The AI-in-healthcare conversation is organized around tools. Scribes, imaging models, agents, decision support. A patient experiences none of that. They experience a sequence that starts when a physician says treatment should begin soon, and ends weeks later when it actually does.
Organize the question around that sequence instead and the answers move. Some of the places AI is ready are places nobody calls clinical. Some of the places it is being sold hardest fail in ways you will not notice for a year.
A referral arrives by portal, by email, or by fax. A fair number of the faxes are handwritten. Someone in the office reads what came in, decides whether the handwriting says what they think it says, and types it into the EMR. Then they check their own entry, because a transposed digit in a dose is not a clerical error. Then they check whether the referring office included everything required. Often it did not, so there is a call, a voicemail, and a callback the next morning. Only then does the benefits work start.
None of that is a decision. It is legibility, transcription, omission and hold music. The patient is at home the whole time and nothing clinical is in question.
This is where the days are. A system reads all three channels, pulls out what is there, checks it against what a complete referral for that therapy needs, and flags the gaps before anyone picks up a phone. Working with infusion centers, we built that path end to end. Time from referral to a scheduled chair went from three to four weeks down to three to four days.
The economics follow the calendar. A chair that sits empty on Tuesday cannot be sold back on Wednesday, so capacity lost to paperwork is gone permanently. A referral that takes a month is also a referral that can be lost, because the patient gets treated somewhere with a shorter queue. And the date treatment starts is the date revenue starts. None of that requires a new line item. It is the same volume, moved earlier.
The failure sits right next to the win. Where handwriting is ambiguous, the system has to stop. The temptation is to resolve it and keep moving, because a pipeline that flows looks better than one that pauses. Guessing a dose is the error that matters here.
In practice that means a confidence threshold per field rather than per document, because the drug name and the dose carry different consequences than the referring physician’s phone number. Below the threshold the system does not fill the field. It marks it, shows the nurse the original image next to the extracted value, and she settles it in seconds. Then you track how often staff override what the system did accept, by field. A rising override rate on one field tells you where the model is drifting long before anyone files a complaint.
That is also the part to test before launch. Not the path where everything works. The path where the system hesitates. Where does that case go, who picks it up, and how long does it sit on a short-staffed Friday.
Prior authorization is the one place in this sequence where both sides are automating at the same time. Providers are building systems to assemble and submit. Payers are building systems to review. That changes the risk in a way neither side sees on its own.
Two reviews get discussed as one here. Eligibility is a question about coverage and plan terms, and it is mechanical. Medical necessity is a clinical judgment against the payer’s criteria. They need different data and carry different consequences when they are wrong.
Used well, AI on the provider side is an evidence assembler. It checks a request against the payer’s published criteria before submission, which clears out a large share of denials that were never about necessity at all. They were about a missing attachment. On the review side it can approve, and approving faster is a real clinical good.
Denial is the line. A system can review and approve. A person owns a denial. Several states have landed on that rule independently, which suggests it is where the reasoning ends up when anyone follows it through.
The quiet failure is what happens when both sides scale without it. Two automated processes trade outputs at volume. Throughput rises on both dashboards. No one on either end can explain why a particular patient waited five weeks, because no person was ever in a position to form that explanation. The appeal still lands on a physician’s desk, and peer-to-peer is still a phone call between two people, one of whom now has to argue with a decision nobody made.
Ambient documentation has more published evidence behind it than anything else on this list. It is also the one most often justified for the wrong reason.
A study in JAMA Network Open last October followed 263 clinicians across six health systems using ambient scribes for thirty days. Burnout fell from 51.9% to 38.8%, with drops in cognitive load and after-hours documentation. Self-reported, no control group, thirty days, so hold it loosely. The direction matches what we hear from clients.
It gets sold as time saved. What it buys is attention. A clinician who is not typing is looking at the person in front of them, and what a patient volunteers when they are being looked at is different from what they volunteer to the top of someone’s head. That does not appear on a dashboard, which is why organizations that justify the spend on minutes per note struggle to explain why their clinicians would fight to keep it.
The failure is treating the draft as the record. A clinician signing twenty notes at the end of a clinic reviews a clean draft differently than a messy one, and the note is not only a clinical document. It is the basis for the code, the claim, and the medical necessity argument if that claim is questioned. It is also what a plaintiff’s attorney reads years later. A plausible note with a wrong detail is a harder problem than an incomplete one, because nothing about it invites a second look.
