Automation is moving from checklist status into preparation
The useful boundary is simple: an AI system may prepare a reconciliation package and challenge its evidence, but the qualified accountant remains responsible for treatment, resolution, review, and sign-off. Without that separation, faster preparation can become faster propagation of an incomplete or incorrect balance.
BlackLine made Verity Prepare generally available on July 27, 2026. Its first-party description shows where close automation is moving: ingest spreadsheets, CSV and text extracts, PDFs, images, and email; structure the data; execute groupings, aggregations, and calculations; and prepare account reconciliations. The company reports that early adopters reduced preparation time by up to 94%. That number is vendor-reported, does not establish performance across accounts or organizations, and should be validated in a controlled pilot.
The product release matters even for teams that never buy BlackLine. It turns an abstract “accounting agent” claim into a concrete operating question. If software can assemble the package, what evidence proves that it used the right population, period, currency, rules, and support? If the package is wrong, can a reviewer reproduce the path without trusting the model's narrative?
Practitioner comments explain why those questions matter. Accounting teams often keep preparation logic in inherited spreadsheets, then upload the workbook and support into a close platform. Users value centralized ownership, access, approval history, and a clean close checklist; they also report drag when reconciliation volume rises or when the platform adds clicks without replacing manual tie-outs. AI only helps if it removes mechanical preparation while preserving flexibility, evidence, and challenge.
PCAOB AS 1105 is written for auditors, not as a product specification for accounting teams. Still, its evidence principles are directly useful for management's preparation design: accuracy, completeness, precision, relevance, reliability, source understanding, recalculation, and reperformance. An AI-prepared package that cannot support those questions is not review-ready.
A reconciliation is not complete because the difference is zero. It is complete when the population is controlled, the rules are known, exceptions are resolved, and an authorized reviewer can reproduce the conclusion.
Use a six-layer close control architecture
1. AuthorityApproved tool, permitted accounts, data classification, named process owner, preparer, reviewer, and sign-off authority.
2. Frozen sourcesLedger and supporting populations tied to entity, account, period, currency, cutoff, extract parameters, hashes, counts, and totals.
3. Deterministic preparationVersioned matching, grouping, normalization, aging, and tolerance rules execute outside free-form model reasoning.
4. AI challengeThe model organizes evidence, finds contradictions, classifies exceptions, and drafts questions without creating accounting conclusions.
5. Independent reviewA qualified reviewer re-performs material and risk-selected paths, inspects support, and resolves or escalates exceptions.
6. Exact-version sign-offApproval binds one output version to one source set, exception disposition, reviewer, time, and authority.
Keep calculations and language in different lanes. Exact matching, amount totals, row counts, date arithmetic, currency conversion, threshold tests, and aging belong in deterministic code or controlled platform functions. An LLM can explain a rule or summarize an exception, but its prose should never be the authoritative calculation.
Likewise, keep preparation and certification separate. The IIA's Three Lines model distinguishes operational responsibility, oversight, and independent assurance. A practical close workflow applies the same separation at smaller scale: the preparer owns source and explanation quality; the reviewer challenges and re-performs; the controller or authorized owner resolves policy and signs off; internal audit remains independent.
| Role | May do | Must not do alone |
|---|
| AI preparation service | Structure data, execute approved tools, list exceptions, draft questions | Change source data, invent support, post entries, approve the account |
| Preparer | Confirm sources, explain items, correct evidence, propose resolution | Approve own reconciliation when independent review is required |
| Reviewer | Reperform, challenge, reject evidence, accept policy-permitted exceptions | Rely only on the AI summary or preparer inquiry |
| Controller/process owner | Own policy, materiality, exceptions, entries, close status | Delegate accountable sign-off to the model or vendor |
| Internal audit | Provide independent assurance on governance and controls | Operate the reconciliation workflow it later audits |
Freeze a source register before matching anything
A model can produce a convincing reconciliation from the wrong extract. Prevent that by building a source register first. The register should answer which system produced the population, who ran it, which parameters and cutoff were used, whether the file was modified, and how the reviewer can retrieve the same evidence.
source_id: BANK-USD-2026-07-31-01
entity: Example Holdings US
account: 101200 Cash - Operating
period: 2026-07
currency: USD
system: Approved Bank Portal
report: Transaction Detail
extract_timestamp_utc: 2026-08-01T02:14:22Z
parameters:
from: 2026-07-01
to: 2026-07-31
row_count: 1842
amount_total: 12844302.17
file_sha256: 5c4b...
manual_modification: false
owner: treasury.operations@example.com
retention_class: FIN-CLOSE-7Y
Then tie control totals before record matching. Confirm the general-ledger balance against the authoritative ledger report, not a copied cell. Confirm the subledger or external population against its own control total. Record signs, currency, duplicate keys, missing identifiers, blank dates, and excluded rows. A zero difference after uncontrolled filtering can hide an incomplete population.
