AI has gone from a side experiment to a working part of daily accounting life in a very short span of time. In fact, Deloitte shares that 87% of CFOs expect AI to be extremely or very important to finance operations in 2026. Firm partners, controllers, and finance leaders now have to decide where AI truly belongs, whether it be in the close, in reconciliations, in forecasting, in compliance work, or in reporting.
A lot of that conversation gets stuck on raw model capability. But for a profession that runs on accuracy and accountability, capability alone was never going to be the real test. The harder, more useful question is whether the systems, controls, and data sitting underneath the AI can actually support it once it’s running in a live environment.
Accounting only works if every figure can be traced back, reproduced, and defended to parties like a client, a board, or a regulator. AI that sits outside that structure can look impressive and still be useless for anything that ends up in financial statements. Talking with finance and accounting leaders, I keep running into the same three roadblocks that keep off-the-shelf AI tools out of real accounting workflows.
1. Generic AI doesn’t understand how accounting actually works
Large language models are mostly trained on broad, public, general-purpose data. That’s great for things like drafting an email or summarizing a document. It’s nowhere near enough for the level of precision accounting requires.
Accounting runs on context these models were never given. A generic model has no idea what a particular chart of accounts looks like, how an entity is structured, where materiality thresholds sit, or what the internal control rules are. It doesn’t know which accounts get extra scrutiny, how intercompany eliminations get handled, or what came out of last year’s audit.
Strip away that context and what’s left is a guess, dressed up as an answer or a statistically likely response, instead of one grounded in actual accounting logic. Plenty of business problems can tolerate that kind of variability. Accounting can’t. The same question, asked the same way twice, has to produce the same answer both times.
That’s exactly why the AI projects that actually work tend to start somewhere unglamorous: cleaning up and standardizing the data first. Without reconciled, controlled data, even a genuinely strong model is going to produce results nobody can trust. Data scattered across spreadsheets, disconnected systems, and manual handoffs will sabotage even the best AI on the market.
2. AI’s “black box” nature doesn’t play well with audit requirements
The second wall shows up the moment AI touches anything that needs to be controlled.
Audit and compliance aren’t optional in accounting. SOX, internal controls, and financial reporting obligations all require being able to explain exactly how a number came to be and prove the right steps were followed. Every control needs to be written down, repeatable, and ready to review or audit.
That’s not how most generic AI tools were built. They produce answers based on patterns learned from training data, not a documented chain of decisions. Ask why a model landed on a specific output, or which inputs mattered most, and there often isn’t a clean answer to give.
In accounting, that gap is a real liability. The moment AI proposes a journal entry, flags something unusual, or finishes a reconciliation, someone needs to be able to point to:
- The data that went into it
- The rules that were applied
- The controls that were enforced
- The person who reviewed and signed off
No audit trail means no basis for relying on that output in financial reporting. This single issue explains why so many promising AI pilots in accounting never make it past testing. The tool looks great in a demo. Then someone in review or audit asks how the answer was actually produced, and the whole thing stalls because there’s nothing solid to point to.
3. Treating governance and security as an afterthought doesn’t work
Few categories of data are as sensitive as financial data. General ledger detail, payroll, tax records, revenue figures all come with strict rules around storage, access, and sharing, and that holds true whether you’re a two-partner firm or a multinational finance department.
Lots of generic AI tools route data through external infrastructure that was never designed with regulated information in mind. That immediately raises questions about who can access what, how duties stay segregated, where the data physically lives, and whether compliance can actually be enforced.
Sometimes client or company data legally can’t leave the system of record at all. Other times, there’s no way to prove the data stayed protected the entire time it was being processed. Either scenario makes it very hard to justify plugging a generic AI tool into core accounting work.
That’s why governance has to come before the model, not after it. Access permissions, audit logging, workflow controls, and security policy all need to be settled before AI ever enters the workflow. Skip that step, and the risk to financial reporting integrity gets too large to justify.
The real need is trust infrastructure, not a smarter model
These three gaps add up to something a lot of accounting leaders are starting to recognize on their own: a sharper model isn’t the bottleneck. What’s missing is the financial infrastructure that would let any model, sharp or not, operate safely inside a controlled process.
For AI to actually be trustworthy in accounting, a few things have to already be in place:
- Data that’s governed and reconciled
- Systems that are truly integrated with each other
- Strong internal controls that are documented, not assumed
- Access and approvals tied to defined roles
- A complete audit trail behind every action taken
To make this work, it takes accounting, IT, internal audit, and security working together, whether that’s inside a firm or a corporate finance function. Decisions that used to be made one department at a time now get evaluated across the whole organization, especially once AI enters the picture. This is because the stakes, like regulatory exposure, reporting risk, and the cost of a mistake, are all higher than they used to be.
More organizations are responding by setting up formal AI governance, pulling finance, IT, legal, and compliance into the same room to make sure new tools fit existing control frameworks before anyone deploys them. That structure is what keeps things from sliding into fragmented tools, inconsistent data, and what a lot of people are now calling “AI sprawl,” or new technology showing up faster than anyone can actually govern it.
In a field built on getting the numbers right, that kind of sprawl just doesn’t hold up.
Control, not experimentation, will define AI’s future in accounting
AI is going to end up woven into nearly every part of accounting work. But the firms and finance teams that benefit most won’t be the ones that ran through the longest list of tools.
They’ll be the ones that integrate clean data, consistent processes, real controls, and interoperable systems. So, when AI does show up, it’s operating inside the same rules accounting has always run on.
ABOUT THE AUTHOR:
Michelle Stalick is chief accounting officer at BlackLine.
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
