Healthcare does not get the same AI playbook

Most of what the world has learned about deploying AI was learned somewhere else.

The consumer and enterprise playbook is well understood by now. Ship early, put the product in front of a lot of people, let them find the edges, fix what breaks. It is a genuinely good model. It has produced remarkable progress at remarkable speed, and it works because the cost of any individual failure is low, and the person who encounters it can simply try something else. That playbook still serves traditional software well. AI is different because its predictive nature creates risks of error that traditional automation does not.

Applied to AI, that model is disqualifying in healthcare.

Not because our industry is more cautious by temperament, though it often is. Because the cost of failure is carried by someone who did not opt in. When a coverage or payment decision is wrong, the person absorbing it is a patient who never chose to be part of anyone's iteration cycle. They cannot try a different model. They cannot route around the error. They frequently do not even know the decision was made by a system rather than a person.

So the edges cannot be found in production. They have to be found before. What that demands is AI that arrives already accountable.

That constraint is why we built Lyric around healthcare decision intelligence rather than around a model. Healthcare decision intelligence means the system carries the policy, the data, the deterministic logic, and the reasoning together, so a recommendation arrives with its justification attached, and a person can act on it or overrule it. It is a category built for exactly the problem this moment presents: decisions that have to move faster and remain explainable at the same time.

Here is what that means in practice.

Behind every claim is a person. It is your mother, your father, your sibling, your spouse, and your child. The work of payment accuracy is complex, technical, and at times arcane, and that complexity makes it easy to forget the human at the end of it. We do not have the luxury of forgetting.

Our clients ask us to use AI. We do, at scale and with real ambition. We are committed to using it in ways that are responsible, ethical, governed, auditable, and traceable. At Lyric, ambition and responsibility are the same commitment, and we will not trade one for the other.

The human makes the decision

Lyric builds tools, and claim decisions stay with our clients. Health plans use those tools to verify claims against policies, codes, and documentation, and they keep control of their own processing. Our systems surface evidence, context, and the reasoning behind a finding. People decide what to do with it.

The AI industry calls this human-in-the-loop. For us, it is the architecture itself. Design a system that takes the human out, and you have designed a system nobody should trust with a patient's care.

Our architecture reflects a choice we made deliberately. Either a person is part of producing a deterministic, traceable outcome with AI supporting the work, or AI is making a human review faster and more accurate. There is no third path where the system decides on its own. Our systems exist to take waste out of healthcare, and AI is what lets us do that at speed without giving up the deterministic answer.

Content authoring is the clearest example. We have spent decades building payment integrity rules from government and industry sources, applying our proprietary clinical and policy expertise to reach a determination that holds. The technology is architected around that process. AI identifies which sources matter, tracks what changed and where it came from, and assembles what a rule requires. Our clinical and policy experts make the determination, and every rule is tested comprehensively on Lyric42 before it goes live. That speed and accuracy would not be reachable without AI. The confidence comes from the fact that every rule carries its lineage.

This is what I call decision lineage, and it should travel with the recommendation rather than sit in a document somewhere. It is what separates an answer that sounds right from one that is. From the outside the two can look identical, which is exactly the danger. An assured answer that turns out to be wrong is worse than an honest acknowledgment of uncertainty. A system that cannot show its work does not belong in a payment decision.

Lineage also has to hold as things change. Research does not stand still, and the volume of evidence driving improvements in healthcare will only grow. When the inputs behind a determination change because the industry or the government publishes something new, that change has to move through our data model and our logic with the same traceability as the original answer. Architecture either supports that from the  start, or it does not.

Governance is a daily practice

Responsible AI is easy to declare and hard to operate. We operate it, and what we operate is governed AI. We  hold robust regulatory compliance and governance frameworks. We make specific commitments to our clients in contract about how AI is used on their behalf. And we review our AI work on a set cadence, at the most senior levels of the company, because real oversight has to reach beyond the team that built the system.

What we hold ourselves to is a short set of questions we should be able to answer at any moment. Is this system doing the job it was built for? Where does it stop being reliable? Has anything drifted since the last time someone looked, and who looked? Those have real answers, and a governance program is only as credible as its ability to produce them.

We also do not rely on our own view of ourselves. Lyric holds HITRUST r2 Certification and the HITRUST AI Security Assessment and Certification, and maintains SOC 2 Type II compliance, which confirms our controls work over time rather than on a single day. Independent, continuous verification matters because these systems change, and so do the conditions they run in.

The test we apply

Our AI deployments are evaluated against three questions. Does it remove burden rather than judgment? Can we show our work? Would we be comfortable explaining it to a client, a regulator, and a patient?  

Those questions rule out the thing I worry about most in this moment: taking a judgment call away from a person and hoping the model is right more often than not. AI extends our capability. Human judgment stays at the center. And our clients decide what happens next, on their own implementation, on their own terms. That decision is not ours to make.

What this makes possible

Here is what I think gets missed in the debate about how fast AI should move. In healthcare, traceability is the precondition for speed.

Consider how payment works today. Pre-payment editing already catches many errors before a claim is paid, but much of the work still happens afterward. A claim is submitted, paid, and then examined, sometimes months later, and the correction travels backward through appeals, recoveries, and abrasion between plans and providers. The system spends a striking share of its energy arguing with its own past.

Now imagine more of those decisions resolving earlier. Right the first time. Then earlier still, at the moment of submission, with the reasoning visible to both sides, so a provider has something to respond to instead of a denial with no explanation. Eventually, before the claim is written at all, when the question is still what should happen rather than what already did.

Every step in that direction requires the decision to move faster and be more explainable at the same time. Speed without lineage is unacceptable, because nobody will let an unexplained system make consequential calls in real time. Lineage without speed is just a better audit trail on a slow process. Together they change what the system is capable of.

What an operator owes the industry

No company should write these rules alone. What an operator can contribute is experience. We have run governed AI in a regulated environment for years, now at scale, and that surfaces things a design review rarely anticipates. Running these systems in production shows you where controls hold and where they strain, what changes once real volume is behind them, and what can fail without warning long after anyone last examined it. Knowing where those gaps open up is exactly the kind of knowledge the industry should be pooling.

So we will bring it. Through our work with industry associations, and anywhere else it is useful. And we intend to hold ourselves to whatever standard the industry arrives at, including where it asks more of us than what we do today. Our clients’ needs and the needs of patients everywhere will be our final arbiter.

We have been in this business since 1989. More than 100 health plan clients rely on us, and 60 percent of them have been with us for a median of 24 years. Trust like that is earned in increments and lost all at once. We take great steps to guard and protect the trust our clients invest in us. It is also, right now, the only thing that will let healthcare move as fast as it needs to.

Trust, truth, and transparency are the prerequisites that makes the ambitious version of AI in healthcare possible.