The Filing Cabinet and the Brain

A Dive into the AI-Native ERP | The Ledger and the Machine - Part Two of a Four-Part Summer Series

Last week, I wrote about the chatbot that could not count, but how its development has led to a movement in accounting. The chatbot was proof that the machine we’ve been relying on for many years was getting much smarter and could now read, reason, and respond. But, for the machine to truly help accountants, the building of valuable tools needs to be further down into how businesses run.

For this week, while you are hopefully somewhere with sand between your toes, I want to speak with you about what I mean by building further down into the “plumbing” of a business.

Every business we interact with, whether that’s a restaurant, grocery store, dentist, or the company you work for, runs on a piece of software. It is often referred to as a system of record or more frequently, a three-letter acronym, ERP, which stands for enterprise resource planning. It came into existence to connect the different functions into one system. In a way, it is the filing cabinet where the official version of everything inside a business lives. Every sale, every invoice, every paycheck, every dollar in and out is captured in the ERP.

For thirty years, owning the system of record has been an incredibly “sticky” business in software. Companies like SAP, Oracle, and NetSuite are some of the bigger brand names, but there are many others that serve as the backbone of a business. The reason I refer to it as “sticky” is because in a way, it’s an electronic version of a safe, your entire financial history lives in this “filing cabinet.”

There are very interesting research results on surveys done to ascertain ERP satisfaction. The stats come back paradoxically. Market data regularly shows that 30 to 35% of companies are actively planning to transition away from or replace their current ERP/legacy solution, yet when surveys are sent to ask about ERP satisfaction, nearly 80% of businesses report satisfaction. That math makes no sense.

Moving on from an ERP is extremely difficult. One venture investor described replacing an ERP as ‘open-heart surgery while the patient is running a marathon.’ Companies would rather live with software they do not love rather than deal with the pain of starting over.

Another interesting take on the ERP is that it’s often looked upon as the Accountant’s Ledger, as if it’s this beautiful piece of software that is written specifically to make accounting easy. Jeopardy buzzer sounds. Sorry, the answer “What is an ERP” is not the answer to name a popular accounting ledger.

For many accountants, working inside of the ERP is nearly impossible to perform all of the accounting procedures necessary to produce the reporting that management wants to see. An example of this is the accounting ritual we refer to as the close. Every month-end, accountants are tasked with ensuring the books agree with reality by reconciling every account and chasing down every mismatch. It is slow, it is manual, and it has been a part of the “slow” rhythm of the profession for a century. The ERP holds the record, but the accountant is the one who makes the record trustworthy.

The question being asked today now that AI is upon us is: what happens to the ERP? A few months ago, Salesforce sent a strong signal to the market when it announced it was going ‘headless.’ They conceded that, in the future, we are not going to click around the way we have been inside software. Instead, we are entering into a new world where we will be interacting with machines that aren’t static. AI agents will read and write the data directly. The software that forced you into using the software in certain ways with their screens, their dashboards, their buttons, referred to as the ‘head’, will not matter anymore. The data underneath all of this is what matters and the ways we interact with this data are changing.

For years, accountants have had to be trained in software to perform its craft. Please understand this. In many industries, tools were developed and innovated to make craft easier, for the finished product to be more reliable. This has not been the case in accounting. In fact, many people would argue that they had more control and visibility with handwritten ledgers than dealing with the difficulties and complexities navigating ERPs. Entire careers have been built on mastering the clicks of one system or another, or being an expert in the software, not even the craft itself!

Seema Amble, an investor at Andreessen Horowitz, wrote an essay shortly after the Salesforce release that it was going headless. She posed a question in her essay: if you strip away the screens, what is actually left? In other words, what makes the filing cabinet valuable once nobody is opening its drawers by hand? Her answer:

“Agents may kill muscle memory as a moat, but they do not kill operational logic and context as a moat. If anything, they make that logic more important, because agents need explicit rules, permissions, and process definitions in order to act safely.”

This is what is wonderful about that answer. The actual craft, the intelligence of an accountant, takes over again. Our accountants’ minds are back in control. Can I hear a Hallelujah?

