Skip to content

AI uses for records

Accrued is not invoiced: a finance estimate uncertainty dataset

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A finance estimate uncertainty dataset must record each amount's status, estimated or confirmed, along with its basis, its as-of date, the evidence on file that day and the document that later confirmed it. Those fields decide whether a later difference is a revised estimate or an error. A large public finance benchmark barely tests this: only 16 of 16,546 TAT-QA questions pair an accrual or estimate with a revision.

Document stack

Accrued is not invoiced

This page covers accrued is not invoiced: a finance estimate uncertainty dataset.Illustrative

Key takeaways

  • An accrual is a dated judgment, not a fact; the dataset has to keep the judgment, its basis and the later confirming document side by side.
  • Under IAS 8, a later difference is a change in estimate when new information caused it and an error when available information was ignored.
  • Tax applies its own rule to the same accrual, so every label needs to name the book it belongs to.
  • In TAT-QA, 35.5% of questions sit on text about estimates or accruals, but no question record stores whether an amount was estimated or when it was known.
  • An agent that states an estimate as an amount owed fails the task even when the invoice later matches its figure.

What should a finance estimate uncertainty dataset record about each amount?

It should record five things beside the number: whether the amount is estimated, confirmed or settled; the basis of the estimate, such as usage times a tariff or a historical claim rate; the as-of date; the documents on file that day; and the document that later confirmed the amount, with the difference. A ledger balance keeps only the number, which is why most finance data cannot teach this.

The task this serves is common and easy to get wrong. A controller asks an agent what the company owes a supplier for June, three days into July. The honest answer is an accrued estimate with a stated basis and the invoice still expected; a confident single figure presented as owed is a different answer, even if it happens to match the invoice later.

The rest of this guide shows why accounting rules make the evidence date decisive, what a public finance benchmark already tests, a casebook of six illustrative accruals, a rubric for reviewers and the fields to ask a data supplier for.

Why the evidence on file decides between a revised estimate and an error

Because accounting standards classify a later difference by what was knowable when the estimate was made, not by its size. IAS 8 defines accounting estimates as monetary amounts in financial statements that are subject to measurement uncertainty, and states that a change in an accounting estimate does not relate to prior periods and is not the correction of an error; such changes come from new information or new developments.

Errors are defined by the same evidence. IAS 8 describes prior period errors as omissions or misstatements arising from a failure to use, or misuse of, reliable information that was available when the statements for those periods were authorised for issue and that could reasonably be expected to have been obtained and considered. The same 650 difference can therefore be routine or a correction, and only the dated evidence trail tells which.

IAS 37 sets the target the estimate aims at. Paragraph 36 says a provision is the best estimate of the expenditure required to settle the present obligation at the end of the reporting period. A dataset that wants to grade whether an estimate was reasonable has to show what the preparer could see at that date, not what arrived a month later.

Companies that report under US GAAP follow its own codified rules, which this guide does not summarize. Record the framework each source company reports under as a field, and have reviewers label against that framework's text.

Tax keeps a separate clock for the same accrual

The same accrued liability can follow a different rule for tax, so a label that is right for the books can be wrong for the return. Under 26 CFR 1.461-1(a)(2), an accrual-method taxpayer generally takes a liability into account in the year all the events fixing it have occurred, its amount can be determined with reasonable accuracy and economic performance has occurred.

The regulation also handles uncertainty directly. If the exact amount cannot yet be determined, the portion that can be computed with reasonable accuracy may be taken into account, and when the exact amount is later fixed, the difference is taken into account in the later taxable year.

For a dataset, that means one more field: the book each figure and label belongs to. An agent that applies the tax true-up rule to a financial statement question, or a book reclassification to a tax question, has mixed two clocks, and the grading rubric should catch it.

What a public finance QA benchmark tests: 16 of 16,546 questions

TAT-QA, a public finance benchmark, tests reading and arithmetic on reported figures and almost never an amount's status. Built from real-world financial reports and released under CC BY 4.0, it holds 16,546 labeled questions on 2,756 table-and-text contexts in its train, development and labeled test files at commit 870accc; its README reports 16,552, the total when the unlabeled test file is counted instead, since the labeled test file holds six fewer questions.

Estimates and accruals appear throughout the source text. Of the 2,756 contexts, 463 (16.8%) contain an estimate word such as estimated, assumption or judgment, 694 (25.2%) contain an accrual-account word such as accrued, provision, allowance or reserve, and 978 (35.5%) contain one or the other. Those contexts carry 5,869 questions, also 35.5%.