The tool works when the review is real, which means it has to be fast enough that reviewing beats rewriting.
Further along the path, the bar changes, and two different things get discussed as one.
The first is software regulated as a medical device. Imaging triage, risk stratification, anything making or supporting a diagnostic claim. That pathway is slow because it requires clinical validation and prospective evidence, and the slowness is the point. Tools that clear it carry evidence behind them that administrative tools never have to produce.
The second is a clinician opening a general-purpose model on their own and asking it about a case. That happens daily, it sits outside every governance structure in the building, and it is not covered by any procurement decision. The gap between those two is where most organizations have no position at all, and it is a more urgent question than which vendor to evaluate.
What we watch for is a team whose adjacent project went well assuming the credential carries over. Getting intake right is a software problem with a measurable answer. Clinical decision support is a medical evidence problem, and success at the first does not qualify anyone for the second.
Three things sit beneath every section above. They get treated as compliance questions. They are engineering decisions, and they get made early.
What the system can see. Scope it backward from the decision being made, not forward from what the integration exposes. An eligibility check needs coverage and plan terms. It does not need the clinical note. On a payment accuracy platform we build for health plans, the AI features run inside the security boundary the platform already had, with no outside model hosting, so adding AI added no new data exposure. That is the shape to aim for.
This matters more than the compliance framing suggests, because the thing at risk is not the record. People tell a clinician things they would not tell an employer or an insurer, and that willingness is what makes a chart accurate. It is also conditional and silent. A patient who suspects their disclosure feeds something they cannot see starts managing what they say. Nothing in the audit log shows it. Every decision downstream is then made on a slightly less accurate record.
How the systems connect. Interoperability is in better shape than the usual complaint suggests. FHIR APIs are widely deployed, national exchange frameworks are live, and pulling a chart from another organization is no longer the project it was ten years ago.
The problem is what those interfaces were built to do. They move documents and events for clinical review and for billing. An AI workflow needs something narrower and faster: a specific field, at the moment of the decision, with a path to write the result back. Reading is largely solved. Writing back into the source system is where these projects stall, because write access is governed tightly and for good reason, and it tends to be the last mile nobody scoped at the start.
The gap is wider between providers and payers than between providers. Prior authorization still runs through portals, faxes and phone trees, which is why that part of the sequence stays slow even at organizations with excellent internal integration.
So the plan that fails is rarely the one that assumed systems do not connect. It is the one that assumed the connection that exists delivers what the workflow needs.
Two risks follow. Connecting systems to make one use case work tends to widen what the AI can reach well past what that use case needed, and nobody revisits it afterward. Scope creep in healthcare AI rarely arrives as a new system, it arrives as a new feature on an old permission. The second is propagation. A wrong value used to sit in one place. Once systems read each other, it is taken as a fact downstream and confirmed by the system after that. The error is no more likely than before. It travels further before anyone looks at it, which argues for correction being cheap and immediate rather than for another review board.
Who answers for it. Organizations name an owner for the system and then find that nobody owns the decisions it makes. Those are different jobs. Before anything touches patients, three answers have to be yes. Can our people see what the system saw and what it produced. Can they correct it. Can they stop it.
Then name a person, not a committee, who answers for what it concluded about a specific patient on a specific day. If that person works for the vendor, you have not deployed a tool.
Before any of this gets decided, find out what is already happening. In most organizations people are already using AI on their own, outside any approved list, because the work is hard and the tool is a browser tab away. Finding out which of those touch patient data tells you more about your exposure than a formal assessment will.
The pattern across all five is the same. AI proposes and a person decides. Not a person in every step, which would put the bottleneck back where you found it, but a person who owns the decision, sees what the system saw, and can correct it or stop it.
AI is ready where a mistake is recoverable, where someone is positioned to catch it, and where somebody will notice if things get better. It is not ready where the error is quiet, where the evidence is clinical, or where nobody can be named afterward.
Start with the part of the sequence your patients spend the most time waiting in. It is usually not the part on the roadmap.
83% faster delivery. Under 10% rework. See exactly how Xceptor got there.
