What counts as a reversal?

A reversal is a deployment the company had to undo: rolling the AI back, rehiring humans, or walking back the message. Each case must be substantiated at a named source. Cases that could not be substantiated were dropped rather than softened, and fields the reporting does not carry are marked as not published rather than filled in. That makes this a collected set of reported cases, not a representative sample. The base rate of reversal is measured in Theme 03, not here.

How is this different from vendor AI ROI dashboards?

This ledger records what companies took back out, not what a vendor reports going in. Every row is a documented case with dates, sources, and where disclosed, the cost, and the record is independent of the vendors involved. Most rows have no published deployment date and no published dollar figure, because companies disclose the retreat far more readily than the spend that preceded it. The page is updated monthly and corrections are published at /corrections.

The Reversal Ledger

Every AI deployment we can document that a company had to undo.

Name: The Reversal Ledger
Creator: The Revenue AI Report
License: https://creativecommons.org/licenses/by/4.0/

One row per case. Filterable by outcome and function, with a permanent anchor on every row so a single reversal can be cited on its own. Fields the reporting does not carry are left marked as not published rather than filled in.

15: documented cases in the ledger, each verified at a named source
2: cases where humans were replaced and then rehired
3: cases carrying a disclosed financial figure

Every row in this ledger is a company that put AI into revenue-adjacent or customer-facing work and had to take it back out.

The count is going up, not down. Enterprises with mature AI governance report higher rollback rates than those without, because governance is what surfaces failures early enough to reverse. Fifteen public cases are recorded here. The private cases outnumber them.

If your team is planning a deployment, use this list to name the failure modes that will be quoted back to you if yours ends up on it. Filter by function, seat, or outcome. Every row carries dates, sources, and where disclosed, the cost.

This page is the live, browsable record. The definition of the framework and how to run it on your own team is at /frameworks/reversal-ledger. The machine-readable dataset and CSV download are at /data/reversal-ledger. The verification standard is at /methodology/reversal-ledger.

Updated monthly · Last update: August 2026 · Corrections at /corrections

Sort by

Showing 15 of 15 rows.

