AI is now on both sides of your DME claims

Most of what gets written about AI in DME billing is aimed at the supplier: buy this, automate that, watch your denials fall. The more consequential development in 2026 has been on the other side of the desk. Medicare itself now runs an AI-assisted review model, and a handful of states have started legislating who is allowed to let software decide a claim.

That reframes the question. It is no longer only "should we automate?" It is "what does our documentation have to survive now?"

Medicare is running the pilot

The WISeR model (Wasteful and Inappropriate Service Reduction) launched on 1 January 2026 in six states: Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. It runs for six years. Prior authorization requests were accepted from 5 January 2026, and services rendered on or after 15 January 2026 became subject to it (DLA Piper).

The participants are private health technology companies, not providers. They use AI and machine-learning tools to make prior authorization and pre-payment review determinations, and, in the part that has drawn the criticism, they are paid a share of the savings associated with services that are denied. CMS requires a human clinician to give a second opinion before a request is denied, and vendors face payment adjustments or removal for inappropriate denials (KFF).

WISeR’s 2026 service list is not DMEPOS, but it is adjacent: skin and tissue substitutes, electrical nerve stimulator implants, incontinence control devices, epidural steroid injections, and knee lavage. Covered places of service include the home.

It very nearly did not survive. In May 2026 the GAO determined that WISeR meets the Administrative Procedure Act definition of a rule, which opened a Congressional Review Act window. The Senate defeated the resulting resolution of disapproval 46-50 on 17 July 2026, so the model is still running (Healthcare Dive).

Clean submissions are now worth money

WISeR uses the same carrot CMS just applied to DMEPOS: providers who maintain high approval rates can earn exemptions from future prior authorization requirements. On the DMEPOS side that programme went live on 1 June 2026, and the threshold is a 90% provisional affirmation rate on at least ten initial requests, scored per PTAN and per DME MAC jurisdiction.

The business case for automation in DME is no longer a vague efficiency argument. It is the distance between your current affirmation rate and 90%.

That is a calculable thing, and it is the right place to start. If a supplier is sitting at 82% and a pre-submission completeness check catches the missing face-to-face note on one request in ten, that is the difference between doing prior authorization all year and not doing it at all. Work out the current rate first. Everything else is guesswork until you have it.

The appeal gap is where the money quietly goes

CMS data on Medicare Advantage makes an uncomfortable point. In 2024, MA insurers made 52.8 million prior authorization determinations and fully or partially denied 7.7% of them, about 4.1 million. Only 11.5% of those denials were appealed. Of the appeals that were filed, 80.7% were partially or fully overturned (KFF).

Read those last two numbers together. Four out of five appealed denials were wrong, and nearly nine out of ten denials were never tested. That gap is not a clinical judgement, it is a staffing one. Appealing takes hours, small teams triage, and the denials that get dropped are the ones nobody had time for, not the ones that deserved to be dropped.

This is the most honest automation argument in the whole category. Automating appeal drafting and tracking does not make you right more often. It removes headcount as the thing deciding which denials get challenged.

1 January 2027 is the real deadline

The interoperability rule known as CMS-0057-F requires impacted payers (Medicare Advantage organisations, Medicaid and CHIP managed care entities, state Medicaid agencies and QHP issuers on the federal exchanges) to run a FHIR-based Prior Authorization API by 1 January 2027. Its operational provisions already took effect for rating periods beginning on or after 1 January 2026: standard decisions within timeframes that cannot exceed seven calendar days, specific denial reasons furnished to providers, and public posting of annual prior authorization metrics. Medicare FFS DME MACs are not covered by the rule.

Separately, more than sixty insurers signed a voluntary pledge in June 2025. AHIP and the Blue Cross Blue Shield Association reported in April 2026 that it had produced an 11% reduction in prior authorizations across covered markets, 6.5 million fewer, including more than 15% in Medicare Advantage. Two further commitments land on 1 January 2027: standardised electronic submissions, and answering 80% of electronic approvals in real time (BCBSA).

One caveat worth stating plainly: that 11% reduction is described as covering services with clear evidence-based clinical guidelines, and the announcement does not break out DME. Do not assume your categories were part of it.

The practical consequence is a question to ask your software vendor in the second half of 2026, not the second half of 2027: can this platform consume a FHIR prior authorization response, or will my team be re-keying data that arrived structured?

And the states now regulate what you submit

Six states enacted laws in 2026 restricting AI in health insurance coverage decisions, with staggered effective dates: Washington SB 5395, Maryland HB 1563, Indiana HB 1271, Alabama SB 63, Utah SB 319 and Georgia SB 544 (Holland & Knight).

Most of them constrain payers. Indiana HB 1271, effective 1 July 2026, is the one that also constrains providers: it restricts submitting AI-generated claims without professional verification. If you bill into Indiana and you have an AI-assisted coding or claim-generation step, you need a documented human review point in that workflow, and you need to know which states each claim is going to.

Where automation actually pays in DME

The category-specific answer is documentation, because that is where the errors are. DMEPOS carried a 21.4% improper payment rate, over $1.9 billion, in the FY2024 CERT report, against an overall Medicare fee-for-service rate of 7.66%, roughly three times the programme-wide figure. The dominant error type is insufficient documentation, not fraud (PAAS National).

The uses HME operators describe in trade coverage this year cluster on exactly that: capturing PAP adherence in real time without manual review, aligning resupply to documented usage so items lacking medical necessity are never billed, keeping time-stamped compliance records of therapy and patient outreach, and segmenting reporting by payer, physician, location and therapist (HME Business).

Notice what is not on that list: nothing about generating claims faster. Automation applied downstream of a bad order produces well-formatted denials at speed. Applied at intake, it fixes the thing CERT keeps flagging.

One last note on evidence, since this is a field crowded with confident numbers. While researching this piece we found no published, independent survey of AI adoption among US DME providers, and the widely-repeated DME denial-rate figures, the fifteen-to-twenty-five percents that appear in a dozen vendor blogs, trace back to no primary source we could find. They are not in this article for that reason. Be equally suspicious when a vendor quotes one at you.