The questions rarely touch them. Only 672 questions (4.1%) name an estimate or accrual term, and only 16 (0.1%) pair one with a revision word such as revised, adjusted, actual, settle or change in estimate. Reading all 16, each asks for a figure, assumption or reason the report already states, such as the amount of a change in estimate relating to prior periods for current taxes; none asks whether an amount was still an estimate or how a later difference should be classified.

The format explains why. Each question record stores the answer, its derivation, the answer type, whether it came from the table or the text, the relevant paragraphs, a comparison flag and a scale (labeled test records add derivation facts and mappings), but no field for status or as-of date. Across all questions, 42.1% are arithmetic, 43.1% single spans, 12.5% multi-span and 2.3% counts. A model can score well on that mix without ever saying the word estimate.

Nested squares drawn to area scale from the TAT-QA count: 16,546 labeled questions, 5,869 on contexts mentioning estimates or accruals, 672 that name such a term and 16 that also name a revision.
TAT-QA at commit 870accc, train, development and labeled test files; counts by SourceX with the keyword rules stated in the article.

Six accruals, one evidence test: the estimate casebook

Each case below fixes the estimate, its basis and as-of date, the document that later confirmed or contradicted it, and the label a qualified reviewer would expect. All amounts and facts are illustrative; they describe no company.

Cases 1 and 2 share the same numbers and differ only in what was on file, which is the point: a dataset that drops received dates cannot tell them apart.

Six accruals, one evidence test: the estimate casebook
CaseEstimate, basis and as-of dateWhat arrived laterExpected labelAnswer that fails
1. Utility accrual12,400 on June 30, from May usage at the current tariffJuly 12 invoice for 13,050, reflecting higher June usageChange in estimate: record 650 in July, no restatementRestates June or calls 12,400 an error
2. Same accrual, invoice on file12,400 on June 30; the June invoice for 13,050 arrived June 28Nothing new: the invoice was received before the June statements were approvedError: available reliable information was not usedTreats the 650 as a routine revision
3. Warranty provisionYear-end provision at a 2.0% expected claim rate from three years of claimsNext year's claims run at 2.6%Change in estimate, if 2.0% was the best estimate on the evidence then heldRebuilds the year-end figure with hindsight
4. Received, not invoiced8,000 accrued at purchase order price for goods received June 29July invoice for 9,200 that adds freightNeeds the contract: freight terms on file make it an error, otherwise a change in estimateLabels the difference without reading the contract
5. Tax treatment of an accrualPortion of a liability computed with reasonable accuracy at year-endExact amount fixed the following yearDifference taken into account in the later tax year, whatever the book labelApplies the book classification to the return
6. Agent asked before the invoiceOn July 3, the June utility accrual stands at 12,400 with no invoiceThe invoice later shows 13,050Answer 12,400 as an estimate, naming its basis and the expected invoiceStates a figure as owed, even one that later matches the invoice

Seven checks a reviewer runs on an estimate answer

Grade every answer on status and evidence first and on arithmetic last. Each check below is pass or fail with the evidence cited; two qualified accountants label independently, a third resolves disagreements, and every label records the rationale and the reviewer's role, not their name.

The seven checks, in grading order:

  • Status: every amount carries estimated, confirmed or settled as of the question date. Fail if an estimate is presented as owed or final.
  • Basis: the answer names the method and inputs behind an estimate, such as usage times tariff, claim rate or purchase order price, and their date.
  • Evidence date: the answer relies only on documents on file at the as-of date it is reasoning about. Fail any classification that leans on later documents.
  • Classification: a later difference is labeled change in estimate, error or needs more evidence, consistent with the evidence-date check.
  • Book: the answer names financial reporting or tax and applies that book's rule. Fail if the tax true-up rule is used for the books or the reverse.
  • History: the original estimate and the confirmed amount both remain visible. Fail if the answer overwrites the estimate with the outcome.
  • Authority: classifying a difference as an error, or proposing a correction, is routed to a named reviewer role. The agent proposes; it does not post.

When the estimate later matches the invoice, was the answer right?

Not necessarily, because a correct number reached without the right status, basis or evidence is not correct treatment. Case 6 shows it: an agent that announces 13,050 as owed on July 3 has guessed, and its later match with the invoice does not make the statement true when it was made.

Auditors face the mirror image and handle it with a time limit. PCAOB AS 2401 directs auditors to perform a retrospective review comparing the prior year's estimates with actual results, looking for possible management bias, and adds that the review is not intended to call into question judgments made in the prior year on the information available at the time. Hindsight is evidence about a pattern, not a verdict on a single estimate.

The grading consequence is direct. Compare outcomes with estimates across many cases to detect bias in a model or a preparer, but grade each answer against the evidence on file at its own date. A rubric that rewards matching the final invoice teaches a model to sound certain.

