Why AI Keeps Failing at Insurance Documentation, Everi Labs at OTWorld 2026

This May, Everi Labs took the stage at OTWorld in Leipzig, the world’s leading trade fair and congress for prosthetics, orthotics, orthopedic footwear technology, and rehabilitation technology. Held every two years since 1973, OTWorld is where the entire field gathers: in its recent editions it has drawn more than 20,000 visitors from over 90 countries, alongside a world congress featuring hundreds of international speakers. For anyone working in O&P, it’s the place to be.

The 2026 edition leaned hard into the themes shaping the industry’s future such as robotics, digital processes, and artificial intelligence. That’s exactly where the Everi Labs session fit in. Delivered by co-founders Erik Dahlqvist and Ludvig Fraenkel, the talk was titled “AI-Driven Reimbursements, a guide to faster payments and fewer denials.”

The problem the talk set out to address

Documentation has never been harder, and the work that does get done is being rejected at record rates. Initial claim denial rates are climbing, and a striking share of orthotic and prosthetic denials trace back to one cause: incomplete documentation. Specialists now spend more hours documenting care than delivering it. Meanwhile, payers have deployed their own AI, pushing first-pass rejection rates up and shifting tens of billions of dollars out of provider revenue over the coming years.

So the industry reached for AI. And mostly, it has failed. A good part of the talk was spent explaining why.

Two approaches, the same mistake

The presentation walked the room through the two dominant ways AI is being applied to reimbursement documentation today.

The first feeds a model thin context, the latest evaluation plus the insurance criteria and asks it to generate justification text. The result is hallucinations, false positives, and invented justifications. Studies of leading models have shown they’ll repeat or even elaborate on planted errors in the overwhelming majority of cases. That’s a liability, not a tool.

The second approach adds more context: pattern data from past cases. It sounds smarter, but it introduces a different failure. The patient stops being a patient and becomes a probability distribution. The argument no longer references this person, it references statistical neighbors. And audits don’t reimburse averages.

As the talk framed it: Approach A invents what isn’t there, while Approach B generalizes away what is there. Both make the same fundamental error, they ask AI to author the case. Neither survives a five-year clawback review.

The principle that actually holds

Across the eight reimbursement systems Everi Labs has studied German GKV, French HAS, Swedish SBU, the British NHS, US Medicare and others, the forms differ but the questions converge. Every payer is really asking three things: Does this patient qualify? Is this the right device for this patient? And is there proof it works, for this kind of patient?

Answering those well isn’t about clever writing. It’s about connecting five things that already exist into one coherent story: patient history, product information, research and registry data, forecast patterns, and live payer policy. Built right, those five become a single connected case, the kind that still holds up when an auditor reviews it years later.

Where AI actually fits

This was the heart of the talk. AI is not the case-builder. It’s the connective tissue.

Give it the right job and the case builds itself: cross-referencing every patient record against every payer policy as those policies change; turning a library of approved and denied cases into searchable precedent; auditing a packet for gaps before submission rather than after denial; and flagging the specific risk a given justification carries for a given payer on a given code, based on how similar cases have actually fared.

What AI never does and this isn’t a matter of preference but, in Germany, of criminal law is write the medical record. The clinician is the author. AI connects the evidence; it does not manufacture it.

The window is now

The talk closed on why this matters at this particular moment. A regulatory wave is arriving; EUDAMED, MDR clinical-evidence deadlines, and EHDS all landing within a few short years and the clinical evidence regulators will demand is the same evidence payers already want. Documentation written today gets reviewed in 2030, under rules that don’t fully exist yet. The window to build the right infrastructure is roughly 18 to 36 months. Not five years. Not ten.

To continue the conversation, reach out anytime at erik@everilabs.com.