Last Wednesday night, one of the founders of an AI unicorn, Harvey, published a research update on X. It was a technical piece with some teeth, but I am so happy I read it though. Even though the target audience was not accountants, there were a bunch of signals for us in that research.
What Harvey did was create and develop “multiple attorneys” who can produce quality work. They did not build a bigger chatbot. Instead, they developed with leading, existing AI models, and put them through an apprenticeship.
They fed the models real-life work assignments. They ran experiments to train it. For example, they fed it a pile of real matter files with irrelevant documents mixed in, because real work means finding the right information, not being handed it. They gave it tools and a rubric, written by experienced lawyers, describing exactly what finished work looks like.
They made the models do the work thousands of times, grading them each time against that rubric, and the grades changed the model so the judgment was baked in.
They trained it the way we train entry-level and staff accountants. Real-life repetition with real files and a senior reviewer who knows what the finished product should look like before going to delivery. This is the process Harvey went through.
The results should help accountants see how companies like Basis will shape the way accounting work gets done. Market acceptance will occur because of the trust earned through the way the models have been trained.
I think it’s helpful to provide an understanding of the new way accounting intelligence will be stacked in layers.
Layer one is foundation intelligence where the models live. OpenAI, Anthropic, Google, the open models. They continue to get better every quarter, and the leader seems to be interchangeable. As one gets better, the others also find ways to improve. This race belongs to the labs.
Layer two is profession-based intelligence. You train and teach AI the work; for example, what reconciliation is, what a close requires, or what a workpaper looks like when it is right. In law, one of the leaders is Harvey. In accounting, this is what companies like Basis are building right now. Layer two is going to be industry-specific and will be bought.
Layer three is firm and client intelligence. This is where the institutional nuances will be captured. How your firm closes the books the way a client’s CFO cares about, which classifications recur, what the lender covenants say, and when something should be escalated. Layer three is your intelligence, your domain knowledge.
Why does this matter?
The labor shortage in accounting is real and is not going away anytime soon. The attrition, retirement, and transient nature in accounting have created a problem because we have historically stored much of our institutional knowledge in people. This is the model that has been in place and is failing us:
Each hire adds more than capacity: they learn why an accrual was handled a certain way, which exceptions recur, what a reviewer expects, and what the auditors asked about last year.
When those seats stay open or experienced people leave, the firm loses both capacity and access to that context.
You may think, “Well, wait. We have an ERP that captures all of this, right?” The ERP solves an important but different problem. It holds the official record: the journal entries, balances, and final outcomes. But the evidence, standards, assumptions, and judgment that make the record trustworthy often live around it. The information contained within requires someone to put context around it. It’s static information that is available to the user, but it’s not intuitive.
An ERP is Frankenstein’s monster before capturing the lightning. Every part is in place, near perfectly assembled, lying still on the table. It has a spine, a skeleton, even a heart! Except, it needs something else to bring it to life. Accountants have been waiting for the day to say, “It’s Alive!” Perhaps the intelligence layer is the lightning we’ve been waiting for.
The intelligence layer works across the whole organization and brings data to life.
As people work in intelligence layers, like Basis, you can set preferences, client-specific rules, and best practices that can be shared across clients, teams, service lines, and offices.
The goal is to standardize what should be common while preserving the unique parts of each engagement.
One of the main benefits of the intelligence layers is the multiplicative nature. As the intelligence layer is used more frequently, the value compounds. The more a firm’s context becomes available to agents, their work becomes more tailored to that firm. Additionally, as AI models improve, those agents can use the same foundation to take on more work and increasingly complex tasks.
In accounting especially, trust is important at every step of the process. Agents should work as though they are being audited along the way: their sources, assumptions, decisions, and workflow should be available for review. It’s important to note that the firm still owns the judgment. Agents execute more of the work; accountants direct, review, and decide.
We have long undervalued institutional knowledge and intelligence. We’ve left ourselves vulnerable over and over again. Think about the way work has been prepared and reviewed. This is the knowledge of how quality work has been done. We’ve put in place review procedures to ensure quality yet allowed the knowledge to evaporate.
Here’s the way it’s been done. A senior looks at a staff accountant’s work. They catch the misclassification. They know this client always books that rebate wrong in the fourth quarter. They know the partner will bounce the memo if the variance explanation leads with the excuse instead of the number. They fix it, explain why over a cubicle wall, and the work is delivered. Two weeks later, they give notice and all the knowledge goes out the door.
We call that quality control. The last step before the deliverable ships. It works, except, most of the time, nobody wrote down why certain things happen on a client. The poor next person up to fill those shoes must dig through piles of prior year workpapers searching for “why.”
This is the training data. The type of data that Harvey used to develop its models. It’s the most valuable kind. Every override, every caught exception, every “the partner will want to see this differently” is professional intelligence being created that will now be captured and multiplied. Review was never just review. It was teaching so knowledge can be transferred and multiplied across the organization. The problem has always been that it walks in and out every day.
Finally, finally, one more time… finally the knowledge stays. It’s ALIVE!
