Traditional review processes in accounting run on a signal that AI does not send. A junior who isn’t sure asks. A preparer who guessed leaves a tentative note in the file. A staffer working past the edge of their training hesitates, and the hesitation is clear: a flagged item, a circled figure, a question in the margin. Together, managers and staff arrive at the correct answer by being clear and collaborative where the most risk is present. AI changes this.
Processing Content
AI severs confidence from correctness. This is the reliability cliff. AI models are trained to present wrong answers with the same fluency, formatting and certainty as right ones. It does not hedge. It does not flag. It does not hesitate, because it was trained to present itself as a magic wand that can handle any problem thrown at it. Anyone reviewing AI output the way they review staff output finds it much harder to ferret out where the risk lies. Finding the areas that require scrutiny is much harder precisely because the output looks excellent everywhere, including in the spots that are wrong.
This is a challenge for the most seasoned of managers. Reviewing work has never been a choice. It is a critical feature of every workflow, and it works efficiently in firms because everyone in the company can be honest about their uncertainty. The new labor is not.
As AI scales the problem of review reliability becomes difficult to manage
Imagine a system that is right roughly nine times out of ten. That sounds, in the abstract, like a win, and inside a pilot it usually feels like one. The pilot is small. A reviewer can look at everything. The cost of looking at everything is hidden by the size of the sample. Now move to production. Errors no longer cluster anywhere detectable, because the signal that used to make them cluster is gone. Finding the tenth requires reviewing all 10. Verification becomes the new labor. At production volume, a “successful” deployment can cost as much as the labor it displaced. The cost didn’t disappear. It moved downstream to the review layer. This is why some firms feel busier, not lighter, after adding AI.
The way out is not more review. The way out is applying the right tool to the job. Most tasks have a single, verifiable answer. If so, a deterministic, rules-based system should be used. That system might or might not have an AI bolted onto it (a hybrid AI system), but at the core the deterministic system does the work. The output is checkable by construction, and the review burden for that stage collapses into the rule itself. If there isn’t a clear, easily verifiable answer to a question, then judgment is the actual deliverable. In that case, AI is a good solution but needs to be paired tightly with human review. The key is not to separate AI work from human work, but to separate work with verifiable answers from work that requires judgment. Then select the tooling to match.
Let’s consider a few examples of matching the right tool to the job:
Tax is dense with single-right-answer stages: classification of documents received, data extraction from forms, data entry into the tax system, and so on. Each of these is a simple unit of work that is verifiable and needs to work the same way for every single job.
Rule-based systems are the clear right answer for each. Advisory is the inverse. Unstructured judgment is the product. Over-constraining the model and applying a static rule set will lead to cookie-cutter answers with no finesse. In these cases, the deliverable is the synthesis, not a lookup. Bookkeeping and CAS sit in between, and are mostly deterministic with narrow pockets of judgment. The right call for most firms will be to employ a combination of both AI and rules-based systems.
As AI continues to grab all of the headlines, prudent firms will judge systems on their merit and fit rather than flash. Given the nature of accounting work, this will lean more toward deterministic, rule-based systems than toward pure AI. Hybrid systems that layer AI on top of rule-based systems offer the best of both worlds, fusing the ease of AI with the predictability and accuracy that accountants require.
In the next installment of this series, we look at AI governance and how to apply existing governance processes to AI and hybrid AI.
Credit: Source link
