# Agent recovery data in business workflows | SourceX

Canonical: https://refer.sourcex.si/guides/agent-recovery-data-business-workflows
Last updated: 2026-11-05

Agent recovery data is the record of a task that went wrong and was put right: the failed attempt, the signal that exposed it, the correction someone with authority approved, and evidence the fix held. Successful tasks show an AI agent the happy path; recovery records show it what to do after a wrong step. Field service callbacks, accounts payable holds, reopened support tickets and engineering rework often contain the full sequence.

## What is agent recovery data, and which business records can illustrate it?

Recovery data is a trajectory with a dip in it. The work starts, a step fails, the failure is detected, a corrected plan is approved and carried out, and the result is checked. Each of those moments leaves a record in an ordinary business system if the company kept it.

Agent builders care about this because their systems fail in the middle of tasks, not only at the start. Appen, which sells agent training and evaluation data, offers trajectories with failure annotations and classifies where agents break down, such as misplanning or pursuing the wrong sub-goal, to decide what data to collect next. Research points the same way: the Reflexion paper showed that language agents improved when feedback from a failed attempt was turned into a written lesson and carried into the next attempt. A business record of a failed job, its diagnosis and its fix is a human-written version of that loop.

Records that commonly hold recovery sequences:

| Workflow | Failure signal | Authorized correction | Verification |
|---|---|---|---|
| Field service | First visit closed as incomplete; wrong part on van | Supervisor approves substitute part and revisit | Completion signature and no callback within a set window |
| Accounts payable | Invoice placed on hold for price mismatch | Buyer approves revised price or credit note | Invoice released and paid; vendor statement reconciles |
| Customer support | Ticket reopened by the customer after closure | Escalation lead approves a different fix | Ticket closed again and not reopened |
| Software engineering | Deployment rolled back after an alert | Reviewer approves a revised change | Redeploy passes checks and the alert stays clear |
| Order fulfillment | Shipment refused or returned | Customer service authorizes reship or refund | Delivery confirmed or refund posted |

## The three segments of a recovery record

A useful recovery record has three segments, and each one answers a different question for the model developer.

1. **Initial failure.** What was attempted, what state the work was in, and how the failure surfaced. The signal matters: a customer complaint, a failed check, a hold code.
2. **Authorized correction.** Who diagnosed the cause, what alternative was chosen, and who approved it. Approval is the part most often missing, and it is what separates a real recovery from an improvised workaround.
3. **Verified recovery.** Evidence that the corrected result held: a later check, a period with no repeat, a reconciled balance.

Missing any segment weakens the record. Failure without correction is an error log. Correction without verification is a hope. Verification without the failure is a success story with the useful part removed.

## A worked example: one service job, failed and recovered

Illustrative. A fictional 240-person HVAC service company handles a commercial rooftop unit repair. All details are invented.

| Day | Record | System | Segment |
|---|---|---|---|
| 1 | Work order opened: unit not cooling; compressor fault suspected | Field service management | Starting state |
| 2 | Technician visit closed as incomplete: compressor on van does not fit this model | Mobile work order notes | Initial failure |
| 2 | Return-to-stock entry for the wrong part | Inventory | Initial failure |
| 3 | Service manager notes cause: model number mistyped at intake | Work order internal note | Diagnosis |
| 3 | Substitute compressor approved under warranty; revisit scheduled | Approval log and dispatch board | Authorized correction |
| 5 | Repair completed; customer signs off | Mobile work order | Verified recovery |
| 35 | No callback logged for the unit in 30 days | Field service management | Verified recovery |

Seven entries across four systems tell a complete recovery story. The day 3 entries are the part a model cannot guess: that the root cause sat in intake, not in the repair, and that the substitute needed a named approver. The day 35 check is what lets an evaluator call the recovery verified rather than claimed. The [field service histories page](/data-types/field-service-histories) covers the systems involved, and [root cause analysis records](/data-types/root-cause-analysis-records) cover the diagnosis step in more formal settings.

## How is this different from failure data, exceptions and corrections?

The guide to [why AI labs value records of mistakes and rework](/guides/failure-data-ai-training) explains why failed attempts are worth keeping at all. Recovery data, as this page uses the term, adds two tests: a named approval for the fix, and a later check showing the fix held. Exception handling records explain how a case outside the normal rules was decided; the [exception handling guide](/guides/exception-handling-ai-agents-data) covers them. Correction workflows, covered in [assessing human review records](/guides/how-to-assess-human-review-records-and-correction-workflows), focus on a reviewer fixing someone else's output. Recovery data is narrower and sequential: the same task, attempted, failed, re-planned and confirmed. One recovery record can contain an exception and a correction, but the test is whether the whole arc is linked to one job, ticket or invoice. The [agent trajectory data explainer](/questions/agent-trajectory-data) defines trajectories more generally.

## Company qualification questions

Keep discovery at metadata level. The partner learns whether the arc exists, not what any job said.

| Ask | Strong answer | Why it matters |
|---|---|---|
| When a job, ticket or invoice fails, is the rework linked to the original record? | Same ID, or a parent-child link | Without the link the arc cannot be rebuilt |
| Does the system record who approved the fix? | Named approver with timestamp | Separates authorized corrections from workarounds |
| Is there a later check, such as a callback window or reopen flag? | Yes, reported monthly | Supplies the verification segment |
| How many years of history are kept, and in which systems? | Five or more years in named tools | Depth and continuity |
| Who can authorize a licensing discussion? | The CEO, usually the owner in an owner-run company | The right sponsor |

The introduced company needs 50+ full-time employees at peak (contractors excluded), a documented operating history, rights to license the records 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.

Special situations can matter here. Investors and advisors handling a wind-down often hold years of rework history in systems about to be switched off; the [special situations investor page](/for/special-situations-investor-data-assets) covers that case.

What stays unknown until SourceX reviews it: current buyer interest in this record type, the rights review result, redaction treatment of customer and technician names, and any offer to the company.

## Limits and unanswered questions

SourceX has not published figures on how often recovery records are requested or what they are worth, and this page makes no such claim. The research cited shows that feedback from failures can improve agents; it does not show that any particular archive will be licensed. Failure records can also be sensitive. Callback notes may describe safety issues or customer disputes, and the company decides scope before anything is reviewed.

Company compensation and referral rewards are separate. The company is paid under its own agreement with SourceX; the partner's reward is calculated on what SourceX actually pays the company and is funded separately, so it never reduces the company's payment. Illustrative only: on a $120,000 payout, the reward would be $56,000. See [how rewards work](/rewards).

## 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 client links rework to the original job and records who approved the fix, [register as a partner](/register), check the company against [who qualifies](/who-qualifies) and read the [program terms](/terms) before introducing its decision-maker. More agent data topics are in the [guides hub](/guides).

## Sources

- https://www.appen.com/agentic-ai
- https://www.appen.com/agentic-ai-trajectory-analysis
- https://arxiv.org/abs/2303.11366
- https://refer.sourcex.si/terms
