What the Note Leaves Behind

On ambient AI, patient navigation, and the work that begins in the exam room

By Luis Cisneros, CEO of Serelora

There are things a person will tell a physician only after they feel safe enough to say them. Not the official history. The other one. A question about a missed prescription starts as an apology, drifts through some story about being busy, and lands, eventually, on the number that appeared on the pharmacy screen. She has been stretching the medication. She meant to call. She was embarrassed, or overwhelmed, or hoping she could make it to the next visit without having to explain how ordinary a hard month had become.

Give that answer a little room and you learn something no intake form ever managed to catch. Medicine allows a strange intimacy between people who would otherwise remain strangers. Someone examines her body and, in the same half hour, hears the circumstance she has been working hardest to keep private. That is not sentiment. That is the most valuable data the clinic is likely to see all day, if anyone is still listening when it arrives.

It also has a use. The physician can attach what she said to the treatment under discussion, sort out what happened, and ask whether she wants help. A private worry becomes the start of a response, with the patient still in the decision. The rest of the team can work from something that was actually established in the room rather than inferred from a checkbox.

An ambient scribe belongs in that exchange. It keeps the conversation while the physician stays with the person talking. Watch a clinician split herself between a patient and a keyboard for ten years and the appeal is not mysterious. A decent note holds the explanation, gives the next clinician a chance, and spares another evening spent reconstructing a visit that already happened.

The industry noticed. Ambient AI went from novelty to furniture in a stretch of time that still feels slightly unreal. Documentation burden dropped. Clinicians looked less hunted. Hours came back. Then the business case arrived, right on schedule, because nothing travels faster in this business than a tool that returns time. The early results are real. Fine. Now what happens to the encounter data?

Informatics has been living with that question for a long time. Structured documentation gave cleaner material for reporting, quality programs, population health, research, and the rest of the downstream machinery. It also turned visits into a tour of boxes, which is the exact fatigue ambient AI is now being paid to relieve. The pendulum swung. Systems can now produce fluent clinical narrative at a scale no template ever managed. Good for the room. Less good if everyone treats the finished note as finished work and assumes some later model will make sense of the pile. Traditional analytics was built to read fields. It is about to inherit industrial quantities of prose and pretend that is the same thing.

The prose itself is not a neutral container. Physicians used to write short because writing a note was an act of judgment. What went in had been selected. What stayed out had been weighed. Ambient drafts tend to run the other way. Evaluations keep finding the same trade. The generated note is often more thorough and better organized. It is also less succinct, more prone to invention, and easier to misread as complete. Length is not richness. It is sometimes only volume.

That volume is produced by a cycle the clinic has lived with in slower form for years. Speech is compressed into a transcript. The transcript is expanded into a polished note. The note is contracted into codes, orders, and later summaries. Another model then expands the result again when someone asks what the chart means. Copy-forward and templates were an earlier version of the same motion. Reuse a paragraph, trim what looks stale, let the next writer inflate it to fit a new visit. Each pass changes the material. A relative time becomes a calendar date. An unexamined system review becomes a string of negatives. A dose is halved. A working diagnosis falls out of the assessment while the plan that depended on it remains. A sentence about memory or adherence appears that nobody offered. The error does not look like an error. It looks like a note.

Two paths from here, and the cycle sits inside both of them. In one, ambient AI sits close enough to the record to recognize what was said, parse it, and map it into standardized elements without sanding the conversation down to dust or completing a pattern the visit never established. You keep the narrative. You also keep dates, sources, and uncertainty attached to the facts they belong to, so a later contraction does not have to guess. In the other, the chart fills with free text at speed and a second layer of language models is asked to interpret it later. That version sounds modern. It is also how you lose reproducibility, granularity, provenance, and any reliable sense of which sentence still belongs to the patient. Read it Tuesday and it means one thing. Read it Friday and it has become a different story with the same nouns.

The first path is already partly visible, and so is its half-finished version. Tools draft laboratory orders, suggest ICD-10 codes, assemble prior authorization material. Take that seriously. Moving work out of the after-hours fog is not a small thing, and the patient has every reason to want it done cleanly. Independent evaluations of ambient platforms have also found substantial rates of omitted or distorted clinical elements, fabricated findings, and notes accepted with little editing. The hope is not fake. It is incomplete.

