# GL coding and reclass histories as AI training labels

Canonical: https://refer.sourcex.si/data-types/gl-coding-history-ai-training-data
Last updated: 2026-10-20

GL coding history is useful AI training data because every coded invoice and every reviewer reclass is a labeled example with a correction signal. For CAS leads the licensor is the client company, not the firm: the firm introduces the client and checks AICPA and state fee rules first.

## Why do GL coding and reclass histories count as AI training labels?

Every invoice coded to an account and every reviewer correction is a labeled example: here is the transaction, here is the account a human chose, and here is the change a second person made. A company that has years of this has a large set of classification examples with a built-in correction signal, which is exactly what bookkeeping and close agents learn from.

For a CAS or bookkeeping lead, the supplier in this conversation is the client company and its own books. Your firm introduces the client. Your firm never licenses, exports or describes client ledgers, and the client decides everything about its own records.

## What do the records look like?

| Record | What it contains | Why it carries signal |
|---|---|---|
| Coded bills and receipts | Vendor, description, amount, account, class or department | The raw input-to-label pair |
| Bank-feed categorization | Rules, suggested categories, accepted or changed | Shows where automatic suggestions failed |
| Reclass journal entries | Original account, new account, preparer note | A correction with its reason |
| Chart-of-accounts changes | Merged, split or retired accounts, mapping notes | Explains why older labels differ from newer ones |
| Reviewer comments | Month-end notes, questions to the client | Context a transaction line does not carry |

The reclass is the high-value piece. A model that only sees final coding learns the average; one that sees the first guess and the fix learns where judgment is needed.

## Where do they live?

Mostly in the client's accounting system (QuickBooks, Xero, NetSuite, Sage or similar), your practice's workpapers, and the review comments in email or a task tool. Strong candidates have several years of unbroken history in one chart of accounts, or a documented crosswalk when it changed.

## What must a CPA check before raising it?

Start with your own rules, because this is a referral for which a fee may be paid. The AICPA Code's [commissions and referral fees rule (ET 1.520)](https://mn.gov/boa/assets/AICPA%20Code%20of%20Professional%20Conduct_tcm1233-764395.pdf) restricts accepting a commission for recommending a product or service to a client when the firm also performs certain attest work for that client, and permitted referral fees must be disclosed. The contingent fees rule (ET 1.510) sits alongside it.

State boards can be stricter than the AICPA Code. The [New Jersey Society of CPAs summary of commissions and contingent fees](https://njcpa.org/stayinformed/hubs/topics/commissions-and-contingent-fees) shows how a state can add its own limits. Confidentiality is a separate duty: client information is not yours to describe to a third party without consent. This is general information, not legal, tax or financial advice. Confirm with your own state board, firm risk partner or counsel before acting.

| Situation | What to check | Likely route to confirm |
|---|---|---|
| Client is an attest client | ET 1.520 and your state rule | Firm risk partner decides whether a referral fee is allowed |
| Client is CAS and tax only | State rule and written disclosure | Disclose any fee to the client in writing |
| Firm prefers no fee | Your firm policy | Introduce without accepting a reward |
| Client unsure | Whether it wants any outside party told | Let the client apply itself at sourcex.si/apply |

## The "clean books" screen

Use this inside a normal annual planning conversation with the owner.

- [ ] The books have 5+ years in a consistent chart of accounts or a documented mapping.
- [ ] Reviewer reclasses are visible as entries, not overwritten silently.
- [ ] Coding is done by people, not by a single outsourced pipeline the client does not control.
- [ ] Vendor and customer names exist in the system, which means redaction is a defined task.
- [ ] The business has 50+ full-time employees at peak (contractors excluded).
- [ ] Management is willing to consider a license with exclusivity for AI training.

A company that fails the last two does not qualify no matter how tidy the ledger is.

## How does the introduction work?

1. Ask the owner for permission to mention the idea. Do not open the books.
2. Share your referral link or submit the company through the referral form, giving basic fit information only.
3. SourceX qualifies the company on size, history, breadth and rights.
4. The company completes a data inventory, listing systems and record types rather than contents. Its [data inventory builder](/tools/data-inventory-builder) can help.
5. Price and terms are agreed, SourceX review, and nothing is binding until the company signs.
6. On closing, SourceX arranges delivery under the executed agreement, with redaction scope agreed beforehand.

> What to say: "Your close process leaves a trail of coded transactions and the corrections your team made. Some AI developers license that kind of history. I cannot see or share your books, but I can introduce you to a company that explains how it works. Would you like that?"

## What are the common mistakes?

| Mistake | Why it hurts | Fix |
|---|---|---|
| Offering client ledgers yourself | Breaches confidentiality and rights | Introduce only; the client is the licensor |
| Quoting a fee to the client | Terms are agreed with the company, not promised | Say that nothing is binding until signed |
| Treating a bookkeeping outsourcer's client books as the firm's data | The data belongs to the clients | Look for the firm's own books, or client consent |
| Skipping fee disclosure | Rule and trust problem | Disclose in writing if a fee is permitted |

## How do rewards work?

A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details. The reward rate is 25% of Actual Supplier Payout. The cumulative cap is $100,000 per referred company. The partner reward is calculated on the actual supplier payout and never reduces what the company receives. No reward is guaranteed.

## Related records

GL histories pair well with [capex requests and post-investment reviews](/data-types/capex-request-post-implementation-review) and [SQL query history](/data-types/sql-query-logs-ai-training-data) when a client has a deep finance function. The [accountants referral page](/for/accountants) covers the firm-level view, and [exception handling records](/guides/exception-handling-ai-agents-data) explain why corrections matter. For background on the category, read [what AI training data is](/proof/what-is-ai-training-data).

## Next step

If your policy allows it and a client qualifies, [register as a partner](/register) and start with one introduction. Check the baseline on the [who qualifies](/who-qualifies) page first.

## Sources

- https://mn.gov/boa/assets/AICPA%20Code%20of%20Professional%20Conduct_tcm1233-764395.pdf
- https://njcpa.org/stayinformed/hubs/topics/commissions-and-contingent-fees