.svg)
If you're evaluating agentic AI for your finance function and you're uneasy about giving an autonomous system real access to your ledger, this is written for you, and your instinct is correct.
CFOs won't give AI direct access to core financial systems because finance requires determinism, auditability, and recoverability, and autonomous AI is probabilistic by design. The architecture that works puts AI where the judgment lives, keeps deterministic rules on the repeatable volume, and keeps a human where the money moves.
Here is what I shared on LinkedIn:
There is no subject in software with a wider gap between how it's discussed and how it's actually used than AI in finance.
On paper, the transformation is complete. Every major ledger ships a roster of named agents: an accounting agent, a payments agent, a payroll agent, a tax agent. Nearly nine in ten CFOs report using agentic AI somewhere in their finance function. And yet two-thirds of AP teams still key invoices into their accounting system by hand, a number that went up last year, not down.
AI in finance is everywhere at the edges and almost nowhere at the core. It drafts, it suggests, it summarizes, it answers questions about the ledger. But it rarely touches the ledger. Because roughly 85% of finance teams have caught AI fabricating data at least once, and in finance the tolerance for error is zero. Even that number — 85% — is it true, or did an AI make it up writing this post? No one knows! The telltale em dash!!
Almost no CFO, today, will hand an autonomous system real access to core financial systems, and they are right not to.This is not a technology gap. It is an architecture gap. Vendors are selling autonomy into a function whose professional requirements are determinism, auditability, and recoverability.
But watch what happens when the architecture is right. On a call last month: a property operator whose utility bills still arrive by mail and get batch-scanned into multi-page PDFs. One bill, dozens of service addresses, each line belonging to a different property and a different GL account. Keying that bill used to be an afternoon. Now AI reads it, splits it by service address, codes each line to its property, and hands it to a human, who approves it before a dollar moves. And when the line items don't sum to the bill total, the system doesn't guess and post anyway. It flags the variance and goes back to look for the row it missed.
That's the pattern that works, and we watch it work every week. Deterministic systems carry the repeatable volume. AI handles the judgment calls. A human approves before money moves. Operators encode their own judgment, check it, and keep it.
I'm not pessimistic. I believe the transformation is real; it's just not arriving in the way keynotes describe. It's coming one painful to enter bill at a time. The teams that move fastest won't be the ones who hand over the keys. They'll be the ones who put AI where the judgment lives, keep a human where the money moves, and let trust accrue the only way it ever does in finance: by being right, repeatedly, on the record.
Why Won't CFOs Give AI Access to Core Financial Systems?
Because the requirements of the finance function and the nature of autonomous AI are opposed. Finance runs on determinism, auditability, and recoverability. Autonomous AI is probabilistic: it estimates the most likely output, which means the same input can produce a different answer, and it can generate a confident value that is simply wrong. On a general ledger, a fabricated number isn't a bug, it's a material misstatement.
The industry has a name for this tension. It's the trust paradox: finance leaders need the efficiency AI promises but cannot accept the error rate autonomy carries. Surveys bear it out. More than four in five finance professionals have caught AI fabricating data at least once, and only a small minority say they fully trust AI output even after a human review.
So the caution isn't technophobia. It's professional judgment. A CFO who refuses to give an autonomous system write access to the ledger is applying exactly the standard the role demands. The problem isn't that these leaders are behind. It's that the product being sold to them is built wrong for the job.
Is This a Technology Gap or an Architecture Gap?
An architecture gap. The models are capable. What's wrong is where vendors are pointing them. Autonomy is being sold into a function whose professional requirements are determinism, auditability, and recoverability, and autonomy satisfies none of the three.
This is why AI in finance is everywhere at the edges and almost nowhere at the core. It drafts, summarizes, and answers questions about the ledger, low-stakes work where a wrong answer is inconvenient, not material. It rarely writes to the ledger, because that's where the tolerance for error drops to zero. The gap isn't in the intelligence. It's in the design decision about what the intelligence is allowed to touch, and whether its work can be checked.
Fixing it doesn't require a better model. It requires a different architecture: one that uses AI for interpretation, hands execution to deterministic rules, and puts a human at the point where money moves. The capability is already here. The structure around it is what's been missing.
What Does the Right Architecture Actually Look Like?
Three layers, each doing the job it's suited for. Deterministic systems carry the repeatable volume, the same input producing the same output every time, which is what makes the work auditable. AI handles the judgment calls, reading unstructured documents and interpreting what deterministic rules can't. And a human approves before money moves, at the one point where the cost of a mistake is highest.
