# Agent recovery data review packet contents | SourceX

Canonical: https://refer.sourcex.si/resources/agent-recovery-data-review-packet
Last updated: 2026-11-04

An agent recovery data review packet is a short set of metadata-only documents that lets SourceX judge a company's recovery records without seeing any of them. It holds eight items, in order: a cover sheet, the task specification, the failure definition, the allowed recovery actions, a map of linking fields, an acceptance check written in advance, a volume and span sheet, and rights notes. Each is prepared by the person who already knows the answer.

This contents list belongs to the guide to [agent recovery data in an ordinary workflow](/guides/concept-agent-recovery-data), which follows one failed chat transfer from breakage to verified fix. That guide explains what a recovery record is. This page is a working document: what the company assembles, in what order, and who holds the pen for each part.

## The packet in order

Read top to bottom, the packet moves from what the task is, to how a failure is recognized, to how a fix is proven, to how much exists and who may license it. A reviewer reads it in the same order, because each item depends on the one above.

| # | Document | What it shows | Who prepares it |
|---|---|---|---|
| 1 | Cover sheet | Company size, systems involved, years of history, named owner and contact | Owner or COO, with the introducer's help on the metadata |
| 2 | Task specification | The job the records describe, in one or two sentences | Head of support operations |
| 3 | Failure definition | The event that marks a task as broken, and where it is logged | Support operations with the bot or messaging platform admin |
| 4 | Allowed recovery actions | What staff may do to fix a case, with limits and approvers | Support operations, with finance for refund and credit limits |
| 5 | Linking-field map | The IDs that join a failed attempt to its repair across systems | Help desk or IT administrator |
| 6 | Acceptance check | The pass or fail test for a recovery, written before anyone looks at cases | Head of support operations, reviewed by quality assurance |
| 7 | Volume and span sheet | Failed cases per month, months of history, platform change dates | Help desk administrator |
| 8 | Rights and sensitivity notes | Personal data present, vendor terms, contracts that limit use | Legal counsel or the CFO |

Nothing in the table is a record. Every entry is a description, a count, a date or a name.

## What a good version of each document looks like

**1. Cover sheet.** One page. Full-time headcount at peak, the help desk and messaging tools in use, how far back each one's history runs, and who can authorize a licensing discussion. It is the only item an introducer helps with, and only from what the owner says.

**2. Task specification.** Plain and narrow. For the failed-handoff case: "Recover a customer request after a bot-to-human transfer did not complete, so the customer's issue is resolved and confirmed." A vague task ("handle unhappy customers") cannot be checked.

**3. Failure definition.** The exact event, named in the system's own terms. In the illustrative case it is a conversation where the bot passed or tried to pass control and no human accepted it. Zendesk's documentation describes how messaging conversation control can pass a conversation to the next configured integration or release control back to its default state; the failure definition says which of those recorded events counts as broken.

**4. Allowed recovery actions.** The menu staff choose from: agent outreach, a refund within a stated limit, escalation to a supervisor above that limit, a goodwill credit under policy. Limits matter because a recovery outside policy is a different kind of record.

**5. Linking-field map.** Conversation ID, ticket ID, customer account ID, refund or credit transaction ID, and which system holds each. If the conversation ID is not copied onto the follow-up ticket, the map should say so plainly.

**6. Acceptance check.** Covered in its own section below, because it is the item most often skipped.

**7. Volume and span sheet.** Rough numbers are fine: failed handoffs per month, months of conversation and ticket history kept, and the dates of any help desk or bot platform migration that might have cut history short.

**8. Rights and sensitivity notes.** Which personal data appears (customer names, payment details, employee names), and whether any vendor contract limits use of conversation data. This is a list of questions for review, not legal advice.

## The acceptance check, written before anyone looks

Without an acceptance check, a recovery is just a later event on the same customer. With one, each case has a pass or fail that a reviewer can test. AI developers want exactly that distinction: Appen's service page, checked October 18, 2026, describes a trajectory analysis service in which experts classify where an agent went wrong and, for failed runs, supply the correct continuation. A company's own recoveries are only useful as continuations if someone can tell a good one from a bad one.

An illustrative check for the failed-handoff task, all conditions required:

- The follow-up ticket is solved with a recovery reason code.
- Any refund or credit settles in the payment system.
- The customer makes no repeat contact on the same issue within seven days.
- An approver is recorded for any amount above the refund limit.

Write it before counting cases. A check drafted after looking at the archive tends to describe whatever the archive happens to contain. Zendesk stores an audit for each ticket update, with events such as field changes and comments, so a check that refers to reason codes and approvers can usually be tested against ticket history later.

## An Illustrative cover sheet for a failed-handoff archive

Everything below is fictional: a 380-person home insurance broker with a chat assistant in front of its claims desk.

| Field | Entry |
|---|---|
| Full-time employees at peak | 380, contractors excluded |
| Systems | Messaging bot, help desk, billing platform |
| History kept | Help desk 6 years; bot conversations 26 months |
| Platform changes | Bot vendor changed 26 months ago; older conversations not retained |
| Failed handoffs | Roughly 300 to 450 per month, from the daily handoff report |
| Linking fields | Conversation ID copied onto tickets since the bot change |
| Acceptance check | Drafted by the head of support operations; not yet reviewed |
| Authorized contact | CEO, with the COO as internal sponsor |
| Open questions | Bot vendor terms on conversation data; payment details in transcripts |

Note how much the sheet says without a single record attached, and how openly it lists what is unknown. A 26-month conversation window is a limit, not a disqualifier, and stating it up front saves everyone time.

## What stays out of the packet, and what it cannot settle

The packet never contains transcripts, ticket exports, screenshots or refund records. Partners do not ask for them, and the company does not send them to a partner. Discovery stays at metadata level: what the records are, how many, how long, which system, who owns them.

Ownership usually splits three ways. Support operations owns conversation and ticket history, finance owns refunds and credits, and IT administers the tools. Authority to discuss a license sits with the CEO, who in an owner-run company is usually the owner. The company itself must qualify: 50 or more full-time employees at peak (contractors excluded), 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 the packet cannot settle: how many cases pass the acceptance check, how customer data will be de-identified, whether vendor terms permit use, and whether any buyer wants this task type. Those are decided in SourceX's review. To check whether a company is worth introducing at all, use the [agent recovery data qualification questions](/questions/does-agent-recovery-data-data-qualify); to describe weak spots honestly, see [agent recovery data evidence gaps](/questions/agent-recovery-data-evidence-gaps). Related reading: [AI-to-human handoff data](/data-types/ai-to-human-handoff-data) covers handoffs that worked, [agent trajectory data](/questions/agent-trajectory-data) covers the wider category, and [questions to ask before licensing company data](/resources/questions-to-ask-before-licensing-company-data) helps an owner prepare.

## 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 owner can fill in a cover sheet like the one above, with failed handoffs logged and linked to follow-up tickets, [register as a partner](/register), confirm the company against [who qualifies](/who-qualifies), and read the [terms](/terms) before you introduce the CEO.

## Sources

- https://www.appen.com/agentic-ai-trajectory-analysis
- https://support.zendesk.com/hc/en-us/articles/5514406080538-About-Sunshine-Conversations-in-Zendesk-Suite
- https://developer.zendesk.com/documentation/ticketing/reference-guides/ticket-audit-events-reference/
- https://refer.sourcex.si/terms
