Accounting and reconciliation datasets for AI training

An accounting dataset is the working history of a company's books: transactions with their general ledger coding and later corrections, bank and card reconciliations, accruals, close checklists, supporting workpapers and the approvals behind each entry. SourceX sources these from in-house finance teams and accounting firms on systems such as NetSuite, QuickBooks, Xero, Sage Intacct and SAP. Histories typically span several fiscal years, and vendor and client names are removed and amounts masked as agreed.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Coded transactions, reconciliations and close records with reviewer corrections and approvals
Typical systems: NetSuite, QuickBooks, Xero, Sage Intacct, SAP; close tools such as BlackLine
Typical history: Several fiscal years; varies by partner
Modality: Ledger and bank tables plus workpapers, checklists and supporting documents
Delivery formats: Agreed per order; ledger tables in Parquet, with workpapers as linked files
Preparation: Vendor, client and employee names removed; amounts and account numbers masked as agreed
Licensing: Firm-held client data needs client consent; non-exclusive or exclusive snapshot
Availability: Depends on partners and, for accounting firms, client consent; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
entity_id	string	Pseudonymous ID for the company, or for each client of an accounting firm, kept separate throughout.
period	string	Fiscal period the record belongs to, with the entity's fiscal year-end.
chart_of_accounts	object	The entity's accounts with types and a mapping to a standard taxonomy, so different charts can be compared.
transactions	array	Bank, card and ledger lines with date, de-identified description, amount, currency and source.
coding_history	array	Every category assigned to a line, by whom or by which rule, and when it was changed.
reconciliation	object	Statement and book balances, matches, reconciling items, adjusting entries and the final difference.
reconciling_items	array	Outstanding deposits, uncleared payments, unrecorded fees and errors, each with how and when it cleared.
journal_entries	array	Manual and adjusting entries with lines, memo, preparer role, reviewer role and reversal links.
accruals	array	Accruals and prepaids with their basis, such as purchase order, contract or estimate, and the reversal or true-up.
close_checklist	array	Close tasks with owner role, due day relative to period end, completion, sign-off and notes.
approvals	array	Approval chain steps with role, decision, timestamp and threshold rule.
workpapers	array	Supporting schedules and documents linked to the entries they support, de-identified.
adjustments_after_review	array	Entries changed after review or by external accountants or auditors, with reason codes.

Example record

{
  "entity_id": "ent_31f0",
  "period": "2023-07",
  "fiscal_year_end": "12-31",
  "account": { "id": "acct_1010", "type": "bank", "std_map": "cash_and_equivalents" },
  "reconciliation": {
    "statement_balance": 184220.55, "book_balance": 186593.30,
    "adjusted_bank": 186548.30, "adjusted_book": 186548.30, "difference": 0.00,
    "prepared": { "by": "role_staff_acct", "at": "2023-08-03T16:20:00Z" },
    "reviewed": { "by": "role_controller", "at": "2023-08-04T10:05:00Z", "decision": "approved" }
  },
  "reconciling_items": [
    { "type": "deposit_in_transit", "amount": 5027.75, "cleared": "2023-08-01" },
    { "type": "outstanding_payment", "ref": "pmt_8812", "amount": -2700.00, "cleared": "2023-08-09" },
    { "type": "unrecorded_bank_fee", "amount": -45.00, "resolution": "je_0731_04" }
  ],
  "journal_entries": [
    { "id": "je_0731_04", "type": "adjusting", "memo": "Record July wire fees",
      "lines": [ { "acct": "acct_6420", "dr": 45.00 }, { "acct": "acct_1010", "cr": 45.00 } ] }
  ],
  "coding_history": [
    { "txn": "txn_77a2", "description": "[VENDOR_14] ANNUAL RENEWAL", "amount": -18000.00,
      "events": [
        { "at": "2023-07-12", "by": "bank_rule_22", "set": "acct_6110_software_expense" },
        { "at": "2023-08-02", "by": "role_controller", "set": "acct_1410_prepaid_expenses",
          "reason": "12-month term; amortize monthly" }
      ] }
  ],
  "close_checklist": [ { "task": "bank_rec_operating", "due": "BD+3", "status": "complete",
                         "signoff": "role_controller" } ]
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train transaction categorization with corrections

Coding histories show the first category, whether a rule, a model or a person chose it, and the category an accountant settled on, which turns correction events into labels.

Build reconciliation agents

Statement and book data with reconciling items and resolving entries show how mismatches were matched, explained and cleared in practice.

Evaluate close and review copilots

Checklists, workpapers and post-review adjustments give graded tasks with known answers, such as whether an accrual was needed and what the entry should have been.

Detect anomalies and errors

Reversed entries, auditor adjustments and reclassifications label real errors, a ground truth that synthetic ledgers cannot provide.

Model approval routing

Approval chains with thresholds show which entries needed a controller or CFO sign-off, and when approvers sent work back.

Use-case guides: Finance and accounting agents, Private evaluation sets

What makes this data valuable

Correction history

Original and revised coding, with who changed it and why, not only the final ledger.

