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.
| Step | What happened | Record it leaves | Annotation |
|---|---|---|---|
| 1. Starting state | Invoice for 120 units meets a receipt for 100 | Invoice, purchase order and receipt lines | The mismatch is the task: 20 units billed but not received |
| 2. Hold placed | Validation stops the invoice | Matching hold with reason and timestamp | System action, no human judgment yet |
| 3. Investigation | AP clerk asks the warehouse whether 20 more units arrived | Exception note or email thread | The clerk could have asked the supplier for a credit instead; this is the decision point |
| 4. Correction | Warehouse finds a second delivery note and records the receipt for 20 units | Receipt line with user, date and attached delivery note | Authority matters: the warehouse, not AP, records receipts |
| 5. Release and posting | Revalidation clears the hold; invoice posts for $2,220 | Release record and posted voucher | Acceptance 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.
| Field | Where it sits in the example | Why the trajectory depends on it |
|---|---|---|
| Purchase order lines | 120 units at $18.50 | The agreed starting point |
| Receipt lines with user and date | 100 units, then 20 more at step 4 | The physical facts the invoice is checked against |
| Invoice lines | 120 units, $2,220 | The claim being tested |
| Hold type and reason, with timestamp | Matching hold at step 2 | Shows the control fired and why |
| Exception note or thread | Clerk's question to the warehouse at step 3 | The reasoning a model would need to imitate |
| Attached delivery note | Second delivery note at step 4 | Evidence behind the correction, not just the correction |
| Release record | Revalidation at step 5 | Shows the hold cleared by rule, not by override |
| Posted voucher | Final posting at step 5 | The 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
- Questions to ask a client about golden trajectories: a short discovery script for checking whether a company's exception records could qualify.
- Record gaps that make golden trajectories hard to assess: a guide to the missing middle steps, silent overrides and unclear authority that weaken a trajectory.
- Golden trajectories review packet: a list of the metadata a company prepares before SourceX reviews its exception history.
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.