Running an AI Reversal Post Mortem

• Source: Report framework library

Encodes a Revenue AI Report framework so the agent applies the published method instead of improvising one.

1. 1.Copy the SKILL.md text below, or download the raw file.
2. 2.Create a folder named exactly running-an-ai-reversal-post-mortem in your agent's skills directory.
3. 3.Save the file inside that folder as SKILL.md.
4. 4.Ask the agent one of the trigger requests below.
5. 5.Check the output against what you already know before it leaves your desk.

• We killed the AI SDR platform in July, write the post-mortem and code the reason
• Add our rolled-back support agent to the reversal ledger with the sunk spend
• The board wants to know why we pulled the agent out of production and what it cost

• Our AI pilot ends Friday, should we go, fix, or stop
• What is our organization-wide Proof Gap for the quarter

---
name: running-an-ai-reversal-post-mortem
description: Turn one shut-off AI deployment into a verified, coded, publishable Reversal Ledger row with disclosed sunk spend.
---

# Running An AI Reversal Post-Mortem

Turn one shut-off AI deployment into a verified, coded, publishable Reversal Ledger row with disclosed sunk spend.

## When to use this skill

- A production AI tool has been switched off and the organization needs the record before the knowledge leaves with the people.
- A Proof Gap review has flagged a specific tool for scrutiny and the tool has since been pulled (https://www.therevenueaireport.com/frameworks/proof-gap).
- Preparing a renewal or vendor consolidation review and needing the internal reversal history in one schema.
- Contributing an entry to the published Reversal Ledger dataset under source protection (https://www.therevenueaireport.com/data/reversal-ledger).
- A board asks why the agent was pulled and the answer currently exists only as hallway consensus.

## Inputs to collect

- Tool or vendor name and the contract, from procurement records.
- Company size band and the function area, from HR headcount and the org chart.
- Deployment date and the date the tool was actually switched off, from admin console logs or the vendor termination notice.
- The revenue seat that carried the work before and after the change, from the org chart.
- The stated reason in the deciding party's own words, from the decision memo, Slack thread, or a named operator interview.
- Sunk spend components from finance, covering seat licenses, platform fees, integration costs, and enablement spend specifically attributable to the tool.
- At least two independent verification inputs from the published set of a named operator source under NDA, a public statement, a finance record shared under confidentiality, or a vendor acknowledgment (https://www.therevenueaireport.com/methodology/reversal-ledger).
- Evidence on what the agent could read, what it could write, who could see the output, and what happened to the transcript, from the security review.

## Process

1. Apply the inclusion test before anything else. A reversal requires all three published conditions, so confirm the tool was in production use rather than pilot or proof-of-concept, that the organization made a deliberate decision to discontinue, and that the tool was actually shut off rather than merely downgraded (https://www.therevenueaireport.com/methodology/reversal-ledger). Artifact: a one-line inclusion verdict.
2. Apply the exclusion test. Drop the case if it is a vendor-initiated product sunset, a migration between two tools in the same category where the motion continues, a pilot non-conversion, or a temporary pause under ninety days. Artifact: the exclusion reason, or a note that none applies.
3. Assign exactly one primary reason code from the published six. Artifact: the code plus the sentence of evidence behind it.
4. Calculate sunk spend to the published rule. Total cost from contract signature to shutoff, covering licenses, platform fees, integration costs, and attributable enablement spend, excluding internal headcount hours. Publish only disclosed figures and estimate nothing (https://www.therevenueaireport.com/methodology/reversal-ledger). Artifact: the sunk-spend line with its components.
5. Verify against two independent inputs from the published set. Artifact: the two named inputs, held under source protection.
6. Write the row against the published dataset schema. Artifact: one complete row.
7. Give the vendor advance notice of a named entry, minimum fourteen days, and reserve space for the reply at equal prominence (https://www.therevenueaireport.com/methodology/reversal-ledger).
8. Write the operating lesson as one change to the next contract, naming the metric, the threshold, and the single owner who can pull the agent.

## Decision rules

The six published reason codes, verbatim (https://www.therevenueaireport.com/frameworks/reversal-ledger).