Artisan
#artisan
Walked back the message
Function
Vendor
Seat affected
AI SDR
Deployed
Not published by the source.
Reversal recorded
Aug 2026
What happened
The stop hiring humans campaign was retired. Artisan publicly announced it was hiring its first human BDR, and customers ended up with more humans, not fewer.
Disclosed cost
Not published by the source.
Source: Company statements and coverage, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Taco Bell, Yum Brands
#taco-bell-yum-brands
Rolled back or shut off
Function
Operations
Seat affected
Drive-thru ordering
Deployed
Not published by the source.
Reversal recorded
Aug 29, 2025
What happened
A public rethink of voice ordering after an order for 18,000 water cups circulated.
Disclosed cost
Not published by the source.
Source: Contemporaneous coverage, verified per case in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Commonwealth Bank of Australia
#commonwealth-bank-of-australia
Replaced humans, then rehired
Function
Customer support
Seat affected
Contact centre
Deployed
Not published by the source.
Reversal recorded
Aug 21, 2025
What happened
Call volumes rose after the voice bot went in. 45 redundancies were reversed and the bank apologized.
Disclosed cost
Not published by the source.
Source: Company statement and coverage, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Klarna
#klarna
Replaced humans, then rehired
Function
Customer support
Seat affected
Support agents
Deployed
Assistant marketed as doing the work of 700 agents
Reversal recorded
May 2025
What happened
Cost became a too predominant factor and quality fell. Agents were rehired and the company settled into a hybrid model with AI on about two-thirds of inquiries.
Disclosed cost
Customer service and operations cost rose to $50M in Q3 2025 from $42M a year earlier
Source: BBC coverage and Klarna Q3 2025 earnings
Verified via BBC coverage and Klarna Q3 2025 earnings
Share on LinkedInShareable card. The mark and citation link travel with the image.
IBM with In-house agents
#ibm-hr
Held the line
Function
HR
Seat affected
HR operations
Deployed
Not published by the source.
Reversal recorded
May 2025
What happened
About 200 HR roles moved to agents. Total employment went up and the savings funded sellers.
Disclosed cost
Not published by the source.
Source: Company statements, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Duolingo
#duolingo
Walked back the message
Function
Marketing
Seat affected
Company-wide positioning
Deployed
Not published by the source.
Reversal recorded
Apr to May 2025
What happened
An AI-first memo was followed by the CEO saying he does not see AI as replacing what employees do.
Disclosed cost
Not published by the source.
Source: Company statements, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
11x
#11x
Vendor repositioned
Function
Vendor
Seat affected
AI SDR
Deployed
Not published by the source.
Reversal recorded
Mar 24, 2025
What happened
Reported churn of 70 to 80%, with ARR that counted three-month break clauses as full-year contracts.
Disclosed cost
Not published by the source.
Source: TechCrunch investigation
Verified via TechCrunch investigation
Share on LinkedInShareable card. The mark and citation link travel with the image.
Coca-Cola
#coca-cola
Held the line
Function
Marketing
Seat affected
Brand advertising
Deployed
Not published by the source.
Reversal recorded
2024 and 2025
What happened
Backlash on generative holiday advertising did not change the approach. The campaign was repeated the following year.
Disclosed cost
Not published by the source.
Source: Company statements and coverage, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Salesforce with In-house agents
#salesforce-support
Held the line
Function
Customer support
Seat affected
Support
Deployed
Not published by the source.
Reversal recorded
2025
What happened
Support headcount moved from 9,000 to about 5,000, with agents on 50% of Service Cloud interactions and no reversal disclosed.
Disclosed cost
Not published by the source.
Source: Company statements, verified in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
McDonald's with IBM
#mcdonalds-ibm
Rolled back or shut off
Function
Operations
Seat affected
Drive-thru ordering
Deployed
Not published by the source.
Reversal recorded
By Jul 26, 2024
What happened
Automated order taking was ended across the test estate after order-accuracy failures at more than 100 restaurants.
Disclosed cost
Not published by the source.
Source: Contemporaneous coverage, verified per case in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Air Canada
#air-canada
Rolled back or shut off
Function
Customer support
Seat affected
Customer service
Deployed
Not published by the source.
Reversal recorded
Feb 16, 2024
What happened
A tribunal rejected the argument that the chatbot was a separate legal entity and held the airline to the bereavement fare its bot described.
Disclosed cost
C$812.02 awarded
Source: Civil Resolution Tribunal decision, BBC coverage
Verified via Civil Resolution Tribunal decision, BBC coverage
Share on LinkedInShareable card. The mark and citation link travel with the image.
DPD
#dpd
Rolled back or shut off
Function
Customer support
Seat affected
Customer service
Deployed
Not published by the source.
Reversal recorded
Jan 2024
What happened
The support chat swore at a customer. The post was viewed 800,000 times in 24 hours and the AI element was disabled.
Disclosed cost
Not published by the source.
Source: Contemporaneous coverage, verified per case in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Chevrolet of Watsonville
#chevrolet-of-watsonville
Rolled back or shut off
Function
Sales
Seat affected
Dealership sales chat
Deployed
Not published by the source.
Reversal recorded
Dec 2023
What happened
The site chatbot was talked into agreeing to a one dollar Tahoe. The feature was removed.
Disclosed cost
Not published by the source.
Source: Contemporaneous coverage, verified per case in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.
Zillow Offers with In-house pricing models
#zillow-offers
Rolled back or shut off
Function
Operations
Seat affected
Pricing and buying operations
Deployed
Not published by the source.
Reversal recorded
Nov 2, 2021
What happened
Algorithmic home buying was wound down after the Homes segment posted a loss before tax, with a workforce reduction of about 25%.
Disclosed cost
$421.6M Q3 2021 Homes segment loss before tax
Source: Zillow Group Q3 2021 results
Verified via Zillow Group Q3 2021 results
Share on LinkedInShareable card. The mark and citation link travel with the image.
Amazon with In-house recruiting engine
#amazon-recruiting-engine
Rolled back or shut off
Function
HR
Seat affected
Recruiting
Deployed
Built from 2014
Reversal recorded
By the start of 2017
What happened
500 models across roughly 50,000 resume terms penalized resumes containing the word women's. The engine was disbanded.
Disclosed cost
Not published by the source.
Source: Reuters reporting, cited in Theme 04. Case record: Theme 04
Verified via case record
Share on LinkedInShareable card. The mark and citation link travel with the image.