Shift dates together: preparing estimate packets without losing the test

Build each training or evaluation item as an authorized, de-identified task packet that keeps the relative order of every date, because the evidence-date check depends on it. A packet holds the estimate entry with its memo, the supporting calculation, each document with its received date, the confirming document, any posted adjustment and the reviewer's notes.

When dates must be masked, shift every date in a packet by the same offset, never independently, so that an invoice received two days before approval still arrives two days before approval. Remove names of employees and supplier contacts, keep the account, amount and document type, and record which accounting framework and which book each figure belongs to.

Selection rules matter as much as fields. Record how cases were chosen, how many estimates were reviewed to find them, and how many were set aside because the evidence trail was incomplete; a casebook built only from clean trails will overstate how often an agent can decide.

What to put in a data request for estimate histories

Ask for estimate-and-confirmation histories, not ledger extracts: accruals and provisions with their basis, as-of dates, received dates for every supporting document, the confirming document, the adjustment and the reviewer's classification, plus the framework and book. State the intended use and the acceptance test, for example the seven checks above scored by two accountants on a sample.

SourceX licenses business data from companies and licenses it to buyers in its own name, and datasets are found for each request rather than held in stock. Its request form collects what a buyer needs, from scope and volume to language, format, deadline and license terms, and the supplying company agrees the scope and terms and signs before anything is delivered. Whether a given company's records hold complete evidence trails is something only its own review can show.

Counting rules for the TAT-QA figures, exactly as bin/tatqa_count.py applies them: the population is every question in the train, development and labeled test files; context text is all paragraph text plus every table cell; estimate words match the starts estimat, assumption, judgment and judgement; accrual words match accru, provision, allowance and reserve; revision words match actual, true-up, true up, settle, revis, final, confirm, change or changes in estimate, and adjust. The word differ is excluded because it matches difference, which in this dataset means subtracting two figures; including it would add 20 questions, which the script lists for review when run with its --list option. Matching ignores case and requires a word start. Percentages are rounded once, half up.

Not measured: how often real accruals change between estimate and confirmation, how agents score on the rubric, or inter-rater agreement on the casebook labels. Those need authorized company records and reviewer time, and results are not yet published.

Frequently asked questions

What is a finance estimate uncertainty dataset?

It is a set of finance records and tasks in which every amount carries its status, estimated, confirmed or settled, along with the basis of the estimate, its as-of date, the evidence on file then and the document that later confirmed it. It lets a model be graded on whether it keeps an estimate and a confirmed amount apart.

Is a difference between an accrual and the invoice an accounting error?

Not by itself. Under IAS 8, a difference caused by new information is a change in estimate; it is an error only when reliable information that was available, and could reasonably have been obtained, was not used. The dated evidence trail decides.

Do public finance benchmarks test estimate status?

Not much, at least in TAT-QA, which asks for figures already reported and the arithmetic on them. In it, 35.5% of questions sit on text about estimates or accruals, but only 16 of 16,546 pair such a term with a revision, and none asks whether an amount was still an estimate.

How should an AI agent report an amount that has not been invoiced?

As an estimate, with its basis and date and the document still expected, for example an accrual of 12,400 from May usage pending the June invoice. Stating it as owed is a status error even if the invoice later matches.

Does tax treat accrual estimates the same way as the financial statements?

No. Under 26 CFR 1.461-1(a)(2), a portion computed with reasonable accuracy may be taken into account and the difference goes into the later tax year once the exact amount is fixed. A dataset should record which book each figure and label belongs to.

Sources

  • IAS 8 defines accounting estimates as monetary amounts subject to measurement uncertainty, says a change in estimate is not the correction of an error, and defines prior period errors by reliable information available when statements were authorised for issue. Source
  • IAS 37 paragraph 36: a provision is the best estimate of the expenditure required to settle the present obligation at the end of the reporting period. Source
  • 26 CFR 1.461-1(a)(2) sets the all-events, reasonable-accuracy and economic-performance conditions for accrual-method liabilities, allows a reasonably computable portion, and puts the later difference into the later taxable year. Source
  • TAT-QA data files at commit 870accc: 2,756 labeled contexts and 16,546 questions, with the question fields and answer types counted for this article. Source
  • The TAT-QA README describes 16,552 questions on 2,757 hybrid contexts from real-world financial reports and states the dataset license, CC BY 4.0. Source
  • PCAOB AS 2401 calls for a retrospective review comparing prior-year estimates with actual results to assess possible bias, and says it is not intended to question judgments made on information available at the time. Source
  • SourceX's request-data page says it licenses business data from companies and licenses it to buyers in its own name, and that datasets are sourced on request, not held in stock. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies