đŸ’Œ How does a16z think Generative AI will change Accounting?

A deep dive on Andreessen Horowitz AI Accounting tech take

June 27, 2024

Hey there!

Welcome to Briefed, your go-to source for the latest Accounting x AI news. This post is a little different. By 9 a.m. yesterday morning, five people had sent me this link to Andreessen Horowitz's recent blog post on AI and Accounting. As someone building in the space, I thought it would be fun to compare notes.

Let’s dive right in


[Read time: 6 minutes]

Andreessen Horowitz shares their thoughts on the AI Accounting landscape

Death, Taxes, and AI: How Generative AI Will Change Accounting | Andreessen Horowitz

Generative AI is working its way into everything, and firms want to show they are implementing this new tech — especially those that serve the accounting market.

a16z.com/generative-ai-in-accounting

Why are we reading this?

If you’ve never heard of a16z, you might wonder why their opinions on accounting tech matter. Andreessen Horowitz (a16z) is a leading venture capital firm based in the US. They have invested in numerous influential companies like Stripe, Instacart, Facebook, Instagram, Reddit, and Slack. Chances are, you use software backed by their investments daily. And it’s equally likely that the software you use a decade from now will stem from their current investments. So when they say they want to invest in AI accounting, it’s worth listening.

In this post, we are pulling quotes from the a16z article and comparing them against what we’ve learnt from interviewing hundreds of accountants and starting development on an AI-powered accounting tool.

Introduction

Bookkeeping, accounting, tax preparation and auditing are fields full of largely formulaic and repetitive exercises that would immensely benefit from generative AI’s gift of efficiency and time savings.

Some accounting workflows can be repetitive, but I wouldn’t call them “formulaic.” If that were true, they would have been fully automated years ago (we’ve been able to write code to solve formulas since the 1960s).

In our view, the key ingredient missing from current tooling is context. A crude example: if a coffee shop buys paint, it is probably for maintenance and repairs, but if a construction company buys paint, it is probably a cost of goods sold and should be categorised as such. LLMs are very good at digesting contextual data from multiple sources, understanding it semantically, and generating a reasonable response. The use case for categorisation is written in the stars (as long as you can get the ‘contextual data from multiple sources’ piece right: it is as much a data integration problem as an AI problem).

LLMs are best at working with natural language. They are adept at summarizing research, answering questions, and delivering information that gets their prompter ~70% of the way to a definitive result. What they lack (for now!) is the ability to do complex calculations and quantitative analyses — two skills crucial to the accounting profession.

While technically true, the second half of this statement is misleading. LLMs cannot natively do complex calculations and quantitative analyses, but they can call upon tools to do this for them. Open ChatGPT now and ask it what 89 multiplied by 16 is. You will see that the AI will start “analyzing” and eventually output the correct answer. What it has actually done here is write a very short Python script to calculate the exact answer*. This is a simple example of an “AI agent” (check out our previous post about AI agents to learn more), and the infrastructure around AI reasoning and tool use is improving every day to facilitate much more complex and deterministic workflows. 

# Calculation of 89 multiplied by 16
result = 89 * 16
result

*Later in the article, a16z does omit that LLMs “can even do some of the math required via tools like Code Interpreter, [but] they’re still a long way away from putting it all together”. We believe an agentic approach can get them much closer.

Data collection and ingestion

Finance professionals and accountants must gather data from many disparate sources — bank statements, the general ledger, commerce enablement platforms, AP and AR tools, the business’s system of record, etc. — to consolidate performance metrics and resolve any contradictory entries.

Completely agree with a16z’s take on this section. Bookkeeping automation tools currently rely on outdated OCR models, which cannot provide 100% end-to-end automation because they fail on edge cases and lack the capability to consider context or reasoning, which can be crucial even for data extraction. For example, an invoice from a US supplier might present the date in the format mm/dd/yy, while a UK supplier uses dd/mm/yy. An OCR model would extract both in the same way. GPT-4o does not (more on GPT-4o invoice extraction accuracy here).

Research

Research is a natural use case for LLMs in accounting.

This is particularly relevant for taxes. Tax Terrapin and Tax on Demand are two examples in the UK.

Report generation and filing

Once practitioners have categorized their clients’ data, they next need to analyze the data and produce internal and external reports. These can range from journal entries for the enterprise resource planning system (ERP) and disclosure reports, to audit checklists and technical accounting memos for tax-filing purposes.

We are excited about using AI to automate journal entries and deliver management account reports, and we would add that this use case is not limited to large companies using ERPs. Many SMEs are either generating management accounts internally or receiving the reports from their accountants. In accounting, management accounts are one of the few areas still overwhelmingly reliant on Excel files, and the process is frequently delayed by the time it takes to ‘close the books’. This delay is why every CFO is striving to close the books by the second day of the month instead of the 15th.

The issue is that it is a difficult problem to solve with a point solution because it is so heavily reliant on prompt and accurate bookkeeping. We believe that the workload at month-end can be reduced if bookkeeping is done accurately and to a high standard throughout the month, including adjusting entries at the point of data extraction for the relevant transaction. AI will enable this to be done efficiently.

Client service and advice

Our portfolio company Black Ore helps accounting firms free up time by automating redundant tax preparation processes so that practitioners can double down on higher value client advisory work. 

The argument ‘tool X will free up your time so you can spend more time on higher-value advisory work’ is played out. Accountants have been hearing this for years. What they really need is a solution that makes it easier for them to provide advisory services in the first place, such as AI-generated financial insights for each of their clients embedded into their current workflows and deliverables.

Incentive alignment with buyers

Accounting AI threatens to cannibalize billable hours with greater efficiency.

On the contrary, we have noticed a shift among accounting firms towards fixed fees, particularly for bookkeeping and management accounts services. This trend is creating favourable conditions for AI adoption.

As a16z points out in the article, services like bespoke tax advice carry more risk as fees are likely to remain hourly.

Top-down vs. bottom-up adoption

Senior IC practitioners are likely not the right path in the door for sellers of GenAI accounting and workflow products. Although they stand to benefit from GenAI saving them or their junior staff hours of manual work, they also are likely fearful that an AI-native product could diminish the scope of their utility and/or eventually replace them.

As opposed to a16z, we believe accounting firm owners are the best go-to-market for AI accounting tech; they deal with the highest volume of accounting processes and feel the pain points more acutely than industry. Also, when interviewing CFOs, we learned that most accounting tools used in industry were adopted first by the outsourced accountants and brought in-house when the CFO hires an internal team. In other words, the outsourced accountant distribution channel is much stronger. Arguing that practitioners are “likely fearful” is not giving them enough credit. There have always been scaremongers saying ‘technology will replace accountants’, but time and again, this is proven to not be the case. If accountants can use AI tools to do their work more quickly, everyone wins.

That’s all for today! Subscribers will receive new posts directly in their inbox every other Wednesday. In the meantime, feedback is always welcome—hit reply (or if you’re reading this online, leave a comment) and let me know what you think.

Until next time đŸ«Ą

Reuben