AS 1105 paragraph .10A is particularly relevant to electronic information received from external sources and then provided by the company. For audits of fiscal years beginning on or after December 15, 2025, the PCAOB says auditors should understand the source and the company's process for receiving, maintaining, and processing the information, then test the information or relevant controls. Accounting teams can reduce downstream audit friction by retaining that chain during close preparation.
One period and cutoffAll sources identify the same period, timezone, business cutoff, and late-arriving transaction policy.
Immutable referenceHash the file or retain an immutable system report ID; do not overwrite extracts in place.
Control totalsCapture record count, debit, credit, net amount, currency, and authoritative balance before transformations.
Transformation logRecord filters, joins, mappings, conversions, deduplication, and manual edits with before/after totals.
Match records to assertions, not just to each other
Record matching answers a mechanical question. Reconciliation supports financial-statement assertions. The package should show which evidence supports existence or occurrence, completeness, valuation or allocation, rights and obligations, and presentation or disclosure. Not every assertion applies equally to every account, and the process owner should define the expected evidence.
| Assertion | Preparation check | Human challenge |
|---|
| Existence / occurrence | Trace recorded balance or transaction to independent or controlled supporting evidence. | Is the evidence authentic, current, and tied to this entity and period? |
| Completeness | Tie population counts and totals; inspect gaps, duplicates, cutoff, and excluded rows. | Could an omitted source, status, account, or late item understate the population? |
| Valuation / allocation | Recalculate amounts, rates, aging, and approved conversions. | Are assumptions and methods authorized and reasonable for the framework? |
| Rights / obligations | Link ownership, contract, counterparty, or legal-entity support. | Does the entity actually control the asset or owe the liability? |
| Presentation / disclosure | Compare mapping and classification to approved chart and policy references. | Is the item classified and described correctly, including required disclosure? |
Inquiry alone is weak evidence. If the preparer says an aged item will clear next month, retain the explanation as a claim and ask for support: settlement, invoice, bank activity, contract, subsequent receipt, or another policy-relevant artifact. The AI can formulate the question and link the answer; it must not convert confidence into evidence.
Recalculation and reperformance should remain visible. For material items and a risk-based sample, a reviewer should be able to follow source row, transformation, match rule, calculation, exception, support, and disposition. If the path exists only inside a vendor's model narrative, the control is not portable or independently testable.
Make the exception register the center of review
A useful system reduces routine preparation so humans can spend time on exceptions. It does not minimize the count by forcing ambiguous items into matches. Define exception classes, materiality, aging, ownership, due dates, escalation, and acceptable evidence before the run.
exception_id: REC-101200-2026-07-0047
class: unmatched-bank-item
amount: 84250.00 USD
age_days: 11
gl_source: GL-USD-2026-07-31-01#row-492
support_source: BANK-USD-2026-07-31-01#row-1802
assertions_at_risk: [completeness, cutoff]
evidence_present: [bank_reference, posting_date]
evidence_missing: [ledger_posting, owner_explanation]
proposed_owner: cash.accounting@example.com
due_at: 2026-08-03T16:00:00Z
status: OPEN
ai_observation: "Bank item has no ledger match under approved rules."
human_conclusion: null
Keep the AI observation narrow and factual. “Likely timing difference” may sound harmless, but it can bias the reviewer and conceal fraud, duplication, cutoff error, mapping failure, or missing source data. The model can say that the item has no match under rule version 3.2 and list candidate evidence to request. A qualified accountant decides the explanation.
Bind every correction to the evidence chain. If a mapping rule changes, rerun the complete affected population, update control totals, preserve the prior version, and show which exceptions opened or closed. Do not patch only the displayed item and leave the population in a mixed state.