- DATA: data exposure, leakage, or compliance failure
- HALLUCINATION: brand or accuracy risk from generated output
- DIAGNOSTIC: inability to explain what the AI did or why
- ECONOMICS: cost per outcome exceeded manual baseline
- INTEGRATION: technical or workflow break with existing revenue stack
- ADOPTION: reps or CSMs stopped using it

- Code the trigger that ended the deployment, not the story that got told about it. Public embarrassment gets the coverage and cost gets the reversal, so ECONOMICS is more often correct than the narrative suggests (https://www.therevenueaireport.com/research/named-reversals).
- Suspect DATA before HALLUCINATION on any customer-communications agent. In the Sinch sample, data leakage pulled more agents out of production than hallucination did, which makes it a procurement and architecture problem before it is a model problem (https://www.therevenueaireport.com/research/rollback).
- Hold the entry rather than publishing it when only one verification input exists. Entries verified by only one source are held until a second input is obtained (https://www.therevenueaireport.com/methodology/reversal-ledger).
- Exclude internal headcount hours from sunk spend even though the Proof Gap includes AI enablement hours on the spend side. The two frameworks use different inclusion rules on purpose, and mixing them produces a number neither page supports.
- Do not treat a reversal as a verdict on the category. A tool's presence in the Ledger reflects specific implementations that failed, not the tool's category viability (https://www.therevenueaireport.com/frameworks/reversal-ledger).
- Keep held-the-line cases separate from reversals. IBM and Salesforce both held their positions and reported redeployment rather than retreat, and counting those as reversals would inflate the pattern (https://www.therevenueaireport.com/research/named-reversals).
- Treat one reversal per year as planned cost rather than failure, because in a 10-country sample of 2,527 senior decision makers, 74 percent had already rolled back or shut down a deployed AI customer communications agent over a governance failure while 98 percent were increasing AI communications investment anyway (https://www.therevenueaireport.com/research/rollback).
- Read a high internal reversal rate as detection capability, not incompetence. Enterprises describing their guardrails as mature rolled back agents more often, not less, and professional services rolled back at 85 percent against technology at 66 percent (https://www.therevenueaireport.com/research/rollback).
- Label the Sinch figures as vendor research when they appear in any output, because the publisher sells into the market it measured (https://www.therevenueaireport.com/research/rollback).
- Set no threshold for how many reversals is too many. No page publishes one. Set it internally as a share of production deployments per year, and label it a team-set threshold.

## Output requirements

Use the published dataset schema exactly, one row per reversal (https://www.therevenueaireport.com/data/reversal-ledger).

| Column | Definition |
|---|---|
| id | Permanent anchor id. Cite a single reversal as /reversal-ledger#id. |
| company | The organization, as named in the source record. |
| vendor | The vendor or system involved, when the reporting named one. |
| outcome | Rolled back or shut off; replaced humans then rehired; walked back the message; held the line; vendor repositioned. |
| function_area | Customer support, sales, marketing, operations, HR, or vendor. |
| seat | The revenue seat that carried the work before or after the change. |
| deployed | When the AI went live, when the reporting states it. |
| reversed | When the reversal, walk-back, or public position was recorded. |
| reason | Why it was reversed, in the source's own terms. |
| cost | Disclosed financial figure. Never estimated. |
| source_name | The publication or record the entry is compiled from. |
| source_url | Link to the source record, where public. |

Alongside the row, deliver the reason code, the sunk-spend components, the two verification inputs, and the one contract change. Mark any field the source record does not carry as not published rather than filling it in.

## Verification loop

1. Validate inclusion. Re-read the three inclusion conditions against the evidence. If any one fails, the case is not a reversal, so stop and file it as a pilot non-conversion or a pause instead.
2. Validate the reason code. Ask what would have kept the tool running. If the answer is not the thing the code names, the code is wrong, so reassign it and repeat this step.
3. Validate sunk spend. Reconcile the components to finance records line by line. If any component is an estimate, remove it and mark the field as not published, then re-total.
4. Validate verification. Confirm two independent inputs from the published set, each separately attributable. If only one holds, hold the entry and repeat once the second arrives.
5. Validate the schema. Confirm every column is present and that outcome and function_area use only the published allowed values. Fix and re-check any field that does not.
6. Only proceed when inclusion passes on the second read, one reason code survives the counterfactual question, sunk spend contains no estimates, two independent inputs are attached, and every schema field carries a published value or an explicit not-published mark. Publish or circulate the entry only after that point. If any check fails, keep the entry internal and name the missing input.

## Quality checks

- Exactly one primary reason code, drawn verbatim from the published six.
- Sunk spend excludes internal headcount hours.
- No estimated figure appears in the cost field.
- Two independent verification inputs are named and protected.
- Vendor notice of at least fourteen days is logged before a named entry circulates.
- The output does not generalize from one implementation to the whole category.

