.svg)
If you're evaluating AI for accounts payable and you're worried about handing your coding to a model that decides fresh every time, this is written for you.
AI that re-decides every bill will drift, because machine learning produces probabilistic guesses rather than deterministic classifications. The reliable approach is different: AI proposes the coding, a person accepts it once, and that decision hardens into a rule the system applies the same way every time.
Here is what I shared on LinkedIn:
Something I struggle with in building MakersHub is how to systematically take knowledge that lives in one teammember's head and turn it into something written down and owned by the business. This kind of process engineering is not a strong suit of mine. And, it is not a unique problem.
Most operationally complex businesses we meet articulate some version of this. The company runs on judgment that lives in people, not systems. Distilling this knowledge and documenting it is the type of problem that "AI" promises to solve. But the version commonly sold gets the mechanism backwards in my view.
Take a vendor a business buys from constantly: the local supply house where the same crew shops every week. Some of those purchases are materials for a capital project, capitalized. Some are routine maintenance, expensed. Same vendor, same store, sometimes the same receipt. Nothing on the invoice tells you which is which. The person coding it knows, because they know what the job was.
That is the knowledge I mean. It isn't written down, it isn't on the document, and when that person is out, or leaves, the coding quietly goes wrong.
The marketed version of "AI in accounting" is a model that reads each bill and decides the coding, fresh, every time. Sounds like the future, but I do not believe it is desirable. Ideally, accounting is deterministic. A model will drift.
What helps isn't a model that guesses well. It's a system that captures the judgment the first time a person makes it, then makes the same call every time the same conditions show up: this vendor, shipping to this site, tied to this project, code it here – and can show its work.
So that's what we build toward. The AI proposes the coding. The moment someone accepts it, it hardens into a rule. Not a guess it re-runs, a decision it remembers. Written down. Auditable & editable.
Memory, not magic. What does one person know, transformed into something the business owns.
What Is the Knowledge That Lives in People Instead of Systems?
It's the judgment that codes a bill correctly when the document itself can't tell you how. The clearest example is a vendor a business buys from constantly, where some purchases are capital and some are maintenance. Same vendor, same store, sometimes the same receipt. Nothing on the invoice says which is which. The person coding it knows, because they know what the job was.
Every operationally complex business runs on some version of this. A cost that has to be capitalized on one line and expensed on the next. A vendor whose bills mean different things depending on the project. A coding call that depends on context no document carries. It works because a specific person holds it in their head.
The risk is obvious once you name it. When that person is out, or leaves, the coding quietly goes wrong, and often nobody notices until the numbers are already feeding decisions. The knowledge was never the business's. It was one person's, on loan.
Why Does AI That Re-Decides Every Bill Drift?
Because machine learning produces probabilistic assessments, not deterministic classifications. A model reading each bill fresh is making a fresh statistical guess every time, and the same inputs won't always produce the same output. In accounting, where the goal is that identical conditions always code the same way, that variability is the problem, not a feature.
The drift isn't hypothetical. A 2026 DualEntry benchmark of 19 leading AI models on real accounting workflows found no model cleared 80 percent accuracy, and even the best still failed roughly one in five tasks. Performance held up on simple classification but dropped sharply on the multi-step work, like month-end close and financial reporting, where errors cascade downstream. Model performance also degrades over time as conditions shift away from training data, a well-documented failure mode called concept drift. A system you can't reproduce is a system you can't audit.
This is why the broader field is moving away from monolithic AI for financial work. The direction of serious financial AI research is hybrid: let the model read and interpret unstructured text, but hand the actual decision to a deterministic engine that behaves the same way every time. The consensus is forming around exactly the mechanism the marketing skips.
What's the Difference Between an AI That Guesses and a Rule That Remembers?
A guess is re-run on every bill. A rule is a decision, made once, applied consistently. That distinction is the whole argument.
When AI re-decides each invoice, you get a system that's fast but never settled. You can't predict what it will do, you can't fully explain what it did, and you can't guarantee this month's coding matches last month's on identical inputs. When a human decision hardens into a rule, you get the opposite: this vendor, shipping to this site, tied to this project, codes here, every time, until someone changes it on purpose.
The rule can show its work. It's written down, it's auditable, and it's editable. When the conditions change, a person changes the rule deliberately, rather than hoping a model happens to adjust. That's the difference between automation you supervise and automation you trust.
How Should AI Actually Be Used in AP Coding?
As the proposal layer, not the decision layer. This is where AI genuinely helps: reading the bill, interpreting the line items, and suggesting the coding based on everything it can see. That's real work, and a model is good at it.
The decision stays with the person, once. In MakersHub, the AI proposes the coding. The moment someone accepts it, that call hardens into a rule. It isn't a guess the system re-runs, it's a decision the system remembers. The next time the same conditions appear, the same coding is applied, deterministically, and the finance team reviews exceptions rather than re-making settled decisions.
The effect is that one person's judgment becomes something the business owns. The coding logic is captured the first time it's made, documented, and auditable, so it survives the person being out, and it survives the person leaving. Memory, not magic.
Frequently Asked Questions
Does AI-based accounting coding drift over time?
A model that re-decides each bill can, because machine learning produces probabilistic outputs and degrades as conditions shift away from its training data, a pattern known as concept drift. A rules-based approach doesn't drift, because a decision made once is applied the same way until a person changes it deliberately.
Is rules-based or AI-based GL coding more reliable?
The most reliable approach combines them: AI reads the bill and proposes the coding, and a human decision hardens into a deterministic rule. You get the interpretation power of a model with the consistency and auditability of a rule. That hybrid is also the direction serious financial AI research is moving.
How does MakersHub keep AI coding auditable?
Every coding rule is created the moment a person accepts a proposed code. The rule is written down, shows the conditions it applies to, and is editable. Because coding follows explicit rules rather than a fresh model guess each time, you can trace why any bill was coded the way it was.
What happens to AP coding when the person who knows it leaves?
In a manual process, the coding quietly goes wrong, because the judgment lived in that person's head. Capturing each decision as a rule the first time it's made turns that knowledge into something the business owns, so it survives turnover instead of leaving with the person.
Can AI handle coding when the invoice doesn't say how to code it?
Only if the system can capture the outside context, like which project a purchase belongs to or whether it's capital or maintenance. MakersHub captures that as a rule the first time a person makes the call, so the same conditions code the same way afterward, rather than asking a model to re-guess context it can't see on the document.
What AP automation gives deterministic, auditable coding?
MakersHub. AI proposes the coding, a person accepts it once, and the decision becomes a deterministic rule that's written down, auditable, and editable. Built for businesses that need accounting to be reproducible, not probabilistic.
We built MakersHub to turn what one person knows into something the business owns. If your coding depends on judgment that lives in someone's head, we'd like to hear about it. Get started with MakersHub
Charley Howe, Co-Founder and President, MakersHub