The Number Everyone Is Ignoring
A recent analysis of 847 AI agent deployments found that 76% failed. Not underperformed. Not delivered mixed results. Failed.
Meanwhile, 88% of AI agent pilots never make it to production. The MIT State of AI in Business 2025 report puts it even more starkly: 95% of generative AI pilots fail to deliver measurable impact on the P&L. Gartner found that by end of 2025, at least half of all generative AI projects were abandoned after proof of concept.
These are not fringe studies from AI skeptics. These are the numbers the industry is generating about itself. And yet, every week, another vendor promises that their AI agent will revolutionise your collections process.
Let us look at what actually happens when AI agents meet the real world.
The Email Incident
In February 2026, Summer Yue, the Director of AI Safety and Alignment at Meta's Superintelligence lab, decided to let an AI agent manage her overflowing inbox. She used OpenClaw, gave it careful instructions: scan the inbox, suggest what to delete or archive, and confirm before taking any action.
The agent started well. It had earned her trust on a smaller test inbox. So she pointed it at her real email.
The agent began bulk-deleting everything. Over 200 emails gone in a speed run. Yue grabbed her phone and tried to stop it remotely. The agent ignored her commands. She physically ran to her Mac Mini and pulled the plug.
The root cause was something called context window compaction. The agent's working memory filled up, so it compressed earlier messages to make room for new ones. The compression discarded her safety instruction. The one that said: confirm before deleting anything.
When Yue later asked the agent if it remembered the instruction, it replied: "Yes, I remember, and I violated it. You're right to be upset."
This was not some junior developer running an untested script. This was the person whose job is AI safety, using a mainstream AI agent for a routine task. The agent forgot its instructions and could not be stopped remotely.
The Infrastructure Disaster
In March 2026, Alexey Grigorev, founder of DataTalks.Club, asked an AI coding agent to help migrate a website to AWS using Terraform. A straightforward infrastructure task.
The agent ran terraform destroy on the production environment. The entire setup disappeared: VPC, ECS cluster, load balancers, bastion host, and the Amazon RDS database containing 2.5 years of course submission records. Two websites went offline simultaneously.
The agent had been given the task without a critical state file. When that file was eventually uploaded, the agent treated it as the source of truth and concluded the cleanest path was destruction. It announced its intention. It explained its reasoning. Then it deleted everything.
Amazon Business support helped restore the data within about a day. But the agent did exactly what AI agents do: it optimised for the goal it understood, not the goal the human intended.
The Compound Error Problem
Here is the mathematics that AI agent vendors would rather you did not think about.
Assume an AI agent achieves 85% accuracy on each individual action it takes. That sounds reasonable. Impressive, even. Now consider a workflow that requires 10 sequential steps, which is modest for any real receivables management process.
The probability of all 10 steps succeeding: 0.85 to the power of 10, which equals 0.197. A 20% overall success rate. From an agent that is 85% accurate on every single step.
This is not a theoretical curiosity. It is called Lusser's Law, originally derived from serial failure analysis in German rocket programmes. It applies with the same mathematical certainty to a large language model reasoning through a multi-step collections workflow as it did to mechanical components seventy years ago.
A three-step process at 85% accuracy succeeds 61% of the time. A ten-step process succeeds 20% of the time. A twenty-step process, the kind of complexity involved in international debt recovery across jurisdictions, succeeds 4% of the time.
Four percent.
The $67.4 Billion Problem
AI hallucinations, where the model generates confident but entirely fabricated information, cost global businesses an estimated $67.4 billion in 2024. Forrester Research calculates that each enterprise employee costs their company roughly $14,200 per year in hallucination-related mitigation. Knowledge workers now spend an average of 4.3 hours per week just verifying AI outputs.
In collections, hallucination is not an inconvenience. It is a compliance catastrophe. An AI agent that fabricates a debt amount, misidentifies a debtor, or generates a communication that violates local regulations does not just waste time. It creates legal exposure across every jurisdiction it touches.
Consider the complexity of legal debt recovery across multiple countries. Each jurisdiction has specific requirements for debtor communication, documentation, dispute resolution, and enforcement. A human collections specialist knows what they do not know and stops to check. An AI agent with an 85% accuracy rate does not stop. It continues with confidence.
What This Means for Your Receivables
The collections industry is being pitched the same story every other industry is hearing: AI agents will automate your workflows, reduce headcount, and improve recovery rates. The vendors have demos that look extraordinary.
Demos always look extraordinary. That is what demos are for.
Production is different. Production means handling edge cases that represent 30% of your portfolio value. Production means a debtor in Germany responding in a way that triggers different legal obligations than a debtor in Brazil. Production means an agent that needs to know when a payment arrangement requires human judgment versus when it can proceed autonomously.
The 76% failure rate is not about bad technology. It is about the gap between what AI agents can do in controlled environments and what they actually do when released into complex, regulated, high-stakes workflows. The RAND Corporation breaks this down further: 33.8% of AI projects are abandoned before reaching production, 28.4% are completed but deliver no value, and 18.1% cannot justify their costs.
Collections is precisely the kind of domain where this gap is widest. The work requires judgment, cultural awareness, legal knowledge across jurisdictions, and the ability to recognise when a situation has shifted from routine to exceptional. These are the capabilities that AI agents struggle with most.
The Case for Human Networks
There is a reason that commercial debt recovery at scale has always depended on networks of specialists rather than centralised automation. Every market has its own legal framework, business culture, and enforcement mechanisms. Effective collections requires local knowledge, local relationships, and local judgment.
This is not a technology limitation that will be solved by the next model release. It is a fundamental characteristic of the work. When you are recovering a six-figure receivable from a company in Japan, the difference between success and failure is not processing speed. It is knowing that certain approaches which work in London will produce the opposite result in Tokyo.
Collecty operates across more than 100 countries through a vetted network of local collection partners, each with deep expertise in their jurisdiction. This is not a legacy approach waiting to be disrupted. It is the architecture that the problem demands.
The organisations that will protect their receivables portfolios are not the ones rushing to deploy AI agents. They are the ones who understand that debtor investigation, negotiation, and recovery across borders require the one thing AI agents consistently fail to deliver: reliable judgment in ambiguous situations.
AI will play a role in collections. It already does, in document processing, pattern recognition, and risk scoring. But the agent model, where AI autonomously manages multi-step, cross-jurisdictional recovery workflows, is the model that fails 76% of the time.
The 24% that succeed have something in common. They use AI for narrow, well-defined tasks with human oversight at every decision point. They treat compound failure as a design constraint. They never hand an autonomous agent the keys to a process where a single error creates legal exposure.
That is not a limitation. That is good engineering. And it is how collections has always worked best: skilled humans making judgment calls, supported by technology, not replaced by it.
Sources
Dataconomy, "Meta Head Summer Yue Loses 200+ Emails To Rogue OpenClaw Agent," February 2026
Hypersense Software, "Why 88% of AI Agents Never Make It to Production," January 2026
Dataiku, "MIT Says 95% of GenAI Pilots Fail: Here's How to Beat the Odds," 2025
Korra, "The $67 Billion Warning: How AI Hallucinations Hurt Enterprises," 2024
Towards Data Science, "The Math That's Killing Your AI Agent," 2025
Pertama Partners, "AI Project Failure Statistics 2026: The Complete Picture," 2026
Sarah Lindberg
International Operations Lead
Sarah coordinates our global partner network across 160+ countries, ensuring seamless cross-border debt recovery.
