Most denials are decided before the visit starts

Where practices say the money goes

Asked to name the single biggest leak in their revenue cycle, 48% of medical group leaders said denials and appeals. Front-end processes came second at 23%, billing and collections at 14%, coding at 13%. (MGMA Stat, 2026)

That ranking is worth sitting with, because denials and front-end processes are not really two separate categories. They are the same failure observed at two different moments.

What the denial numbers actually are

Be careful with the figures in circulation; most are hospital-weighted and get quoted as though they describe a five-physician practice.

• Practice-level: MGMA's benchmark for first-submission denials in single-specialty practices is about 8% — a figure that had not moved meaningfully since 2019.
• Medicare Advantage, peer-reviewed: across 270 million claim submissions, MA plans denied 17% of claims on initial submission; 14.6% for physician services specifically. About 60% of denials were resubmitted, and roughly 57% of all denials were ultimately overturned. Net effect: a 7% reduction in provider revenue from MA. (Health Affairs, 2025)
• Marketplace plans: KFF found 19% of in-network claims denied in ACA marketplace plans, with insurer-level rates ranging from 3% to 36%. Medical necessity accounted for only 5% of denials; administrative reasons, excluded services and prior-authorisation failures made up far more. (KFF, 2026)

Two of those numbers deserve to be read together. Fifty-seven percent of denials get overturned — and roughly 40% never get resubmitted at all. The denial rate is not the number that costs a practice money. The abandonment rate is.

The causes are upstream

Surveyed on why claims fail, providers put missing or inaccurate data at 50%, authorisation at 35%, and incomplete patient registration at 32% — and 68% said submitting a clean claim was harder than a year earlier. (Experian Health, 2025; vendor-run survey, worth reading with that in mind.)

None of those three are billing failures. They are all things that happened, or failed to happen, at the front desk.

The economics of fixing them there are well documented. CAQH's index of administrative transactions puts the provider cost of an eligibility and benefit verification at $7.97 when run manually and $2.18 fully electronic. A prior authorisation costs $10.97 manual against $5.79 electronic, and takes 22 minutes against 11. (CAQH Index, 2023) CAQH estimates 70 minutes of staff time available per patient visit when the full set of transactions runs electronically — summed across every transaction type, not a single saving.

Four places a small practice actually loses claims

1. Insurance captured as an image, not as data

A photo of a card in the chart is not a payer ID, a group number and a subscriber relationship. Somebody retypes it, usually while a waiting room fills up. Structured capture at check-in — with the card image kept as backing evidence — removes an entire class of transposition error, and makes the Medicare Secondary Payer questionnaire answerable rather than skipped.

2. Eligibility checked once, at scheduling

Coverage lapses between the appointment being made and the patient arriving. The check that matters is the one run on the day, against the plan the patient is actually presenting.

3. The diagnosis in the note and the diagnosis on the claim are assembled separately

This is the quiet one. When the visit diagnosis flows from the encounter into coding automatically, the linkage is defensible and consistent. When a biller reconstructs it from a narrative afterwards, you get mismatches that read as medical-necessity denials.

4. Nobody owns the denial

With 57% of all denials ultimately overturned, an unowned denial queue is the most expensive furniture in the practice. The fix is unglamorous: a worklist with an assignee and a date, not a report nobody opens.

What AI can and can't do here

Honest accounting: only 14% of providers report using AI to reduce denials, and among those that do, 69% report improvement — self-reported, with no disclosed measurement method. (Experian Health, 2025) Among health systems, 63% report using AI and automation somewhere in the revenue cycle — and of those, only 15% report having seen a positive return. (HFMA, 2025)

The useful applications are narrow and boring, which is a good sign. Reading a denial letter and extracting the reason code and the appeal deadline. Drafting an appeal that cites the note content already in the chart. Flagging a claim before submission when the documented diagnosis does not support the ordered service. Catching that a payer added an authorisation requirement for a procedure that did not need one last quarter.

What AI should not do is decide, unsupervised, what to bill. That is the subject of a separate piece, and the published accuracy numbers are not encouraging.

How ChartVoyant approaches it

The design principle is that the claim should be a by-product of the visit rather than a reconstruction of it. Insurance and Medicare Secondary Payer data are captured as structured fields at check-in. Eligibility runs against the plan on file. Intake answers reconcile into the chart rather than sitting in a parallel PDF. The visit diagnosis carries into coding directly. Denials land in a worklist with an owner.

None of that is exciting. It is where the money is.