Your experts' corrections are data labs can't easily make for themselves.

Every time your team rejects, edits or overrides the model's output, that's the valuable part. The model's guess is everywhere. What your expert changed it to, and why, is not.

If you run an AI product for finance back office work, or an AI agent product, this page is about licensing expert correction data from your AI product to AI labs. Pillars plans to turn those corrections into graded tasks. Below is what we'd take, what stays yours, what a lab would get, and what we won't touch.

Finance back office · AP invoice codingExample (synthetic)

Starting workspace

  • invoice_example-co.pdfSynthetic vendor invoice for Example Co.: annual software subscription, billed up front
  • chart_of_accounts.csvExample Co.'s chart of accounts

Task

Code this invoice to the right GL account.

reads

AI draft

Codes the subscription to: struck through: Office supplies

Expert fix · AP clerk

Recodes it to software expense and splits out the prepaid portion.

"annual subscription paid up front; only the current period is expense."

becomes

Graded task

Task prompt and workspace files

Expert fix, kept as the reference

Rubric items below

Rubric items the grader checks

from the workspace from the expert fix

  1. R1Correct GL account (software expense, not office supplies).
  2. R2Prepaid portion split out correctly.
  3. R3Note cites the reason for the recode and the split.
Figure 1. How one correction would become one graded task.Example (synthetic). Shows the planned format only; not customer data.

Where the value is

An AP clerk recodes a software subscription the model filed under office supplies, and splits out the prepaid part.

An analyst moves a number from one quarter to another because the model misread a restatement.

We're also scoping agent traces, coding included: runs where a person rejected, edited or reran what the agent did in finance or software work.

Agent traces · coding agentExample (synthetic)

Workspace

repo/A synthetic repo with one failing test.

src/ tests/

Agent run

Coding agent

  1. Reads the code
  2. Runs the tests
  3. Edits the test file so it passes
  4. struck through: Tests pass

Edits the test instead of fixing the bug.

Expert fix

The developer reviews the run and writes the fix.

DeveloperRejects the run

Fixes the function the test covers.

Rubric

Rubric items the grader checks

  1. R1Test file unchanged.
  2. R2Function fixed.
  3. R3All tests pass.
Figure 2. How a rejected agent run would become a graded task.Example (synthetic). Shows the planned format only; not customer data.

What we take

One slice. One function, one product line, one period of time.

Not your product. Not your customer relationships.

We look for the sessions where a model failed or an expert corrected it on something hard, and drop the ones that only failed because something else broke.

What stays yours

Your product, your customers and your relationships with them.

You can sell elsewhere, you pay nothing up front, and you approve every sample.

What a lab would receive

Graded tasks, not a raw dump of your sessions.

We scope real work into tasks with hidden graders and reference answers. We plan to build each grader from the version the expert finally accepted, and to use that version as the reference answer. The parts the expert changed would become the core of what the grader checks.

An evaluation sample comes first. Samples go to named teams under evaluation-only terms, with a per-team copy we can trace. Graders and held-out splits are never shipped in training data.

Finance back office · restatementExample (synthetic)

Graded task

example

Task

Book the revenue adjustment in the right quarter.

Files

  • quarterly_model.xlsxExample Co.'s quarterly model
  • restatement_note.pdfA restated filing note for Example Co.

AI draft

Books the revenue adjustment in: struck through: the later quarter

Reference

Expert fix · Analyst

Moves it to the earlier quarter the restatement covers, and notes why.

Rubric

R1Adjustment sits in the restated quarter.
R2The later quarter's figures are unchanged.
R3The note cites the restatement as the reason.
What the model sees
What gets corrected
What the grader holds
Figure 3. What one graded task would contain.Example (synthetic). Shows the planned format only; not customer data.

What we need

Thirty minutes with whoever owns the data. Then one month of one slice.

The first call covers what your product does, what it logs, and whether your users correct it. No raw data changes hands at this stage.

It fits if you have an AI product with real usage in finance back office work, or an AI agent product with real usage; logged sessions; some record of whether the output was accepted, edited or corrected; and terms that let you license it, or could.

On price: we don't quote a number before we know the slice.

What we won't touch

  • Tax return prep.
  • Raw card and bank numbers.
  • MNPI and trading data.
  • Identifiable third-party financials.

Rights questions

What if your terms or customer contracts say no training?

Then that is the first thing the rights review reads. We check your terms of service and privacy policy for the whole period, your customer contracts and your model-provider terms before anything moves. Sometimes the answer is "not yet" or "not this part of the data." Rights are reviewed with counsel before anything moves.

Do your model provider's terms matter?

Yes. Every correction starts with an AI draft. If that draft came from a hosted model, we check your terms of service and privacy policy for the whole period, your customer contracts and your model-provider terms before anything moves (see OpenAI business terms, read Nov 16 2026). Some sessions may not be licensable for training. Rights are reviewed with counsel before anything moves.

How AI products in these lanes treat training on customer data.

Talk to us

Thirty minutes, one slice. hello@pillarsdata.com

← All research

Contact

We're early, and talking with AI companies and lab data teams now.

hello@pillarsdata.com

Started by Patrick Ojo · X · LinkedIn · Dallas, Texas