# Agent recovery data: evidence gap analysis | SourceX

Canonical: https://refer.sourcex.si/questions/agent-recovery-data-evidence-gaps
Last updated: 2026-11-04

Most gaps in agent recovery data are broken links rather than missing records. The failure sits in a bot or phone log, the fix in a help desk ticket, the proof in a billing or payment system, and nothing joins them. The six gaps that matter are: no stored failure event, no customer context, no link from follow-up to failure, no record of who approved the fix, no version of the routing setup, and no confirmed outcome. Each narrows what a review can use; none rules a company out.

This page is part of the guide to [agent recovery data](/guides/concept-agent-recovery-data), which follows one failed chat transfer to a verified fix. Its job is a gap analysis: what a review would expect from a history of recoveries, what support archives typically keep, where the chain breaks, and how to describe it in an introduction.

## What a recovery review expects to find

A recovery is assessable when a reviewer can rebuild three things for one case: where the task broke, what someone did about it, and how they knew the fix held. Six kinds of evidence carry that.

| Expected evidence | What it lets a reviewer rebuild |
|---|---|
| Failure event with a time | The exact point the task broke, and why |
| Customer context | What the customer asked, so the task is more than its result |
| Follow-up link | That the later ticket is the repair of this failure, not another case |
| Approver of the fix | That the repair was within policy, for example who approved a refund |
| Routing or configuration version | Which setup failed, if rules changed around the same time |
| Confirmed outcome | That the fix held: a settled refund, no repeat contact, a closing reason |

Agent developers want this chain because a failed run without its repair teaches little. Appen's service page, checked October 18, 2026, describes a trajectory analysis service in which experts classify where agents go wrong and write the correct continuation for failed runs. A company's own linked recoveries are the real-world version of that continuation; unlinked ones are two unrelated events.

## Broken links: the gap table

The pattern below reflects how support stacks are commonly set up, not a survey of companies.

| Gap | What archives usually keep | What is usually missing | Effect on assessment |
|---|---|---|---|
| Failure event | Bot outcome or call disposition | The routing event and its target queue | The break point has to be guessed |
| Customer context | A summary or intent label | What the customer actually said | The task cannot be reconstructed, only its result |
| Follow-up link | A new ticket on the next contact | The original conversation or case ID on that ticket | The repair cannot be attributed to the failure |
| Approver | The refund or credit itself | Who approved it and in what role | Nobody can tell whether the fix was within policy |
| Routing version | The current routing rules | Earlier rules, edited in place | It is unclear which configuration failed |
| Confirmed outcome | Ticket status "solved" | A repeat-contact check or settled payment | The fix might not have worked |

The failure event is easier to capture than many companies assume. Zendesk's messaging documentation, for example, describes conversation control being passed to an integration or released back to a default state; events like these mark where a handoff stopped. The hardest gap is usually the follow-up link, because it depends on an agent copying an ID across systems at the moment of the repair.

## Ten recoveries at a home-services firm (Illustrative)

**Illustrative and fictional.** A 380-person home-services company books plumbing and heating visits through a website chat assistant that hands urgent cases to a dispatch team. When a handoff fails, dispatch usually calls the customer back the next morning. The introducer asks the head of customer operations a simple question for each kind of evidence: out of every ten recoveries, roughly how many keep it?

| Evidence | Owner's estimate, out of 10 | Why it falls short |
|---|---|---|
| Failure event | 8 | Phone-line failures after hours are not logged |
| Customer context | 6 | The bot keeps an intent label; full text only since 2023 |
| Follow-up link | 4 | Dispatch opens a new job without the chat ID unless reminded |
| Approver | 5 | Goodwill credits over a set amount need a manager, logged only in billing |
| Routing version | 3 | Rota-based routing rules are edited weekly and not versioned |
| Confirmed outcome | 5 | Job completion is recorded; repeat contact is not checked |

Read across, the linked core is the four in ten recoveries that carry the chat ID, and within that group the context and outcome are reasonably complete. The routing gap matters less here than it might elsewhere, because dispatch failures are mostly caused by after-hours staffing, not rule changes. That is a useful thing to say plainly. The graphic on this page draws the same six rows as blocks of ten.

## Migration and retention: gaps that open over time

Some gaps are not about how a case was handled but about what happened to the records afterwards.

- **Help desk replacement.** Ticket history can carry a lot. Zendesk's developer reference explains that every update to a ticket stores an audit with events such as field changes and comments, which is what lets the order of a recovery be rebuilt. When a company changes help desk, ask whether that audit history migrated or only the final ticket state did.
- **Different retention for bot and ticket records.** Conversation logs and tickets often live in different tools with different retention settings. If the bot keeps a shorter history than the help desk, older recoveries have a repair but no recorded failure.
- **Restores after an incident.** Records rebuilt from backups may lose links or timestamps; [ransomware-restored data integrity](/questions/ransomware-restored-data-integrity) covers what to ask.

Exceptions in other workflows behave the same way: a [freight audit exception](/data-types/freight-audit-exceptions) is only useful when the original invoice, the dispute and the correction stay linked.

## Saying it plainly in the introduction

| Instead of | Say |
|---|---|
| "Every failure is tracked." | "Chat handoff failures are logged; after-hours phone failures are not." |
| "We link everything." | "About four in ten recovery jobs carry the original chat ID." |
| "Fixes are approved." | "Credits above the limit need a manager; approvals sit in billing, not the job." |
| "We know it worked." | "Job completion is recorded; repeat contact is not checked." |

Give years and shares rather than adjectives, and note who gave each answer. A company that names its gaps reads as more credible than one that claims a complete archive.

## Who owns the records, and what stays unknown

Support operations usually owns conversation and ticket history, finance owns refunds and credits, and IT administers the tools. Authority to discuss licensing normally sits with the COO, CFO or CEO. Bot and help desk vendor terms may affect rights, so a review checks them.

The company test does not change with gaps: 50+ 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.

Unknown until SourceX reviews it: the real share of linked recoveries, how customer data will be de-identified, and whether any buyer wants this task type. To decide whether a client is worth introducing at all, use the [qualification scorecard](/questions/does-agent-recovery-data-data-qualify). What the company would later assemble is listed in the [review packet](/resources/agent-recovery-data-review-packet). For the wider ticket archive, see [using historical tickets for a support agent](/guides/use-historical-tickets-to-train-ai-support-agent), and for step-by-step records generally, [agent trajectory data](/questions/agent-trajectory-data).

## 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 you can tell a client's owner which links in their recovery chain hold and which break, [register as a partner](/register), check [who qualifies](/who-qualifies), and read the [program terms](/terms) before you make the introduction.

## Sources

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