Cognitive surrender makes the incompleteness hard to see. Writing a note used to force a pause in which a clinician decided what the visit had actually been. Reviewing a fluent draft is a different job, and a weaker one. The software reduces the cognitive load of documentation. That is the advertised gift. It is also the mechanism. Once the page sounds like a physician wrote it, the impulse is to sign. Some deployments have left a large share of drafts untouched. Clinicians have described not recognizing the note as theirs and not remembering the patient at the next visit. Automation bias does the rest. The official version is trusted because it is official, and because checking it would spend the time the tool just returned.

Then the woman who could not afford her medication can still leave with an approved prescription she cannot collect. The laboratory order can be perfect while she has no workable way to reach the laboratory. Software will advance every stage it knows how to recognize. The original obstacle will wait outside, unbothered. Next month the chart will contain a polished account of a plan she was never in a position to follow. Sometimes that account will also contain a sentence she never said. The documentation problem got solved. The data problem did not, because the data problem was never just a shortage of codes. It was a shortage of fidelity through each transformation.

Even extraction does not close it. A system can identify the diagnosis, recognize the test, hang the right codes on the visit, and still leave the team guessing how treatment is actually unfolding. The result is tidy. It is useful for a certain kind of administrative work. A computable encounter is not a coded encounter with better manners. Coding says what the clinic wrote down after the last expansion. If the word computable is going to mean anything, it has to say what became of the plan after she walked out, and it has to say so without laundering an invented detail into a structured field.

Think about the record built for a claim. Diagnosis names the condition. Service code names the work. Supporting text helps a reviewer decide whether to pay. All legitimate. Clinics have to account for what they did and get paid for doing it. Trouble starts when that account is treated as a full description of care. A documented test or a paid claim does not tell you whether the result was used well, or whether the patient could carry the next step. Analytics will count the order anyway. Analytics are not shy. They will also count a fabricated finding if someone allowed the finding to become a code.

She may have picked up the prescription and taken it for four days before stopping. She may have collected it and misunderstood the instructions. She may have developed a problem she has been waiting to mention. A dispensing record tells you the medication left the pharmacy. It does not tell you what happened after that. The standards already know the difference. FHIR separates order, dispense, administration, and a report of use, and it lets a patient’s account remain a report, with a source and a level of detail. The standard is not missing. What goes missing is the will to keep those distinctions attached to one another once the visit is over, and to keep a model from turning a report into a fact.

Capture decides what the record can do later. A drug name and a date can sit in perfect fields while the reason she stopped, cost, remains buried in the narrative like a note someone meant to come back to. If the workflow reads only the fields, it will keep treating the prescription as current even after the condition required to carry it out has failed. The relevant information exists in the chart, but somebody still has to find it, and first they have to trust that the sentence they found was hers. That is the first path done to the coding layer, then handed, quietly, to the second.

The relationships need a place of their own. A reported interruption should stay connected to its reason, to the plan it affects, and to whatever is being done about it. Then the next action has a basis. Difficulty paying can lead to a real look at assistance. A later report that the medication was obtained can change the picture. Follow-up can establish how she is using it and whether more help is required. Each new fact should alter what the team is entitled to believe. Without those links you can have elegant coding and abundant narrative in the same record and still not be looking at the same person. Drift loves an unlinked sentence. It can be reused, summarized, and billed without anyone noticing it has changed sides.

Following that history also keeps us from asking too much of adherence as a measure. A patient can follow the plan and still get worse, still hit an adverse effect, still need the treatment reconsidered. You need her response and evidence of how the plan was actually carried out. Miss that and a bad outcome gets blamed on a treatment she never received as intended, or a dead plan keeps running because every official task shows complete. Ambient drafts have already been caught inventing commentary on adherence. Analytics will bless the confusion either way. Completed tasks make clean numerators. Interrupted months do not.

Assistance deserves the same scrutiny. A transportation request submitted is a beginning. Whether the ride happened, and whether she reached the appointment, tells you what the request accomplished. Keep those events connected and the record can show the work that remains. It can also keep the clinic honest about the care it thinks it delivered. An encounter becomes computable when later facts can return to the same structure that held the original plan, and when a later summary is not allowed to promote a guess. Otherwise you have a diary with billing codes.

