The question is usually framed too broadly. "Can you trust AI with the close" invites a yes or no answer when the useful answer is task-specific.
Some parts of the close are judgment work where a plausible suggestion that a human verifies saves real time. Other parts are arithmetic that must reconcile, be reproducible, and survive an auditor asking where a number came from. Applying the same technology to both is how firms get into trouble.
A spreadsheet formula error usually announces itself. The total does not tie, a balance sheet does not balance, a reconciliation leaves a residual. The error is visible because the system is deterministic and something visibly broke.
A language model does not break. It produces fluent, correctly formatted, plausible output regardless of whether the underlying claim is right. If it states that prepaid amortisation for the month is $4,200 when the schedule says $4,800, nothing in the output signals a problem. The number is in the right place, in the right format, with the right level of precision.
This is the core issue. In accounting, errors that look like errors are manageable. Errors that look like correct answers are not.
The governing question: if this step produces a wrong answer, would anything downstream visibly break? If yes, AI assistance is lower risk. If the wrong answer would flow silently into the financials, it needs deterministic logic.
Working through a typical month-end close, the split is clearer than the general debate suggests.
| Close task | Appropriate approach | Reason |
|---|---|---|
| Detecting candidate prepaids in the GL | AI-assisted, human confirmed | A suggestion queue with evidence. The accountant approves each item, so a false positive costs a click. |
| Calculating the amortisation schedule | Deterministic | Must reproduce exactly each period and tie to the balance. |
| Posting the journal entry | Deterministic | Feeds the financial statements directly. Must be idempotent and traceable. |
| Classifying an unfamiliar vendor | AI-assisted, human confirmed | Pattern matching on description text, reviewed before it affects the accounts. |
| Intercompany reconciliation | Deterministic | Matching must be exact and the difference must resolve to zero. |
| Consolidation and eliminations | Deterministic | Errors compound across entities and are difficult to trace after the fact. |
| Foreign currency translation | Deterministic | IAS 21 prescribes the method. There is no judgment to apply. |
| Balance sheet sign-off | Deterministic workflow, human judgment | The reconciliation is arithmetic; the sign-off is a professional attestation. |
| Summarising a lease agreement | AI-assisted | The accountant reads the summary against the document before applying a treatment. |
| Writing the variance narrative | Deterministic figures, human framing | The numbers must derive from the ledger. The framing is judgment. |
AI appears where there is a human checkpoint before the output affects anything, and where the cost of a wrong suggestion is a rejected suggestion.
Deterministic logic appears wherever output flows into the financial statements without further review, or where reproducibility is a requirement rather than a preference.
Detection is the clearest case of the first category. Scanning a general ledger for transactions that look like they should be on an amortisation schedule is exactly the kind of fuzzy pattern matching where a suggestion engine earns its place — provided it presents evidence and confidence, and provided a human confirms before anything posts.
Calculation is the clearest case of the second. Once the accountant has confirmed a $24,000 annual insurance premium starting in March, the monthly amortisation is $2,000 and there is nothing for a model to contribute.
A practical way to test any tool is to imagine the audit conversation.
An auditor selecting a sample of adjusting entries will ask what supports each one. An acceptable answer identifies a schedule, shows the calculation, and names the person who approved it with a timestamp. An unacceptable answer is that the system generated it.
The same standard applies in due diligence, where a buyer's accountant will select figures from your reporting and trace them. If the trail ends at "the model produced this," the number carries no weight and the credibility cost extends to figures that were correct.
One test worth running on any tool you are evaluating: process the same period twice and compare the output byte for byte.
Deterministic logic returns identical results. Anything that involves generation may return a different emphasis, a different set of highlighted items, or a different figure. In a controls environment, non-reproducibility is a finding in itself — it means the process cannot be re-performed, which is the basis on which most substantive testing works.
AI is a genuine productivity gain in accounting when it is pointed at detection, classification, summarisation, and drafting — all tasks where a professional reviews the output before it matters.
It is a liability when it is pointed at calculation, posting, consolidation, or any output that a third party will rely on without independent verification.
Vendors that describe their product as AI-powered without distinguishing these categories are either not thinking carefully about the distinction or are hoping you will not. Either way, the question to put to them is simple: which specific steps use generation, and which use fixed logic? A clear answer is a good sign. An evasive one tells you what you need to know.
AI can assist with parts of the close but should not own any step that must reconcile or be audited. Suitable tasks include suggesting classifications for unfamiliar transactions, summarising supporting documents, and drafting narrative that a human verifies. Unsuitable tasks include calculating amortisation, posting journal entries, computing consolidations, and determining account balances. These require reproducible logic because the output must tie and be defensible to an auditor.
The primary risk is undetectable error. A language model produces output that looks correct regardless of whether it is correct, so mistakes do not announce themselves the way a formula error or an out-of-balance journal does. Secondary risks include non-reproducibility, meaning the same period can produce different results on different runs, and lack of traceability, meaning a figure cannot be tied back to source records during audit or due diligence.
Yes, for the right tasks. AI is genuinely useful for research, document summarisation, drafting communications, and exploring unfamiliar accounting treatments as a starting point for verification. The discipline is to keep AI on tasks where a human checks the output before it matters, and to keep deterministic logic on anything that produces a number a third party will rely on.
Ask which specific tasks use probabilistic generation and which use fixed logic, and require the vendor to be specific. Then test three things: whether every output figure can be traced to source records, whether running the same period twice produces identical results, and whether the tool discloses the assumption behind any quantified recommendation. A vendor that cannot separate these clearly is marketing AI rather than applying it.
Fynease uses detection to find what needs scheduling, and deterministic logic for every calculation, journal entry, and consolidation. Every figure traces to source.
See Fynease Automate Start free trial