The machines can take over the clicking, but they will not take over the knowing (at least not for now). The knowledge of how a specific business actually runs becomes more valuable in the world of AI, not less. The machines need to function safely and reliably, and it needs people to direct and oversee them.

The most valuable asset in the age of AI is the thing accountants have been accumulating and rarely ever recognized and celebrated for: context and knowledge.

This is where things have been moving in accounting.

Here are some examples for you: Rillet is rebuilding the accounting ledger from scratch with their AI-native focus, so that transactions file themselves. Basis is building AI agents for accountants that are verifiable and traceable. Pylon is building intelligence for customer support. One interesting finding is that startups are discovering in their own data that AI working alone produced measurably worse outcomes than AI working with humans.

The future is not a chatbot, but instead, a new layer of intelligence that will sit on top of the old filing cabinets/ERP where it will read, reason, and act, with accountants holding the judgment.

So, wait, what do we do with our ERP?

There are options. The right one depends less on the software than on what you see for your business and how you want to design it for the future.

To frame this, I want to reference a recent piece from Azeem Azhar’s Exponential View. Here, he illustrated a study that modeled three kinds of companies adopting AI to show that short term financial results are misleading. In the model, the three companies share the same economics and the same hit rate on projects but different learning habits. Two years in, all three are losing similar money. Five years in, the eventual loser looks best. It takes roughly eight years to tell from the outside who made the right call. I feel this is important context because it helps explain the moment we are living through. At this point in the curve, success and failure look identical. The winners are not the companies with the best first project. The winners are the ones which the model calls system builders, where the learning from every experiment carries into the next one. The losers stop experimenting after their first success or run pilot after pilot with no memory and iteration, each new project starting with the same odds as the last.

Option one: wait. Keep your current ERP and let the AI come to you. They are working hard to catch up. At its Sapphire conference this year, SAP announced what it calls the autonomous enterprise with more than two hundred agents. NetSuite announced NetSuite Next, a rebuilt platform with an assistant you can question in plain English, available in preview within the next twelve months.

I get it waiting feels responsible, or prudent, even. I want to at least caution you that simply waiting may limit your ability to adapt within the AI movement. Just know that waiting is outsourcing your learning to a vendor’s release schedule. There are many pieces in your business, which are unique to all, that you will still have to learn how AI changes. Your close, your controls, your team. You don’t want to start learning it years later, while your competitors are compounding experience when they choose to start now.

Option two: replace. Move to an ERP that is AI-native and has a better brain. With an AI-native system, transactions are designed to file themselves where the books are maintained on a more current basis making the close easier and timelier. Of the options, there is a higher ceiling with this one because it comes with a built-in brain in the core. But, you may want to be at a natural breaking point to move in this direction. Perhaps you have outgrown your system, you are hiring a new accounting leader, or you are fundraising and require reporting that your older system can’t produce. You want to think of it from the standpoint of replacing it in a natural rhythm of your business’s life cycle, not in a panic.

Option three: layer. Keep your ERP but put a brain on top of it. This is the system-of-intelligence architecture I noted earlier. The intelligence layer is where the agents will read, reconcile, chase, and draft, while the old ERP keeps doing the one thing it is genuinely great at: maintaining the record. You want to make sure that the intelligence layer that you run tracks every action so you have a trail an accountant can verify. For most companies in the middle market, this option allows the AI and the learning that goes with it to start now. Starting with the intelligence layer allows you to capture context now. Your business “DNA” begins to be tracked and verified. Every rule you teach an agent is a piece of your business’s “DNA” moving out of someone’s head and into an asset the company owns. The layer is how you stop storing it in people’s heads and start storing it in something that compounds.

There is a danger with this option too. Exponential View calls it the project accumulator: the company that keeps launching pilots but never learns, so nothing carries forward. You don’t want to run twelve AI pilots and not capture the memory. With the intelligence layer, you want to think about linear growth in the beginning. Adopt one rule, make it work by enforcing it, and allow it to be captured effectively. Let the process be mapped.

The accountant is moving to the top of the machine, where we have always belonged in the first place.

Join me next week in Part Three, where I’ll talk about the machine that could finally do the work, and the day it moved into Excel.