## Limitations

- Reason codes compress a multi-cause failure into one label. Record the secondary cause in the reason field in the source's own terms so the nuance survives.
- The published case set is a documented case list, not a representative sample. The base rate lives in the rollback theme, measured on 2,527 decision makers (https://www.therevenueaireport.com/research/named-reversals).
- Sunk spend understates true cost because it excludes internal headcount hours by rule.
- Post-mortems run after the fact rely on memory and on people who may have left the role.
- The base-rate source is vendor-published. It is used because the sample is large and the field dates are stated, and the incentive is disclosed (https://www.therevenueaireport.com/research/rollback).

## Example input

Illustrative case, synthetic figures. An AI customer-communications agent went live in March 2025 across support and inbound sales at a 1,200-person company. It was switched off in July 2026 after a security review found it could read a shared drive containing customer contracts. Finance discloses $310k in seat and platform fees plus $85k in integration cost. Two inputs exist, a finance record shared under confidentiality and the vendor's written acknowledgment of the termination.

## Example output

Inclusion: passes. Production use, deliberate decision, actually shut off.

Reason code: DATA, for data exposure, leakage, or compliance failure. Not HALLUCINATION, because output quality was never the trigger.

Sunk spend: $395k, comprising $310k licenses and platform fees plus $85k integration. Internal headcount hours excluded by rule. Nothing estimated.

Verification: two independent inputs, a confidential finance record and a vendor acknowledgment. Publishable.

Row: outcome is rolled back or shut off, function_area is customer support, seat is Customer Success, deployed 2025-03, reversed 2026-07, cost $395k.

Contract change for the next agent: diligence the data boundary before the model, because leakage ends more deployments than hallucination does (https://www.therevenueaireport.com/research/rollback). Name the metric, the threshold, and one owner who can pull the agent without a committee.

Comparable published cases for context: Klarna, whose assistant was marketed as doing the work of 700 agents and whose customer service and operations cost had risen to $50M from $42M a year earlier by Q3 2025, settling into a hybrid model with AI on about two-thirds of inquiries; Air Canada, where the tribunal rejected the argument that the chatbot was a separate legal entity and ordered C$812.02; and Zillow Offers, announced November 2, 2021 against a Q3 2021 Homes segment loss before tax of $421.6M (https://www.therevenueaireport.com/research/named-reversals).

## Rules of conduct

- Write for a Director, VP, or operator. Short sentences. Explain uncommon terms.
- Separate facts from assumptions. Never hide uncertainty.
- Do not invent numbers, benchmarks, quotes, or customer names.
- Do not send messages, change CRM records, or publish anything unless the user explicitly asks.
- Flag when a decision needs human review.

## Evidence

- https://www.therevenueaireport.com/frameworks/reversal-ledger
- https://www.therevenueaireport.com/methodology/reversal-ledger
- https://www.therevenueaireport.com/data/reversal-ledger
- https://www.therevenueaireport.com/research/named-reversals
- https://www.therevenueaireport.com/research/rollback
- https://www.therevenueaireport.com/frameworks/proof-gap
- https://www.therevenueaireport.com/skills/review-ai-pilot

## Cite this framework

Kvarfordt, Jonathan. "The Reversal Ledger." The Revenue AI Report. https://www.therevenueaireport.com/frameworks/reversal-ledger

What does the Running an AI Reversal Post Mortem skill do?

Turn one shut-off AI deployment into a verified, coded, publishable Reversal Ledger row with disclosed sunk spend.

Where does the Running an AI Reversal Post Mortem skill come from?

Report framework library. It was written by The Revenue AI Report against a 12 criterion quality rubric and graded in an independent scoring pass.

Why was the Running an AI Reversal Post Mortem skill chosen for this library?

Encodes a Revenue AI Report framework so the agent applies the published method instead of improvising one.

When should the Running an AI Reversal Post Mortem skill not be used?

Do not use it for: Our AI pilot ends Friday, should we go, fix, or stop Or: What is our organization-wide Proof Gap for the quarter

How do I install the Running an AI Reversal Post Mortem SKILL.md file?

Download the file, create a folder named exactly running-an-ai-reversal-post-mortem inside your agent's skills directory, and save the file inside it as SKILL.md. The agent loads it when a request matches the description.