SEC EDGAR's full-text search, queried July 12, now shows a pattern boards should read before their next disclosure cycle opens. Public companies are filing agentic AI into the record twice: once as an announcement, and far less often as a risk.
In the first half of 2026, 561 public companies used the phrase “agentic AI” in a filing. In all of 2023, the count was zero. Of this year's mentions, 316 appear in 8-Ks, the form reserved for material events investors need to know about now, not filings drafted months later under full legal review. In the same 90-day window, only 217 filings across the entire market carried language closer to “our use of artificial intelligence” framed as a risk factor, and most of those sit beside the same company's own capability claims made elsewhere in the document.
The Split Is the Signal
The split matters more than the raw growth. An 8-K is a company choosing to tell the market something happened today. A risk factor in a 10-Q or 10-K is a company's counsel describing what could go wrong. When more than five times as many companies are willing to announce agentic AI capability as are willing to name it as an exposure, the filings are not incomplete. They are a decision. Management is narrating a growth story and leaving the governance story to whichever committee gets to it eventually, if it gets to it at all.
CACI International's March 8-K described its acquisition of ARKA Group as adding “space-based sensors and agentic AI-based software” to its geospatial intelligence business, a defense contractor speaking to shareholders in procurement language, not risk language. Marine Products Corp, a recreational boat manufacturer with no obvious AI product to announce, filed formal AI risk warnings the same quarter. Those two filings sit at opposite ends of the same market. The gap between them is the story.
The same EDGAR data carries a promotional tail worth naming. Microcaps, a SPAC merger prospectus, and a company holding a large treasury position in a cryptocurrency token now use agentic AI language in their own material-event filings, including a Form D registered under a name built entirely from the phrase. The vocabulary that established operators use to announce real capability is now also doing promotional work for issuers with no deployment behind it. That makes the phrase itself a due diligence flag in both directions: absence of governance language around a genuine deployment, and presence of the vocabulary around an issuer with nothing to govern.
Governing Implication
Delaware has not yet ruled on a Caremark claim built specifically on AI oversight failure, but the standard it will apply is already settled. Marchand v. Barnhill established that boards owe a heightened duty of oversight over mission-critical operations, and a board that let agentic AI enter its own 8-Ks as a strategic asset without ever naming who owns its authority, controls, and shutdown procedure has created its own evidentiary record of the gap. The filing that looked like good news on the day it was announced becomes the filing opposing counsel reads first.
Friction Point
A company's own material-event filings can now be read against it. An 8-K announcing agentic AI capability with no corresponding named owner in governance disclosure is not neutral. It documents that the board knew the system existed and chose not to describe who was accountable for it.
Action Point
Before the next 10-Q closes, pull your own company's most recent 8-K alongside its risk factor section. If the 8-K describes agentic AI as a capability and the risk factors are silent, generic, or copied from the prior filing, that gap is no longer an internal governance question. It is public record, and it is dated.
That gap is the argument developed in The Single Owner Protocol for Agentic AI: every agentic system needs one named human executive attached to its authority, controls, and shutdown procedure before the next material filing goes out, not after the claim arrives.