# Golden trajectories in enterprise workflows | SourceX

Canonical: https://refer.sourcex.si/guides/golden-trajectories-enterprise-data-referrals
Last updated: 2026-11-04

A golden trajectory is a well-demonstrated, checked path through a task: the starting context, each action, the tools used, what came back and the verified outcome, often with the reasoning behind each step. AI developers use them to show agents how a task should be done and to score whether an agent did it. For an introducer, the useful question is narrower: does a company keep connected records of real work with checked outcomes, and can someone authorize a licensing discussion? An archive alone does not qualify.

## What is a golden trajectory in an enterprise workflow?

In an enterprise workflow, a golden trajectory is one complete instance of a business task, done correctly and confirmed as correct, recorded in enough detail that each step can be followed. "Golden" signals trust: the path has been checked against a known good outcome, not merely logged.

Vendors who build agent training data describe the idea in similar terms. Appen describes golden trajectories as expert step-by-step demonstrations annotated with the reasoning behind each step, used as an imitation-learning starting point before reinforcement learning, and notes that it collects several paths where a task has more than one valid solution. Toloka's tool-use benchmark extension pairs each test case with a task description, user instructions and a single correct golden trajectory. Both uses point to the same requirement: the outcome must be checked, and the steps must connect.

That is what separates this page from the broader definition of [agent trajectory data](/questions/agent-trajectory-data), which covers any recorded sequence, good or bad. A golden trajectory is the subset someone would hold up as the reference.

## One fictional task, step by step

Illustrative only. A 260-person industrial parts distributor handles a request to raise a customer's credit limit above policy. The names, systems and figures are fictional.

| Step | What happened | Where the record lives | What makes it trustworthy | Rights question to raise |
|---|---|---|---|---|
| 1. Context | Sales rep asks to raise a customer's limit for a large seasonal order | CRM opportunity note | Timestamped, linked to the account | Does the note contain customer personal data? |
| 2. Gather | Credit analyst pulls payment history and open invoices | ERP aging report | System-generated, not retyped | Is the ERP data the company's own? |
| 3. Tool use | Analyst checks an external credit report | Credit bureau portal | Report reference stored on the case | Does the bureau license forbid reuse of reports? |
| 4. Decide | Analyst recommends a temporary increase with a 60-day review | Approval workflow comment | Reason written at the time | Who authored the reasoning, and is it internal? |
| 5. Check | Finance controller approves; 60 days later the account is paid on time | Approval log and payment record | Outcome confirmed by a second person and by later events | None new; this is the step that makes the path golden |
| 6. Close | Limit reverts at review; case closed with a note | ERP and CRM | Closure linked back to step 1 | Can the linked chain be de-identified? |

Three observations from the example:

- **The chain is the asset, not any single file.** Step 4 alone is a comment. Linked to steps 1 through 6, it shows a decision, its reasons and whether it worked. The guide on [workflow context instead of isolated files](/guides/how-to-identify-useful-workflow-context-instead-of-isolated-files) covers this pattern.
- **The check is what makes it golden.** Step 5 holds both a second person's approval and a later outcome. Records without any check are still trajectories, but not golden ones.
- **The reasoning is the scarce part.** Scale AI's discussion of agent-ready data makes the point that much valuable enterprise knowledge is held in people's heads. Records where the person wrote down why, at the time, are more useful than records of clicks alone.

## Which company records look like golden trajectories?

Look for repeatable tasks where three things are recorded together: the request, the steps taken, and a confirmed outcome. Common candidates include:

| Record family | Typical system (plain-text label) | Where the "check" usually comes from |
|---|---|---|
| Support escalations | Zendesk, ServiceNow | Customer confirmation and ticket closure |
| Approval exceptions | ERP or workflow tool | A second approver, then later results |
| QA and inspection | Quality management system | Inspector sign-off and rework records |
| Engineering changes | Jira, PLM tool | Review approval and release outcome |
| Claims or disputes | Case management system | Final resolution and any appeal |

The [tool-use training data](/data-types/tool-use-training-data) page lists further examples where the steps include system lookups, and [why enterprise AI needs enterprise data](/proof/why-enterprise-ai-needs-enterprise-data) explains the demand side for owners who ask.

## How should a partner qualify the company?

Stay at metadata level. You recognize the pattern; you never collect the records.

1. **Which task repeats?** Name one workflow, such as credit exceptions, warranty claims or change orders.
2. **Are the steps linked?** Ask whether one case ID ties the request, the actions and the outcome together in one system or across two.
3. **Is there a check?** A second approver, a customer confirmation or a recorded result.
4. **How much and how long?** An approximate count of closed cases and the years covered.
5. **Who owns the records?** The company's own operational records are the easier case. Third-party reports, client-owned material and personal data need review.
6. **Who can authorize?** An owner or executive who can open a licensing discussion.

For SourceX, the company also needs 50+ full-time employees at peak (contractors excluded), a documented operating history, rights to license and an authorized contact. 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.

What stays unknown until SourceX reviews it: whether those records match current dataset demand, what the rights review finds, which records the company chooses to license and what it is offered. SourceX licenses approved business data directly from companies for AI training and evaluation and manages the agreement, rights review, data preparation, delivery and payout. It does not build agents or turn records into golden trajectories itself. The related explainer on [golden datasets](/questions/what-is-a-golden-dataset) covers how reviewed examples are used in evaluation.

## Limits

Not every archive is useful. Records with no outcome, steps scattered across personal inboxes, or material the company does not own may not support a license at all. We have no published figures on how many companies hold usable chains, so treat the checklist as a screen, not a prediction.

## 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 a company you know keeps linked, checked case histories and has an owner willing to talk, [register as a partner](/register), confirm the basics against [who qualifies](/who-qualifies) and read the [program terms](/terms); more background is in the [guides hub](/guides).

## Sources

- https://www.appen.com/agentic-ai-golden-trajectories
- https://toloka.ai/datasets/tau-bench-dataset-extension
- https://scale.com/blog/hitl-ep3-agent-ready-data
- https://refer.sourcex.si/terms
