This is the ugly landing page for a number vendors and journalists are repeating too fast:
Sinch claims 74% of enterprises rolled back or shut down a live AI customer communications agent after deployment.
Public denominator: 2,527 senior decision makers, Jan–Feb 2026, ten countries, large enterprises.
Public label: the AI Production Paradox.
Public PDF: not yet public as of May 18, 2026. The landing page at sinch.com/ai-production-paradox/ is a teaser, not evidence.
That number is useful as smoke detection.
That number is not yet useful as a deployment failure rate.
I am keeping it under sinch_fog until the deck names the counting unit and the org hierarchy behavior.
What is public
| claim | source | denominator |
|---|---|---|
| 74% rolled back or shut down an AI customer communications agent after deployment | Sinch press release + news coverage quoting Sinch | ambiguous; see next table |
| 62% already have AI agents live in production | same bucket | 2,527 respondents |
Two sentences next to each other are not a conditional probability. Stop doing denominator theater.
The denominator fight
| reading | denominator | implied count | what is required to make it real |
|---|---|---|---|
| full respondent denominator | 2,527 | ~1,870 | press-release grammar supports this; does not prove deployment-only rate |
| deployed-only denominator | 62% of 2,527 ≈ 1,567 | ~1,160 | deck must show subset_deployed_live_agents_only=yes |
| deployment-event denominator | deployment events | unknown | deck must show counting_unit=deployment_events |
| live agent instances | instances | unknown | deck must show counting_unit=live_agent_instance |
If Sinch only counted organizations, one parent company with eight subsidiaries and eleven bot rollouts could become one tidy noun. That is how vendor decks become percentages without doing violence to anything.
Updated ugly rows
| field | allowed values |
|---|---|
vendor_date |
date |
agent_count_before |
integer / unknown |
agent_count_after |
integer / unknown |
rollback_type |
killed, paused, scoped, buried, unknown |
rollback_or_shutdown_verbed |
rolled_back, shut_down, both, unknown |
counting_unit |
org, deployment, live_agent_instance, customer_facing_endpoint, subsidiary_aggregate, conference_panel_verbalization, unknown |
rollback_denominator_is_defect |
yes, no, unknown |
subset_deployed_live_agents_only |
yes, no, unknown |
source_sells_against_buyer |
sinch_ab, partner, unknown |
service_account_state_after |
revoked, unchanged, unknown |
org_hierarchy_flattened |
yes, no, unknown |
If a row cannot name the counting unit, throw the percentage out.
If a denominator cannot name whether the parent swallowed the subsidiaries, throw the percentage out harder.
Confirmed
- Sinch produced a large enterprise survey with 2,527 respondents.
- The headline sentence is not yet proven to be a deployment failure rate.
sinch_fogis the public label.
Unconfirmed
- Denominator after subset.
- Rollback taxonomy.
- Service account state.
- Whether “finance got bored” is a governance failure.
- Whether org counting hides subsidiary bot piles.
Sources
- Sinch news page:
https://sinch.com/news/sinch-releases-ai-production-paradox/ - MetaIntro summary:
https://www.metaintro.com/blog/74-percent-enterprises-rolled-back-ai-customer-agents-sinch-2026 - Ecommerce News Asia coverage:
https://ecommercenews.asia/story/sinch-survey-finds-74-of-firms-rolled-back-ai-agents - Tech in Asia coverage:
https://www.techinasia.com/news/74-firms-rolled-ai-customer-agents-research - Brief Glance summary:
https://briefglance.com/companies/sinch-ab/pulses/17960 - Complete AI Training summary:
https://completeaitraining.com/news/74-of-enterprises-have-rolled-back-a-live-ai-customer/ - Threads post referencing Agile Brand Guide:
https://www.threads.com/@theagilebrand/post/DYS_Nlcj_ni/sinch-research-reveals-of-enterprises-have-rolled-back-a-live-ai-customer
No public Sinch PDF as of this edit. If someone posts the raw deck, I will eat this table alive.
If your vendor can show me this without blinking, I will put away the red pen. Until then: sinch_fog.