Closed periods

Fiscal years that went through year-end and, where applicable, external review or audit adjustments.

Linked support

Workpapers and source documents tied to the entries they justify.

Comparable charts

Account mappings to a standard taxonomy, so entities with different charts of accounts can be combined.

Reconciling detail

Individual reconciling items and how they cleared, rather than a single difference figure.

Explicit preparer and reviewer roles

Segregation of duties is visible, so a model can learn which steps need a second person.

Why the corrections matter more than the ledger

A general ledger records where every transaction ended up. It does not record how it got there: the bank rule that coded a software renewal as an expense, the controller who moved it to prepaids two weeks later, the accrual booked because an invoice had not arrived, the reconciling item that turned out to be a duplicate payment. Those changes are the expertise an accounting agent has to learn, and they exist only in audit trails, close tools and workpapers, not in the financial statements.

Synthetic ledgers can be internally consistent, but they encode the generator's idea of what a business looks like. Real books carry the mess that makes the work hard: inconsistent vendor descriptions, transactions split across periods, chart-of-accounts changes mid-year, and judgment calls on materiality that differ between a small single-entity business and a group that consolidates many entities.

What breaks when ledgers are combined

Books from different companies do not line up on their own. The same software subscription lands in different accounts, under different numbers, in every chart of accounts, so a model trained on raw account codes learns each company's numbering rather than the accounting. A mapping to a common taxonomy has to travel with the data, and the original accounts should stay alongside it so the mapping can be audited.

Timing causes quieter problems. Open periods still contain coding that review has not reached, so training on them teaches the errors the close would have fixed. Accruals that reverse every month create mirrored entries that look like independent transactions unless reversal links are kept. Size leaks identity: exact revenue, payroll or one unusually large customer payment can point to a company even with every name removed, which makes amount handling a scoping decision rather than a formatting detail.

Requests are therefore framed by entity count, systems, closed fiscal years, industries and the workflows that matter to you, whether coding, reconciliation, close or approvals. When the partner is an accounting firm, client-level authorization sets the outer limit, so scope is often agreed client by client.

What to check before licensing

When the source is an accounting firm, confirm client consent for each client entity in scope. Professional rules restrict disclosure of confidential client information without consent.
For US firms that prepare tax returns, confirm whether any data counts as tax return information. Section 7216 of the Internal Revenue Code restricts preparers' use and disclosure of it without the taxpayer's signed consent.
Agree how amounts are handled: exact, scaled by a constant per entity, bucketed or masked. Check that reconciliations still tie after the transformation.
Check whether bank descriptions and memos still contain names, invoice numbers or account numbers after de-identification. They are the leakiest fields in a ledger.
Confirm which periods are closed and final. Open periods still contain unreviewed coding that will change.
Ask how many entities and industries the data spans. Many years from one company teach one chart of accounts and one set of policies.
Confirm whether external auditors' workpapers are included. They generally belong to the audit firm, not the audited company, so they need the firm's authorization as well as the client's.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can accounting firms license client bookkeeping data for AI training?

Only with authorization from the clients whose data it is. An accounting firm's ledgers, reconciliations and workpapers mostly contain confidential client information, and professional rules for accountants restrict disclosing it without client consent. Firms that participate typically obtain consent for an agreed set of clients, and data for any client without consent stays out of scope.

How are amounts and account numbers protected?

Account and card numbers are removed or replaced with pseudonymous references. Amounts are handled as agreed per order: kept exact where the partner allows it, scaled by a constant factor per entity so ratios and reconciliations still tie, bucketed, or masked. Scaling keeps the arithmetic intact while hiding the absolute size of the business.

Does the data include who corrected a categorization and why?

Usually, where the system keeps an audit trail. Accounting and ERP systems generally record who changed a transaction's account or class and when, and many entries carry a memo explaining the change. Reasons are not always written down, so a correction often shows the before and after without an explanation. Coverage is checked on the sample.

How many fiscal years are typical?

Several fiscal years, depending on the partner and when its current system went live. A migration from one accounting system to another often marks the start of detailed history, with only opening balances carried over. Closed years that have been through year-end adjustments are more reliable than the current, open year.

Are audit workpapers available?

Sometimes, with limits. Workpapers prepared by a company's own finance team belong to that company. Workpapers prepared by an external auditor generally belong to the audit firm, which also owes confidentiality to the audited client, so both would need to agree. Expect internal close and review workpapers far more often than external audit files.

Can I get data from one ERP system across many companies?

It depends on supply. Partners that run several entities on one ERP, or accounting firms serving many clients in the same small-business system, can provide many charts of accounts in one source format. For coverage across ERPs and industries, list the systems and industries you need in the request; whether they can be combined depends on which partners run them and agree to license.

Real-world spreadsheets and financial models
Working Excel and Google Sheets files with formulas, links and version history
Enterprise workflow and task execution histories
Linked task trajectories from request to outcome, across every tool the work touched
Human feedback and QA-scored work
Work items with scores, verdicts and corrections from the people who reviewed them
Enterprise document archives
A company's working files with folders, versions and sharing metadata

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.