoft Copilot? Who knows ?? I hope it's amazing but nothing I've seen yet leads me to believe it will be.
Anything I missed? What do you anticipate, or as an accountant, what problems would you love to see AI solve for you?
Tax + Tech | Focused on Data Science, AI, and ML
10 个月I would like to see more emphasis on machine learning, using valuable data to identify new service opportunities and insights.
The API for AI bookkeeping | AzoraOne
10 个月Great vid on the path to "fully" automated bookkeeping. We are solving this problem in Northern Europe right now, but we focus on documents (yes, more context). Not exactly in the way you describe it technically, but something similar. We build models on even smaller scale, per document. The coding/categorisation that the software learns is (in the base case) stored per customer, and we can also share this knowledge between customers within the software. Continuous inheriting of live knowledge between companies. The learning is rapid; looking at a couple of transactions per vendor will do the trick. More complex coding, like multiple accounts and VAT rows, takes a few more interactions with data. Currently we are piloting our next big bet (a technical concept we call Bricks) with a leading accounting firm in Sweden, and we believe we will be at "fully" automated bookkeeping for any of their customers before the end of this year. One can not use fully without the "", since there will always be stuff like a first document from a new vendor and so on. developer.azora.one - our API that any accounting software provider can use Just write me a line if you′d want to learn more!
FinTech, AI & Blockchain | Board Member | Chief & NVTC
10 个月Love it! AI will make it easier to get through complex contracts in a more efficient way to make all the GAAP and IFRS filings easier. This is what we strive for at mangodoc.ai
The Student Advocate of Cloud Accounting | Host of the Accounting Conversations Podcast | Passionate Accounting Student | CPA Eligible Spring 2026 | Associate @ Actuate Cloud Professional Accounting
10 个月As a student I have already seen the huge impacts of AI on research. It is most valuable when it can be linked to specific documents. Complex tax and legal documents can be scanned quickly by AI now instead of having so much information memorized.
Advisor to business owners and families who value planning for their futures
10 个月I have a very Snarky Chat Bot on my website at Neil.tax currently I don’t have the know how to get its API’s talking to anything like Zapier, 365, or TaxDome. I want to know and or have the providers of this stuff figure it out so badly. Like if AI could have a client’s TaxDome folder structure/contents known, and could create a link to a specific file for the client… then the client is hit with the firewall/login, but after entering username, they have there file sitting there. Heck even if I could prompt AI and get the url without having to navigate to files custom url. Wild times.