A record like that can still answer a question after the claim has gone out. How is the agreed plan working in her life? The information can be retrieved, checked, updated. A detail that surfaced in conversation can decide the next move weeks later. Keep the richness of what was said. Make the information usable at scale. Those are not two projects. They are one design problem that the industry keeps trying to split so each half can declare victory.

The relationships have to survive a change of system. Interoperability is the ability of systems to exchange information and make use of what they receive, as ONC puts it. Portability is the ability to obtain that information and take it somewhere else, including when a patient changes clinicians or a practice changes software. They overlap. Neither is accomplished by writing the note faster. A narrative that performs beautifully inside one product is still only a document if the next environment cannot use what it meant, or cannot tell a fact from a fluent interpolation.

A physician can get time and attention back from an AI tool while the information remains a hostage of the room that produced it. The note reads well. The assistant can talk about it. In that setting the improvement is real. Move the record and a different question appears. How much of the usefulness travels? If the answer is that another model will sort it out on arrival, the second path has been renamed infrastructure, and the expand-contract cycle has been given another turn.

Her history is enough to test it. A new clinician should be able to tell the medication that was prescribed from the one she reported stopping, with the reason and the timing still attached. A navigator in another organization may need eligibility information and the status of an assistance request. Each person needs enough context to continue without inheriting a fossil, an invention, or the job of making her tell the same story again. Provenance is not a scholarly garnish. It is how a later reader knows which sentence is still true.

An export can move the pages and leave the next system to reconstruct the relationships. A readable document is not nothing. Using it reliably in another workflow often requires a second act of interpretation, which is another expansion, which is another chance to drift. Even the federal certification criterion for electronic health information export asks for a computable format without naming a single standard or transport method. Successful export, in that thin sense, still leaves someone else to determine what can actually be understood. Pages travel. Meaning does not, unless somebody arranged for it to.

Dates and sources have to travel with the facts. A correction has to remain a correction. An unresolved barrier should stay tied to the plan it is blocking. When work is handed to another organization, responsibility for the next step has to be accepted and made visible. Shipping data does not settle that. It only makes the handoff less of a scavenger hunt.

So the test for AI gets harder, which is overdue. Measure the time returned to the physician. Then look at what happens when the information is shared, exported, or used to continue care somewhere else. Look also at whether the signed note still matches the visit after the last transformation. A system that keeps its meaning through those moves contributes to continuity. One that writes a gorgeous note and goes dark at the threshold has solved a local inconvenience. Time saved in the exam room is a start. What survives after the exam room is the job.

That is the part of the encounter I want the technology to stay with. Social determinants of health are the circumstances in which people live and the access they have to resources and opportunity. The World Health Organization puts those conditions at the center of health inequity. In a single visit they tend to arrive as something smaller and sharper. She cannot pay for the treatment. She cannot get there. Ambient AI is unusually good at catching that sentence. Catching it as prose and then losing it as structure is how the second path wastes the only honest thing the first path found. Completing the sentence into a judgment she never invited is how the same path makes the record slightly worse than silence.

Treat the 80/20 line with some care. The older County Health Rankings model put 20 percent of its weighting on clinical care and spread the rest across social and economic factors, health behaviors, and the physical environment. Those were weights for a population score. They do not prove that social conditions cause 80 percent of every good outcome. What they do describe, if you are paying attention, is interaction. Cost does not replace clinical judgment. It can keep clinical judgment from ever reaching the person it was meant for.

In her case the interaction is already on the table. A treatment has been identified. Money has interrupted it. There may be a program she qualifies for, a benefit she has never used, an organization that could walk her through the application. None of that matters until someone determines whether it fits and helps her get it. A computable encounter that cannot represent that work will keep producing beautiful notes about plans that die between the clinic and the pharmacy.

Getting the help is work. A program wants proof of eligibility, a physician statement, a document she already gave to somebody else. The ride covers her zip code and requires booking two days out. The application sits because one page is missing. From the clinic these look like errands. For the person waiting, any one of them can freeze the entire plan.

Navigation lives in that freeze. A navigator finds the requirement that is still open and moves her through it. A social worker sees support the clinical team never considered. A nurse checks whether the treatment was obtained. A pharmacist finds an affordability option that was sitting in plain sight. The physician still owns the clinical judgment. The conversation in the room finally gets a team. If the record cannot hold what that team learns, or if it holds a polished distortion of what she said, the next person is back in the narrative with a flashlight.

