UnitedHealth wants AI to kill a third of prior authorisations. The test is whether costs bend

The target: a third of prior authorisations

UnitedHealth Group has set the clearest AI target in American healthcare this week. Dan Kueter, the company's chief commercial officer, told A16Z that UnitedHealth has an internal objective to achieve touchless prior authorisation, and aims to eliminate 30% of prior authorisations. The paediatric space is a priority for elimination, Kueter said, because providers and payers agree there. Prior authorisation is the paperwork gate between a doctor's order and a patient's treatment, and it is one of the most hated processes in the system. Cutting a third of it would be felt by every clinic in the country. m:…6411

Kueter's larger claim is bolder. He said AI can become deflationary for healthcare in America when it begins doing things that have not been possible before, and that AI could begin bending healthcare costs within a single-digit number of years. That is a different claim from saving money on administration. Early AI installations in healthcare have been administrative, Kueter said, and have contributed to inflationary dynamics through more comprehensive coding. In other words, the first wave of health AI made billing more thorough, which made costs higher. The second wave has to do something new, not just do the old paperwork faster. m:…6411

A market built for upstarts

Kueter also described a market in motion. He said the commercial insurance market is experiencing a generational replacement cycle that creates opportunities for upstarts, and that self-funding is moving down-market to smaller employers through level-funding and stop-loss financing mechanisms. UnitedHealth is adopting an omni-channel front door strategy, meeting consumers where they are, including through third-party AI agents. Human brokers and consultants will likely persist for a long time but will be accelerated and assisted by AI, Kueter said. m:…6411

The implication for health AI builders is that the customer is changing shape. Smaller employers buying self-funded plans, consumers arriving through AI agents they do not control, and a payer that wants to remove itself from a third of its own approval gates. A health AI product that only talks to the old front door will miss the traffic. m:…6411

The evaluation question

Google for Health published a framework this week for deciding whether health AI is actually good enough. The company published a whitepaper detailing an end-to-end evaluation approach, developed in partnership with Dr. Elizabeth McGlynn and Dr. Dana Gelb Safran. The framework evaluates AI models across three areas: accuracy, safety and usefulness, using closed-question benchmarks, autoraters and expert human review, and is designed to assess both common and high-risk health scenarios. Google says it uses the framework to assess whether its models provide useful and accurate responses and to promote safe interactions in health contexts, and that the goal of sharing the methodology is to support the health ecosystem in building trustworthy AI tools and promote transparency. m:…6035

The framework matters because the prior authorisation bet cannot be judged on speed alone. An AI that approves the wrong care quickly, or denies the right care politely, fails the safety and usefulness tests even if it clears the queue. Google's three pillars are a reminder that accuracy without safety is not a product in healthcare. m:…6035

Training the humans, not just the models

OpenEvidence launched Learn this week, a personalised medical education platform designed for every stage of a clinician's career. It includes three tools: Test Prep, Preceptor and Flashcards. Test Prep contains over 50,000 board-style questions covering USMLE, COMLEX, and specialty and subspecialty boards. Preceptor uses Socratic, case-based tutoring, reveals case findings only when requested, and prevents premature closure by requiring specific checkpoints. Flashcards use spaced repetition to strengthen long-term retention. m:…7432

The Learn tools are powered by the same peer-reviewed research and clinical guidelines as the core OpenEvidence platform, according to the company, and OpenEvidence intends to hold them to a higher standard than engagement metrics. It is partnering with medical schools and residency programs to study the impact of Learn on clinical judgment. The contrast with the payer story is instructive: one side of health AI is removing human gates, the other is training human judgment. Both will be needed if the 30% target is going to survive contact with real patients. m:…7432