What we looked for and did not find

Most rows have no published deployment date and no published dollar figure. Companies disclose the retreat far more readily than the spend that preceded it. Cases we could not substantiate at a named source were dropped rather than softened, so this is a collected set of reported cases, not a representative sample. The base rate of reversal is measured in Theme 03, not here.

Cite this ledger

Jonathan Kvarfordt, The Reversal Ledger, The Revenue AI Report. https://www.therevenueaireport.com/reversal-ledger. Cite a single case with its row anchor, for example https://www.therevenueaireport.com/reversal-ledger#klarna.

The full case write-ups, charts, and per-source confidence tags live in Theme 04: Named reversals, with dates and dollar amounts.

Questions about this ledger

What is the Reversal Ledger?: The Reversal Ledger is a permanent, filterable record of documented cases where a company put AI into revenue-adjacent or customer-facing work and had to take it back out. One row per case, with dates, disclosed cost, and source, and a permanent anchor on every row so a single reversal can be cited on its own. This page is the live, browsable record. The framework definition is at /frameworks/reversal-ledger, the machine-readable dataset and CSV are at /data/reversal-ledger, and the verification standard is at /methodology/reversal-ledger.
What counts as a reversal?: A reversal is a deployment the company had to undo: rolling the AI back, rehiring humans, or walking back the message. Each case must be substantiated at a named source. Cases that could not be substantiated were dropped rather than softened, and fields the reporting does not carry are marked as not published rather than filled in. That makes this a collected set of reported cases, not a representative sample. The base rate of reversal is measured in Theme 03, not here.
How should I cite the Reversal Ledger?: Cite it as: Jonathan Kvarfordt, The Reversal Ledger, The Revenue AI Report. https://www.therevenueaireport.com/reversal-ledger. Cite a single case with its row anchor, for example https://www.therevenueaireport.com/reversal-ledger#klarna. Full case write-ups, charts, and per-source confidence tags live in Theme 04: Named reversals, with dates and dollar amounts.
How is this different from vendor AI ROI dashboards?: This ledger records what companies took back out, not what a vendor reports going in. Every row is a documented case with dates, sources, and where disclosed, the cost, and the record is independent of the vendors involved. Most rows have no published deployment date and no published dollar figure, because companies disclose the retreat far more readily than the spend that preceded it. The page is updated monthly and corrections are published at /corrections.

New reversals land in the weekly issue. Get the Report.

How to cite this dataset

Written by Jonathan Kvarfordt, Founder and Principal Analyst, The Revenue AI Report. Published under CC BY 4.0.

APA

Kvarfordt, J. (2026). The Reversal Ledger. The Revenue AI Report. Retrieved from https://www.therevenueaireport.com/reversal-ledger

MLA

Kvarfordt, Jonathan. "The Reversal Ledger." The Revenue AI Report, 31 Aug. 2026, www.therevenueaireport.com/reversal-ledger.

BibTeX

@misc{kvarfordt2026reversalledger,
  author = {Kvarfordt, Jonathan},
  title = {The Reversal Ledger},
  year = {2026},
  publisher = {The Revenue AI Report},
  url = {https://www.therevenueaireport.com/reversal-ledger}
}

Artisan

Taco Bell, Yum Brands

Commonwealth Bank of Australia

Klarna

IBM with In-house agents

Duolingo

11x

Coca-Cola

Salesforce with In-house agents

McDonald's with IBM

Air Canada

DPD

Chevrolet of Watsonville

Zillow Offers with In-house pricing models

Amazon with In-house recruiting engine

How to cite this dataset

New reversals land in the weekly issue.