Choose the first accounts by judgment, stability, and evidence quality
| Good pilot task | Poor first pilot | Reason |
|---|
| High-volume cash clearing with stable keys and daily feeds | Goodwill impairment | Deterministic matching is testable; impairment depends on significant judgment and assumptions. |
| Intercompany transaction matching before dispute review | Disputed intercompany settlement decision | AI can prepare pairs and exceptions; authorized owners decide disputes and treatment. |
| Prepaid expense roll-forward using approved schedules | New capitalization policy | Existing rules can be reperformed; policy creation requires qualified judgment. |
| Bank-to-ledger matching with controlled external evidence | Unstructured reserve estimate | Source and rules are observable; estimates require methods, assumptions, and specialist review. |
| Aging and stale-item challenge | Automatic write-off or journal posting | Flagging is reversible; posting changes the books and requires authority. |
Evaluate the pilot against a frozen baseline. Compare preparation time, review time, exception precision, missed exceptions, control-total failures, reopening rate, audit questions, and user overrides. A vendor-reported productivity percentage is not your control evidence. Your team needs results by account class, source quality, rule stability, and reviewer effort.
Set stop conditions before the pilot: source mismatch, unexplained control-total difference, unauthorized data exposure, unsupported transformation, missed material exception, incorrect assertion claim, excessive override rate, or inability to reproduce the output. A stop is useful evidence, not a failed transformation program.
Failure modes that create a fast but unauditable close
| Failure mode | What it hides | Control |
|---|
| Wrong period or entity | Plausible tie-out from mismatched populations | Intake gate and source-register key equality |
| Silent filtering | Rows removed to achieve a zero difference | Before/after counts and totals for every transformation |
| Net-zero comfort | Offsetting errors, duplicates, or incomplete populations | Gross debit/credit totals, assertion checks, duplicate and gap tests |
| AI-authored explanation | A guess presented as management evidence | Separate AI observation, preparer claim, support, and reviewer conclusion |
| Self-review | Preparation error survives because the same lane approves it | Independent reviewer identity and reperformance |
| Version drift | Reviewer signs a package different from the one produced or corrected | Immutable manifest and hash-bound approval |
| Broad tool access | Unauthorized files, entities, or posting capability | Least-privilege read access and separate human-controlled posting identity |
| Rule changes during close | Inconsistent populations and unapproved accounting logic | Version, owner, effective period, test set, and formal change approval |
| Evidence that cannot be exported | Vendor lock-in and weak independent challenge | Exportable source register, transformation log, exceptions, and results |
Privacy and retention also matter. Reconciliation packages may contain bank details, payroll, employee, customer, vendor, tax, and legal information. Use an approved tool, minimize fields, restrict accounts and entities, log access, define retention, and prevent training or cross-tenant use when organizational policy requires it. Do not paste sensitive close data into a consumer AI service because the workflow pack is convenient.
Pilot and release checklist
Pick a controlled account setChoose repeatable accounts with stable rules, reliable sources, known owners, and limited judgment.
Freeze baseline evidenceRetain prior manual preparation time, review time, exceptions, corrections, and audit questions.
Validate source ingestionTest every format, field, currency, sign convention, cutoff, row count, and control total.
Test rule determinismRun the same sources twice and require identical calculations, matches, exceptions, and manifests.
Seed known failuresUse duplicates, missing rows, stale items, wrong periods, unsupported mappings, and offsetting differences.
Reperform independentlyHave reviewers rebuild material and risk-selected paths from original sources without the AI narrative.
Review access and dutiesConfirm the tool cannot post entries, approve itself, change source data, or reach unrelated entities.
Bind exact-version approvalRecord sources, rules, output hash, preparer, reviewer, exceptions, time, and authority.
Monitor after releaseTrack overrides, reopened exceptions, missed items, rule changes, review time, incidents, and user complaints by account class.
Do not expand because the demo looks fluent. Expand when the evidence package is reproducible, reviewers find at least the same exceptions as the baseline, source-control failures stop the run, and total review effort falls without weakening accountability.
FAQ
Can AI approve an account reconciliation?
No. It can prepare, calculate through approved tools, organize, and challenge. A qualified preparer and independent reviewer remain responsible for treatment, resolution, and exact-version sign-off.
Which reconciliations are suitable for a pilot?
Start with repeatable, low-judgment accounts with stable sources, documented rules, reliable control totals, and known owners. Keep estimates, complex valuations, policy changes, disputed items, and material manual entries outside the first pilot.
What evidence should be retained?
Source identity, extract parameters and time, immutable references or hashes, counts and totals, transformation rules, assertion checks, matched and unmatched populations, exceptions, questions, corrections, and the exact version approved.
Does this replace audit procedures?
No. It is a management preparation workflow. Auditors determine the procedures required under applicable professional standards and the engagement's risks.