Scope for each practice is confirmed in the engagement agreement, including systems, payers, and queues.
Unreviewed automation is a fast way to be confidently wrong
A model can group denial letters or highlight a missing modifier and still misread a remark code or invent a diagnosis from a thin note. Practices are being offered tools that submit or adjust without a named reviewer. The risk is a scalable version of the old problem: a guess, written in a confident tone. Any assistive tool in this model has a human owner and a log.
What the work includes
- A written list of tasks a tool may draft, such as sorting denial reasons or summarizing a remit code
- A prohibition on unattended code changes, diagnosis changes, and claim submission
- Human acceptance or rejection of each material suggestion
- A log of suggestions and decisions
- A data-handling review before any record is sent to a tool, including whether the vendor will train on it
- A kill switch that returns the queue to a fully manual path
How the work moves
Step 1
Choose a narrow task
The first task is classification or summarization, not payment. If the task can change a balance, it is out of the pilot.
Step 2
Limit the data
The tool receives the minimum text required. Public website forms are never part of the path. A business associate agreement is required before PHI is involved, and vendor training terms are read.
Step 3
Review every material suggestion
A person accepts, edits, or rejects the draft. Acceptance is attributable to a user, not to a shared login.
Step 4
Measure harm, not only speed
The pilot counts incorrect suggestions and any that would have changed a code or a balance if a person had not caught them. Speed alone is not success.
Operational benefits
- Faster sorting of repetitive correspondence when the pilot holds
- A written boundary staff can defend
- Less pressure to adopt a tool because a brochure said it bills
- An exit back to manual work
Common challenges
- Models produce fluent errors
- Vendor terms may allow training on customer data
- Bias in historical coding can be repeated at scale
- Staff may over-trust a suggestion that looks finished
Technology
No specific model is endorsed here. If a tool is used in an engagement, it is named in the agreement, restricted to approved tasks, and kept off the marketing site. This page does not mean the capability is live.
Where specialty changes the work
The safest early use is the same in every specialty: grouping denial text and remit codes. Specialty coding suggestions, if they are ever piloted, need a coder who knows that specialty and a sample audit. They are not the starting point.
Questions
Is AI billing already running?
This page describes a proposed control model. It is not a statement that an unattended or even an assisted model is currently processing claims for clients.
Can we paste notes into a public chatbot?
No. Clinical notes and identifiers do not belong in public tools or in the forms on this website. Any assistive tool would have to be contracted, access-controlled, and limited.
Will a model pick E/M levels?
Not in this proposal. Level selection remains a coding judgment based on the note. A tool that suggests a level would require a separate, audited pilot and still a human decision.