Golden trajectories in a real workflow: one invoice exception from mismatch to posting

Short answer

A golden trajectory is a complete, checked record of a task done correctly, step by step: where it started, every action taken, who was allowed to take it, and a final check anyone can verify. In an ordinary finance workflow, it can be a single invoice that failed three-way matching, was held, investigated with the warehouse, corrected by the person authorized to record receipts, released and posted, with every step leaving a record and the posted result matching the goods received.

Visual briefFull-size diagram
A three-way match grid where the invoice quantity of 120 conflicts with a receipt of 100, followed by five numbered steps that end in a released hold and a posted invoice that passes its acceptance check.
Illustrative invoice exception at a fictional distributor. The trajectory is golden only because the final check passed and every step left a record.

What does golden trajectories mean in an ordinary business workflow?

It means more than a closed case. A golden trajectory keeps the middle of the task, the ordered steps and the reasons for them, and ends in a result that can be checked against the records rather than taken on trust.

The term comes from AI agent training. Appen, a data services company, describes golden trajectories as expert-demonstrated, step-by-step task completions, with each step annotated with its reasoning, that give an agent an imitation signal early in training. Appen also collects demonstrations across different solution paths when a task can be solved more than one way. A golden dataset is the related idea one level up: a reviewed set of examples with known answers, which can be built from trajectories like this one. Real trajectories also carry the detours and judgment calls that synthetic data rarely replaces.

Companies do not create these on purpose. They create them as a side effect of controlled processes. Accounts payable is a good place to look because the software forces structure: an invoice that does not match its purchase order is stopped, someone has to act, and the system records the release.

One invoice exception, step by step (Illustrative)

The case below is fictional. A 300-person building-products distributor buys 120 units of a fastener kit at $18.50 each on one purchase order. The invoice arrives for all 120 units, $2,220 in total. The receiving record shows only 100 units.

Three-way matching is the control that catches this. Microsoft's documentation for Dynamics 365 Finance describes three-way matching as comparing the invoice unit price with the purchase order price, and the product receipt quantity with the invoice quantity. Oracle Payables uses matching hold and release types for the same purpose, and an invoice cannot be paid until its holds are released. Dynamics 365 Finance can also be set so that posting an invoice with discrepancies requires approval, which is why the release step leaves its own record.

StepWhat happenedRecord it leavesAnnotation
1. Starting stateInvoice for 120 units meets a receipt for 100Invoice, purchase order and receipt linesThe mismatch is the task: 20 units billed but not received
2. Hold placedValidation stops the invoiceMatching hold with reason and timestampSystem action, no human judgment yet
3. InvestigationAP clerk asks the warehouse whether 20 more units arrivedException note or email threadThe clerk could have asked the supplier for a credit instead; this is the decision point
4. CorrectionWarehouse finds a second delivery note and records the receipt for 20 unitsReceipt line with user, date and attached delivery noteAuthority matters: the warehouse, not AP, records receipts
5. Release and postingRevalidation clears the hold; invoice posts for $2,220Release record and posted voucherAcceptance check: posted quantity equals received quantity, price equals PO price, no open variance

The final row is what makes this golden rather than merely complete. Anyone can compare the posted voucher with the receipt and the purchase order and confirm the outcome without trusting anyone's memory. If the warehouse had found nothing, the correct path would have been different: a supplier credit for 20 units, then posting for 100. A good set of trajectories includes both branches, because a real exception has more than one valid resolution.

The fields that make the trajectory checkable

The example qualifies as golden only because each step left a field a reviewer can read. These are the fields it relies on.

FieldWhere it sits in the exampleWhy the trajectory depends on it
Purchase order lines120 units at $18.50The agreed starting point
Receipt lines with user and date100 units, then 20 more at step 4The physical facts the invoice is checked against
Invoice lines120 units, $2,220The claim being tested
Hold type and reason, with timestampMatching hold at step 2Shows the control fired and why
Exception note or threadClerk's question to the warehouse at step 3The reasoning a model would need to imitate
Attached delivery noteSecond delivery note at step 4Evidence behind the correction, not just the correction
Release recordRevalidation at step 5Shows the hold cleared by rule, not by override
Posted voucherFinal posting at step 5The result the acceptance check compares

If the hold had been released by a manual override with no note, the posted result could still be right, but nobody could show that it was right for the right reason.

