# Failed AI pilot: a partner playbook | SourceX

Canonical: https://refer.sourcex.si/guides/failed-ai-pilot-data-referral-opportunity
Last updated: 2026-11-04

A failed enterprise AI pilot shows, in writing, where generic software lacked the company's context, and that points to the records of people handling those cases correctly: escalation queues, staff corrections and years of workflow history. The window is short. Once the vendor environment is shut down and the project team disbands, that knowledge fades. This playbook covers the trigger, the signals, the records a wind-down puts at risk, and the conversation that leads to a metadata-level introduction, without asking for a single file.

This playbook is part of [what a failed data monetization or AI pilot reveals](/questions/data-monetization-failed), which explains how pilot error types map to records. This page is the field version: when to act, what to listen for and what to say.

## The trigger: a pilot wind-down

The trigger is not the failure itself. It is the moment the company decides to stop: the sponsor writes up results, the vendor contract is not renewed, and the test environment is scheduled for deletion. NIST's AI Risk Management Framework asks that performance be measured under conditions similar to deployment, which is what a pilot does. Its write-up is therefore the company's own record of where generic software lacked context. Google's People + AI guidebook names this kind of failure: a context error, where the system's assumptions about the situation are wrong.

For a partner, that write-up reframes the question. Instead of "do you have data?", the question becomes "which of your records show people handling the cases the pilot missed?"

## Signals a client is in it

Listen for these in ordinary client conversations. Two or more together usually mean a wind-down is under way.

| Signal | Where you hear it | What it suggests |
|---|---|---|
| "We're not renewing the pilot" | Budget or vendor review | A contract end date, often with data deletion terms |
| A post-mortem or lessons-learned deck is being written | Steering committee, sponsor update | Error categories are being written down now |
| The project team is being reassigned | Resourcing conversations | The people who know the edge cases are about to scatter |
| "It couldn't handle our exceptions" | Any operations meeting | The company's exception handling is the valuable part |
| The test environment is being shut down | IT change calendar | Evaluation sets built for the pilot may be deleted |

ERP and operations consultants often hear these first; the [NetSuite consultant guide](/guides/how-erp-consultants-spot-referral-opportunities-during-netsuite-projects) shows the same listening during a system project, and the [network opportunity finder](/tools/network-opportunity-finder) helps choose which clients to approach.

## The records a wind-down puts at risk

A pilot rarely creates valuable records by itself. It touches them, and the wind-down can delete the pieces that connect them. Five record types, described at metadata level:

1. **The evaluation set.** Cases the team picked to test the pilot, often with the correct answer written by staff. Frequently kept in the vendor or test environment.
2. **Staff corrections.** Where staff overrode or edited the pilot's output, with the right answer. May live only in the pilot's interface.
3. **The escalation queue.** Cases the pilot handed to a person, and how the person resolved them.
4. **The source workflow history.** Tickets, work orders, approvals or claims in the company's own systems, going back years before the pilot. This is the durable asset; the pilot only showed where it matters.
5. **The vendor's copies.** Extracts and outputs held under the pilot contract. Not the company's source records, and counsel's question, not the partner's.

Items 1 to 3 are the ones a wind-down can erase. Item 4 usually survives. A partner's useful contribution is to mention, before the deletion date, that the company may want to keep its own evaluation cases and corrections, whatever it decides about licensing.

## The conversation: four moves

**Illustrative and fictional.** A 300-person regional insurance claims administrator ended a pilot that drafted claim decisions. The sponsor, the COO, is writing the post-mortem.

1. **Acknowledge what it showed.** "The pilot kept missing the exceptions your adjusters handle every day. That tells you something about your own records."
2. **Ask for categories, not cases.** "In a sentence each, what kinds of errors did it make?" The COO lists coverage exclusions misapplied, late-reported claims mishandled, and missing context from prior claims on the same policy.
3. **Ask the metadata questions.** For each category: which system holds the correctly handled cases, how many years, roughly how many a year, and who owns them. The COO answers: the claims system, eleven years, several thousand exceptions a year, owned by the company, with some client-specific terms.
4. **Name the introduction.** "If it were worth exploring, who would need to say yes?" The COO names the CEO, with general counsel reading client agreements.

Partner notes after the meeting: three error categories, one system, eleven years, rough volume, owner, signer, a client-contract caveat, and the test environment deletion date. No files, no screenshots, no copy of the post-mortem.

If the sponsor wants to take the idea internally, the [internal memo template](/resources/internal-memo-template-proposal) gives them a structure; the [AI readiness checklist](/resources/ai-readiness-assessment-checklist) helps if they plan another pilot.

## What the playbook cannot tell you

- **Rights.** Some correctly handled cases quote client contracts or contain personal data. SourceX's rights review decides what can be used.
- **Survival.** Older years may sit in a retired system or be gone.
- **Demand and terms.** Whether any buyer wants these records, and at what price, stays unknown until SourceX reviews them.
- **Eligibility.** The company needs at least 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.

The broader reasons pilots stall are in [why AI pilots fail in portfolio companies](/guides/why-ai-pilots-fail-portfolio-companies). If the company licenses records, it is paid under its own agreement with SourceX; the partner reward is separate, funded by SourceX and never deducted from the company's payout. Sign the Channel Partner Agreement before submitting the introduction.

## 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's sponsor can name the error categories and the system that holds the handled cases, [register as a partner](/register), check the company on [who qualifies](/who-qualifies), and read the [program terms](/terms) before the test environment is gone.

## Sources

- https://pair.withgoogle.com/chapter/errors-failing/
- https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
- https://refer.sourcex.si/terms
