By BRIAN FENGLER
Each evening within the emergency division, the digital well being document hums with small calls for for consideration. A nudge about sepsis. A best-practice advisory. A documentation suggestion that arrives between sufferers and asks for yet another click on.
Most of us deal with these prompts as background noise. Nearly none of us cease to ask whether or not a given nudge is, actually, a question, or whether or not it was constructed to the identical normal a scientific documentation specialist has to satisfy.
In August, that query bought an official reply. And it isn’t the reply many well being IT distributors have been promoting.
The American Well being Data Administration Affiliation and the Affiliation of Medical Documentation Integrity Specialists launched the 2026 update to their Guidelines for Achieving a Compliant Query Practice, formed by suggestions from nearly 200 individuals and groups. The precept is obvious: “Whether or not a question comes from an individual or know-how, the usual for compliance ought to stay the identical.”
Then comes the sentence that ought to finish a well-recognized gross sales declare: “Organizations shouldn’t assume {that a} vendor-supplied or EHR-integrated software produces compliant queries by default.”
For years, “compliant by default” functioned as an unstated promise round AI-enabled scientific documentation and utilization evaluate instruments.
If a system might draft a question, the pondering went, it might draft a authorized one. The 2026 pointers shut that hole. Compliance is not one thing the software program is presumed to get proper. It’s one thing the well being system that buys it has to show.
I’m an emergency doctor. I co-founded EvidenceCare in 2016 as a result of I watched clinicians drown in prompts lengthy earlier than generative AI arrived. What the brand new pointers clarify is that the immediate drawback was by no means solely about alert fatigue. It was additionally about who’s accountable when a machine asks a physician to doc one thing that will change how a case is coded, scored, or paid.
A question is what it does
Part XI of the rules, titled “Question Expertise,” brings 4 classes of instruments into scope: computer-assisted doctor documentation, computer-assisted coding, giant language mannequin and generative AI platforms that draft or ship question language, and EHR-integrated advisories and alerts.
The rules outline a question by perform, not by label. A communication is a question when it “presents a supplier with a selected prognosis or documentation possibility for consideration in reference to a selected affected person encounter,” and that definition applies “no matter how the communication is labeled or the place it seems inside the know-how platform.” Prompts, nudges, advisories, and alerts are named explicitly. A software marketed as passive determination assist doesn’t obtain an exemption due to what the seller calls it.
The definition is narrower than it first sounds. Passive scientific determination assist that surfaces reference data with out steering a supplier towards a selected response for a selected affected person isn’t a question. The sensible take a look at is that this: does the software level a clinician towards a specific reply for a specific affected person? When that’s unclear, the rules say the communication must be handled as a question and held to each relevant compliance normal.
On a busy shift, that distinction isn’t educational. “Listed below are the sepsis standards” is reference materials. “Think about documenting extreme sepsis for this encounter” is a question. Calling the second a nudge doesn’t change what it’s.
The identical guidelines as a handwritten be aware
An AI-generated question is held to the identical normal as one written by hand. It should embrace related scientific indicators sourced from the well being document. It should keep away from subjective interpretation. Its title, monitoring identifier, and topic line should not reference a desired prognosis, a reimbursement affect, or a top quality measure. Reply choices have to be clinically supported, should embrace an “different” possibility or room for an alternate clarification, and should not use formatting, emphasis, or ordering that steers the supplier towards a most well-liked response.
Two provisions are simple to miss, and they’re the place loads of product design quietly fails. Automated templates “shouldn’t auto-populate the reply discipline or recommend diagnoses with out supplier validation.” And if a technology-generated question doesn’t produce the response a company needed, sending a follow-up guide question solely to override the supplier’s earlier reply is noncompliant until new scientific indicators exist.
The rules are candid about the place the chance originates: software configuration, coaching knowledge, and system design, particularly when these selections introduce directional bias or put reimbursement forward of scientific accuracy.
What “compliant by design” would require
The rules set the end result and go away the structure to the seller. Assembly that end result is unlikely to imply operating a big language mannequin over a chart and transport fluent textual content. It often means selections made earlier than the primary question is generated.
Templates must constrain the mannequin, and the mannequin needs to be examined towards these constraints. If the formatting logic can not daring, spotlight, or reorder reply choices, the output is formed earlier than supply as a substitute of cleaned up afterward. Reply fields can not pre-select a response. Medical indicators have to be traceable, so a company can present how every indicator was sourced and why every reply possibility was included. And human evaluate has to dwell someplace within the workflow. The rules permit evaluate earlier than supply or “ongoing structured auditing of question output,” they usually count on common structured auditing when there is no such thing as a pre-delivery evaluate.
From the product facet, that’s the distinction between an AI that writes nicely and an AI that may survive an audit.
The client owns the chance
On accountability, the rules go away little room for interpretation: “Organizational course of homeowners retain accountability for each question delivered to a supplier, together with these generated by automated or AI-assisted processes.” Vendor vetting and compliance evaluate of know-how question instruments sit alongside human evaluate, audit frequency, workers coaching, and the method for reporting and correcting noncompliant queries. The burden of proof sits with the well being system that adopts the software, not the corporate that offered it.
Regulators are shifting in the identical course. As summarized by Arnold & Porter, the HHS Workplace of Inspector Normal listed “AI-generated prompts encouraging unsupported coding” amongst “doubtlessly abusive threat adjustment practices” in its March 2026 compliance steerage for Medicare Benefit.
Question observe drew enforcement earlier than AI entered the image. In December 2024, Independent Health agreed to pay up to $98 million to resolve False Claims Act allegations {that a} subsidiary retrospectively searched medical information and queried physicians for data that will assist further diagnoses and better threat scores. The allegations had been resolved with out a willpower of legal responsibility, and no AI was concerned. What the case exhibits is how intently regulators already study the question course of.
Put that settlement beside the OIG language and the 2026 pointers, and the sample is tough to overlook. Question observe was already an enforcement goal. Synthetic intelligence didn’t invent the chance. It scaled it, then packaged it as a function.
For leaders evaluating an AI-enabled CDI or utilization evaluate software, 4 questions now belong subsequent to the acquainted ones about accuracy, integration, and pace:
- How does the software forestall a number one question, in particular phrases, as a substitute of stating that it does?
- Which instruments in the environment, together with EHR alerts, meet the definition of a question?
- Who opinions queries earlier than supply, or how typically will we audit them afterward?
- What’s our course of for correcting a noncompliant question as soon as we discover one?
Earlier than the following audit cycle
The central level of the 2026 replace is that “compliant by default” was all the time a advertising and marketing declare, and it’s now a declare the adopting group has to check for itself. Question quantity, supplier burden, and the dimensions of documentation evaluate make some automation inevitable. Executed nicely, it stays one of many higher solutions to the stress CDI groups are below.
However “carried out nicely” now has a written definition, and it’s extra particular than many merchandise in the marketplace had been designed to fulfill. Organizations that deal with the replace as a design requirement, and that evaluate each alert, immediate, and AI-drafted question towards Part XI, shall be in a stronger place when the following audit cycle arrives. Extra importantly, they are going to nonetheless deserve the belief that the question course of is dependent upon.
Brian Fengler, MD, is the Chief Medical Officer and co-founder of EvidenceCare, a scientific efficiency enchancment firm he co-founded in 2016.