What the example shows, and what it cannot

  • One branch is not the task. The warehouse found the missing 20 units. Had it found nothing, the golden path would have been a supplier credit and posting for 100. A useful set includes both branches.
  • Logged is not golden. Every step here left a record. Many exceptions close with a one-line note, leaving the middle steps to memory.
  • Authority can be invisible. The example notes that the warehouse, not AP, records receipts. Where system roles are loosely set, the record may not show who was entitled to act.
  • A check is only as good as its fields. Matching posted quantity to received quantity works because both were recorded. A check that relies on a person's sign-off proves less.

Who owns these records, and what stays unknown

The AP team and controller own invoices, holds and postings; the warehouse owns receipts; IT administers the ERP. Supplier agreements that may limit use of pricing data sit with procurement or legal. Authority to discuss licensing usually rests with the CFO or CEO.

The records only matter if the company itself can license them. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. The company also needs 50 or more full-time employees at peak (contractors excluded), documented operating history, rights to license the records, and an authorized contact.

What stays unknown until SourceX reviews it: how many trajectories pass an acceptance check, whether supplier contracts limit use of pricing data, how personal data is handled, and whether any buyer wants this task type. Industry interest in golden trajectories does not mean every AP archive qualifies. Partners never request exports, screenshots or sample invoices. The owner can prepare with questions to ask before licensing company data, and can a company license its data to AI companies answers the first question most owners raise. For the wider exception family, see accounts payable exception handling; more concepts are in the data types hub.

Go deeper on golden trajectories

Next step

Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

If the controller confirms that exception holds, notes and releases are logged in the ERP and can name the executive who owns the records, register as a partner, confirm the company against who qualifies, and read the terms before you introduce them.

  1. Step 1Share your linkSend your personal link to a company you know.
  2. Step 2Company appliesThe company applies itself at /apply.
  3. Step 3Company signs and is paidThe company signs, delivers its data and is paid by SourceX under its agreement.
  4. Step 4You get your rewardthe greater of $50,000 plus 5% and 15% of the actual supplier payout, with no cap. 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.

Common questions

Is a golden trajectory the same as a golden dataset?

They are related but not the same. A golden dataset is a reviewed set of examples with known answers; a golden trajectory is one example where the whole sequence of steps, not just the final answer, is recorded and checked. A golden dataset can be built from golden trajectories.

Does every resolved exception count as a golden trajectory?

No. It needs the starting state, the ordered actions, who was allowed to take them, and a final check that someone can verify against the record. A resolved ticket with a one-line closing note usually lacks the middle steps.

Can I ask the company for a sample exception to show SourceX?

No. Partners stay at metadata level: which system, how many exceptions a year, how far back, and who owns the records. SourceX handles any record review under an agreement with the company.

Sources

  1. appen.com/agentic-ai-golden-trajectories
  2. docs.oracle.com/en/cloud/saas/financials/25b/fappp/invoice-h…
  3. learn.microsoft.com/en-us/dynamics365/finance/accounts-payable/t…
  4. learn.microsoft.com/en-us/training/modules/accounting-distributi…
  5. refer.sourcex.si/terms

Free resources

Know a company with valuable proprietary data?

Become a referral partner, get your link and introduce the company’s CEO (in an owner-run company, usually the owner). Referral partners can join from most countries. We can't work with partners in countries or regions under U.S. sanctions or on U.S. restricted-party lists. The companies you introduce should be based in the U.S., Canada and the EU, or in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia with at least 70% of their data in English.

Refer a company →

I own a business

Explore licensing your company's data directly to SourceX for AI training and evaluation. Start a short assessment; no uploads needed.

Start an assessment