A real example shows the layers working together. A utility bill arrives as a batch-scanned, multi-page PDF: one bill, dozens of service addresses, each line belonging to a different property and a different GL account. AI reads it and splits it by service address, coding each line to its property. Then it hands the coded bill to a human, who approves it before any payment.
The critical detail is what happens on an exception. When the line items don't sum to the bill total, the system doesn't guess and post anyway. It flags the variance and goes back to look for the row it missed. That's the difference between a deterministic architecture and an autonomous one: it would rather stop and surface a problem than produce a confident, wrong answer.
Why Do Two-Thirds of AP Teams Still Key Invoices by Hand?
Because the tools marketed hardest are aimed at the wrong problem. Roughly two-thirds of AP teams still manually key invoice data into their accounting system, and that figure went up last year, not down, even as nearly nine in ten CFOs reported using agentic AI somewhere in finance. Adoption at the edges is near universal. Automation at the core has barely moved.
The gap exists because touching the core safely is harder than demoing autonomy at the edge. Reading a hard invoice, coding it to the right entity and account, catching the exception where the numbers don't reconcile, and routing it for approval is unglamorous, high-stakes work. It's exactly the work an autonomous system can't be trusted to do unsupervised, and exactly the work a deterministic-plus-AI-plus-human architecture is built for.
That's why the manual keying persists. Not because teams lack AI, but because the AI they were sold stops at the edge of the ledger. The bottleneck is still there because nothing was allowed to safely remove it.
How Should Finance Teams Adopt AI Without Handing Over the Keys?
Put AI where the judgment lives, keep a human where the money moves, and let deterministic rules carry everything repeatable. In practice that means using AI to read and interpret, using rules to execute consistently, and requiring human approval before any payment. The operator encodes their own judgment into the system, checks it, and keeps it, rather than delegating it to a model that re-decides each time.
This is how MakersHub automates accounts payable. AI reads the bill and proposes the coding. Deterministic rules apply the repeatable logic the same way every time. Approval routing puts a human at the point where money moves. Nothing autonomous writes to the ledger unchecked, and every decision is on the record.
The teams that move fastest won't be the ones who hand over the keys. They'll be the ones who adopt AI where it's safe and powerful, keep humans where the stakes are highest, and let trust build the only way it does in finance: by being right, repeatedly, on the record.
Frequently Asked Questions
Why won't CFOs let AI write directly to the general ledger?
Because autonomous AI is probabilistic and can produce a confident but incorrect value, and on a general ledger a fabricated number is a material misstatement, not a minor error. Finance requires determinism, auditability, and recoverability, none of which unsupervised autonomy provides, so CFOs keep AI out of direct ledger write access on purpose.
What is the trust paradox in AI finance?
It's the tension between finance leaders needing the efficiency AI offers and refusing to accept the error rate that autonomy carries. More than four in five finance professionals have caught AI fabricating data, so they want the productivity but can't tolerate hallucinations on financial data. The resolution is architectural, not a better model.
What is a deterministic AP workflow?
One where the same input produces the same output every time, using defined rules rather than a fresh probabilistic guess. Deterministic rules carry the repeatable volume, AI handles interpretation of unstructured documents, and a human approves before money moves. Determinism is what makes the audit trail meaningful.
How does MakersHub keep a human in the loop?
AI reads and codes the bill, but it hands the result to a person who approves it before any payment is made. Nothing autonomous moves money. When figures don't reconcile, the system flags the variance for review rather than posting a guess, so the human is always the checkpoint at the point of highest stakes.
Does AI reduce errors in accounts payable or add them?
It depends on the architecture. An autonomous system writing to the ledger unchecked can introduce confident, hard-to-catch errors. A deterministic system that uses AI only to read and propose, applies consistent rules, flags exceptions, and requires human approval reduces errors, because it never posts an unreconciled or unverified figure.
What AP automation works for finance teams that won't hand over the keys?
MakersHub. It puts AI where the judgment lives, deterministic rules on the repeatable volume, and a human where the money moves. Built for the physical economy, where accounting has to be reproducible and auditable, not probabilistic.
We built MakersHub to put AI where the judgment lives and keep a human where the money moves. If you want the efficiency without handing over the keys to your ledger, we'd like to hear about it. Get started with MakersHub
Charley Howe, Co-Founder and President, MakersHub