The AI You Shipped Last Quarter Is Still Running
The person who deployed it moved on. The model didn't.
Somewhere in your company, a model is still making decisions.
It went live six months ago. Or eighteen. Somebody wrote the vendor selection memo, somebody signed the SOW, somebody flipped the switch. The person who did those things has moved to another team, or another company. The model is still running and nobody is on the line.
Cluster 193f0b8d in the NormSense database documents this as a pattern. Seventeen fresh observations in the last seven days. The cluster description is blunt. “AI deployers industry-wide abandon deployed models without named human accountability for ongoing fitness, scope drift, or retirement.”
Six independent source types converged on it this week. Academic preprints, ethics journals, MLCommons, Semantic Scholar, IEEE, GitHub topics. That’s the strongest structural pattern signal in the entire pull.
What the SEC filings show
Nine major public companies filed AI-related disclosures with the SEC in the last seven days. Atlassian. H&R Block. Amcor plc. Western Digital. Applied Industrial Technologies. ResMed. Intapp. Avnet. General Mills.
Each disclosed AI deployment. None of them disclosed who owns the deployment after it’s live.
Atlassian mentions GDPR automated-decision transparency obligations but doesn’t disclose implementing them. Western Digital describes an internal AI policy as a risk-management instrument without publishing its scope or enforcement mechanisms. H&R Block deploys generative AI tax assistance to DIY filers without disclosing error rates or contestability. ResMed puts AI into home healthcare patient engagement without explaining how patients contest algorithmic decisions.
The 10-K genre is designed to disclose risk to investors. It isn’t designed to name the human at the company who owns the AI system’s ongoing behavior. So the filings describe deployment as an event and treat the aftermath as ambient AKA abandonment.
The Paylocity cluster
Cluster cda013f0 had the single largest weekly accumulation in the corpus: thirty-one fresh observations in seven days. Paylocity embeds AI into HR and talent workflows affecting employees. No disclosed model logic. No employee contestability mechanism.
Paylocity isn’t the abandonment story. Paylocity is the shipping-without-accountability story that becomes the abandonment story eighteen months from now, when the person who bought Paylocity is at a different company and the model is still adjudicating who gets flagged for performance review.
The cluster converges specifically on NYC Local Law 144, EEOC charges, and EEOC guidance archive. The regulatory infrastructure catching Paylocity is employment law. The people running Paylocity’s AI on your workforce are your HR SaaS vendor. The person in your company who understood how it was configured left last spring.
Meanwhile, civil society noticed
The NAACP filed a formal Congressional lobbying disclosure this week. Three related observations landed in the corpus from the same LDA filing. The organization is lobbying Congress on AI in criminal justice affecting Black communities. On AI equity in healthcare eligibility. On AI civil rights protections generally.
This is a legal filing under the Lobbying Disclosure Act. A major civil rights organization is now on formal record with Congress about AI deployment gaps.
Institutional actors are entering the debate.
The response forming
Cluster 06d29049 documents pre-execution policy enforcement for AI agents. Nine fresh observations this week, five independent source types. Governance tooling that stops AI actions before they happen, rather than auditing them after.
This is the counter-move to the abandonment pattern. If nobody at the deploying company is going to own the AI’s ongoing behavior, the industry is building tooling that constrains the AI’s behavior structurally at the runtime layer.
It won’t be enough. Runtime constraints don’t answer the accountability question. They just narrow what the abandoned model can do.
But the corpus is capturing the response emerging alongside the pattern. That’s what a live corpus does.
So
The AI you shipped last quarter is still running. The person who shipped it is not.
Nobody is assigned to notice when the model starts scoring differently. Nobody is assigned to retire it when the world it was trained on has changed. Nobody is assigned to know why it declined the loan or flagged the employee or denied the claim.
Once you see the gap, you can’t unsee it. Your bank has it. Your health system has it. Your employer has it. Every 10-K this quarter has it.
The AI is still running…somebody should probably be on the line.
Zach, see you in the cluster pages.