Automating the tax provision with AI support

The tax provision is a data assembly problem wearing an accounting hat. You pull the trial balance, sort book-tax differences into permanent and temporary, rate the temporaries, roll deferred tax balances forward, and write the memos that explain each position. AI shortens the assembly: extraction, classification, reconciliation and first-draft documentation. It does not decide the tax positions, and a preparer still owns the number that lands in the financial statements.

Most of the pain at quarter-end is not the judgment. The judgment happens near the end, when a tax accountant looks at a schedule and decides whether a reserve is deductible this year or next. The hours before that are spent moving numbers between the general ledger, fixed-asset systems, prior-period workpapers and a stack of spreadsheets that each report the same balance in a slightly different shape. That is the copying-and-reconciling layer, and it is where automation earns its place. It is also where a confident wrong number does the most damage, because a mis-tagged difference flows straight into the effective tax rate and nobody catches it until the rate looks strange.

Sort the differences before you rate them

The book-tax difference schedule is the spine of an ASC 740 provision, and it is mostly classification work. Each adjustment has to be tagged permanent or temporary, mapped to an account, and tied to a source. A model trained on your own historical workpapers can do the first pass here, because the mapping is stable quarter over quarter. Depreciation is a temporary difference every time. Meals and entertainment land in the same permanent bucket. The accrual reversals follow a pattern the model has seen across many prior quarters.

• Tag each general-ledger movement against how it was treated in prior periods, and flag any account whose classification the model wants to change. A change is either a real event or a mistake, and both deserve a preparer’s eyes.
• Resolve the same account across systems before you sum it. Entity resolution matters here: the fixed-asset ledger, the GL and last quarter’s workpaper may name the same balance three ways, and reconciling them by hand is where hours disappear.
• Keep a false-positive budget in mind. A sorter that flags every line as suspicious is as useless as one that flags nothing. Tune it so the review queue is short enough that a person actually reads it.

The model should never silently overwrite a prior classification. When it disagrees with how something was treated last quarter, that disagreement is the output worth surfacing. Point-in-time correctness applies: the schedule has to reflect the balances and rules as they stood at the reporting date, not as they read today after a later adjustment. If your extraction pulls a restated figure into a closed period, the rate reconciles to nothing.

Extraction is the input, and it lies quietly

A large share of provision inputs live in documents, not systems. Apportionment data, statutory rates by jurisdiction, tax-attribute carryforwards, and the numbers buried in prior returns and notices. Pulling these reliably is the difference between a provision that reconciles and one that spends a day chasing a variance.

The failure here is rarely a document the model cannot read. It is a document it reads confidently and wrong: a rate transposed, a jurisdiction mismatched, a figure lifted from the wrong column of a table that looked clean. Extraction needs a confidence signal on every field, and low-confidence pulls need to route to a human before they enter a workpaper. Build an eval set from real historical documents with known answers, including the ugly scans and the multi-column tables, and measure field-level accuracy rather than a single document score. Watch for drift when a filing template changes or a new jurisdiction enters the mix, because a model tuned on last year’s forms degrades quietly on this year’s.

Where the extraction feeds a number that hits the effective tax rate, put a reconciliation check between the pull and the workpaper. A carryforward that does not tie to the prior return, or a rate outside the plausible range for its jurisdiction, should stop the pipeline rather than flow through it.

Documentation is where the audit trail is won

The provision is not finished when the numbers foot. It is finished when the workpapers explain the numbers well enough that a reviewer and, later, an auditor can follow each position back to its source. AI writes strong first drafts of this documentation, and that is a real saving because the writing is repetitive and the deadline is fixed.

• Draft the memo for each material position from the schedule and the supporting documents, with the amount, the treatment and the citation the model relied on. A preparer edits and signs; the model does not get the last word.
• Carry lineage on every figure. A number in a workpaper should link to the trial-balance line or the source document it came from, so a reviewer can verify without re-deriving. This is the same discipline that makes the close and the audit go faster.
• Log every classification change with a reason attached. A deferred balance that moved because a rate changed reads differently from one that moved because someone re-tagged an account, and the audit trail has to tell them apart.

The line to hold is simple. Straight-through processing is fine for the mechanical roll-forward of a stable deferred balance below a materiality threshold you set, with sampling on the output. Anything that moves the effective tax rate, changes a position, or touches an uncertain tax position surfaces as a proposal a preparer approves before it becomes final. The model compresses the hours spent assembling the provision. It does not sign it, and the person who does needs every number to carry its source, its history and a reason it reads the way it does.