The final version of the 2026 AHIMA/ACDIS Guidelines for Achieving a Compliant Query Practice Brief has been released. How technology is redefining the query process is an important concept in this updated version, as the industry continues to clarify when a documentation “prompt” becomes a query.
As the brief correctly states, it establishes “industrywide best practices for the clinical documentation query process,” but it is not enshrined in a regulatory framework. However, that could change in the future as the U.S. Department of Health and Human Services (HHS) Office of Inspector General (OIG) continues to identify “potentially fraudulent and abusive conduct” in the risk-adjustment process. That ongoing scrutiny should reframe how everyone involved in documentation integrity thinks about artificial intelligence (AI)-assisted documentation.
Querying has always evolved with its tools. Open-ended queries gave way to multiple-choice formats aligned with coding terminology. Paper queries gave way to queries embedded in the electronic health record (EHR), where an authenticated response became part of the permanent record (and, therefore, usable for code assignment).
Now we are in the era of automated queries, and some of them are difficult to recognize as queries at all. “Back-end” prompts, generated after a note is signed, still look and behave like traditional clinical documentation integrity (CDI) queries, and a human typically approves them before they reach the provider.
“Front-end” or point-of-care tools are different. They shape documentation while the provider is still dictating, often without a CDI or coding professional ever seeing what was suggested, accepted, or declined.
The updated query practice brief takes direct aim at that gap. It defines a query functionally, rather than by format, as any communication that “presents a provider with a specific diagnosis or documentation option for consideration in connection with a specific patient encounter,” regardless of “how it is labeled or where it originates within the technology platform,” full stop. When in doubt, the brief instructs, treat it as one. That is a deliberately expansive test addressing a blind spot within our industry of tools that behave like a query, but are not easily recognized as one, even though it influences coded claims data.
As the industry grapples with the evolution of queries, I suspect we may be entering an era in which regulatory guidance is necessary. The OIG Medicare Advantage (MA) Industry Segment-Specific Compliance Program Guidance addresses, albeit briefly, querying physicians via electronic medical record platforms. Their concern is regarding prompts generated by AI algorithms that encourage providers to add “diagnoses that patients did not have or that did not affect the care, treatment, or management of the patient” as a compliance risk. Read that again closely. OIG is not describing a hypothetical. It is describing, almost word for word, the same front-end prompting behavior the practice brief was written to govern, except OIG is not offering voluntary best practices, but rather identifying a fraud vulnerability.
I do not think this is a coincidence, and I do not think CDI professionals should treat it as one. When a professional consensus document and a federal fraud-and-abuse guidance document independently converge on the same behavior within months of each other, that convergence is itself the signal. It tells us that the industry’s own internal debate about AI query oversight is no longer academic. It has become the government’s business.
So, what do AI documentation integrity tools get right, and what do they get wrong? The answer, according to researchers examining generative AI’s reliability more broadly, is often quite a lot, but not consistently, and rarely transparently. It is this lack of transparency that can be problematic. A Rutgers Business School analysis catalogued the now-familiar flaws of hallucination, source confusion, and limited transparency into how a conclusion was reached that make these tools simultaneously useful and risky. Linguistic fluency has improved dramatically.
Unfortunately, reasoning has improved only marginally. Today’s AI influenced query prompts can sound certain about a diagnosis it should not be suggesting at all.
When humans guide querying, we understand the nuances associated with code assignment. Yes, providers should be documenting using terminology that can be captured within ICD-10-CM, but not all diagnoses in the health record are reportable. This is where providers can be led astray by AI, being prompted to respond with more certainty than can be supported by the totality of the health record.
What may be “true” at the point of documentation may not be accurate once the pieces are assembled to reveal the completed puzzle. This could create situations in which additional queries may be necessary to clarify documentation obtained through prompts, which is unlikely to be a physician satisfier.
That is precisely why the query practice brief insists on human oversight, rather than vendor assurance. It states that organizations “should not assume that a vendor-supplied or EHR-integrated tool produces compliant queries by default,” and that accountability for every AI-generated query stays with the organization, not the vendor.
The OIG guidance makes a structurally similar point about risk adjustment, whereby MA parties are expected to build their own safeguards, rather than rely on a platform’s design, to keep them compliant.
Who inside a hospital is positioned to catch a front-end prompt before it becomes a documented diagnosis with no audit trail behind it? Not coding, which reviews the record after the fact. Not CDI, which is typically structured around post-documentation reviews. The physician advisor (PA) occupies a rare seat at that table: someone who still authors or reviews clinical notes, understands why a query is being made, and recognize a documentation nudge for what it is, even when it never carries the word “query” anywhere on the screen.
Including PA in AI governance is not a new argument. It is, however, a considerably more urgent one, now that the case for their involvement is being made by OIG as much as by AHIMA and ACDIS.
None of this matters only at the organizational level. It is also a personal reputational risk for the physician who signs the note, and providers who treat an AI-generated diagnosis as someone else’s problem are misjudging where that risk actually sits. The American Medical Association (AMA) has called AI liability “a top issue for physicians” and is pushing regulators toward “appropriate apportionment of liability for AI errors.”
Malpractice carriers are blunt about it; CM&F Group, a professional liability insurer, has stated that “the clinician whose name is on the chart is the clinician who bears responsibility for what’s in it.” Providers who wave an AI prompt through to save a few minutes are not protecting their time, but risking something considerably harder to get back than a query response.