This is not theoretical. The CMS Accountable Health Communities Model tested screening, referral, and navigation between clinics and community services. It assumed, correctly, that identifying a resource and reaching it are different jobs. Healthcare AI ought to learn from the people already doing the second job and give them a durable place in the workflow. Screening with no follow-through is another completed task with an unknown ending. The model knew that. The software should not get to discover it as an insight.

A navigator cannot work from a slogan. Reduce her explanation to “nonadherent” and the reason treatment stopped is gone. Cost was the barrier. It was reported on a particular day. She did or did not want help. Something was or was not already tried. Those facts should remain fastened to the original account so the next person does not have to make her begin again. Collapse the story into a label and the first path has produced a field that conceals the only fact worth having. Invent the label from fluent pattern-matching and the concealment arrives pre-signed.

Some of it belongs in fields. A ZIP code narrows the search to organizations that actually serve her. Confirmed eligibility can keep the team from chasing programs that will refuse her on page two. Other things need the space of a sentence. She can only reach an appointment when a relative is free after work. Make the constraint retrievable. Keep enough of the explanation that someone can act without being a fool. The old fight between templates and narrative was never about whether people speak in paragraphs. It was about whether a paragraph can travel without becoming a different paragraph.

Then the information needs a date and a way to change. A ride arranged last month is a rumor about next week. An application can be pending, approved, or dead, and each state creates different work. Keep the history current and the record describes her circumstances as they are, with the earlier version available when someone needs to know how it shifted. Provenance without time is just a name on a sentence. Time without an update is archaeology. Either absence gives the next expansion room to improvise.

Fail to organize it and a request for help becomes another descent into a 228-page chart. Find the passage. Decide whether it is still true. Decide whether it was ever true. Assemble it into a letter. A language model can help with the search. It can also skip a detail or braid together statements that belong to different Tuesdays. That is the second path in a smaller room, another expansion of material that has already been contracted and inflated once. Fluency is not structure. If the facts and their sources were kept as the work unfolded, the next task starts with something you can inspect.

That matters more when the output leaves the building. A letter supporting access should rest on what she reported and what the team established. An uncertain date stays uncertain. Missing information comes back for clarification. The person reading the letter should be able to check the claims without repeating the entire investigation. A model that writes a moving letter from mixed moments has not solved the data problem. It has put a clean shirt on it.

Once the foundation is there, automation can do work worth doing. With her agreement, the system can identify a resource from current information, prepare what the program requires, and send it to the right person. A request that needs a physician statement can return to the physician with the evidence attached. The navigator can see what is still open. The patient can hear the next step in language she understands. That is the first path continued past orders and codes, into the work that decides whether orders and codes ever become care, and past the temptation to treat a draft as a decision.

Keep watching after the referral goes out. Did the organization answer? Did she connect? Did the service move the barrier? A referral marked complete because a message left the system tells you almost nothing about the person it was supposed to help. Record the result and the next handoff has ground to stand on. Completion at the interface is an attractive number. It is also how software applauds itself for hitting send.

Sometimes the result is bleak. No capacity. Wrong eligibility. A service that exists on paper and cannot work in her actual week. Leave the obstacle visible so someone can try another way. If support is simply not there, say so. Honest records show the edge of what a clinic can do with the world around it. A system that cannot represent failure will keep offering the same help and calling the repetition intelligence.

Those limits vanish when the software is built around a single screen. A physician tool can declare success at the completed order. A patient app can issue instructions and leave the carrying-out to someone already underwater. Shared work requires each person to see the part they can do and to know when it is someone else’s turn. Ambient AI made the physician’s screen quieter. That is not the same as making the patient’s course visible, and it is not the same as keeping the physician awake inside the note.

She should be able to use that shared picture herself. After the visit she might ask an assistant what she can do today to move the plan forward. The answer should start from the plan agreed in the room and the circumstances already sitting around it. She should not have to become her own medical records clerk in order to ask an ordinary question. If the assistant has to reconstruct her life from scattered prose, the second path has simply changed addresses, and the reconstruction will carry every prior distortion forward with confidence.

From a coherent record the assistant could explain the next step, look for an earlier slot for an ordered test, or find help for a medication she cannot pay for. With her direction, and the right connections, it could prepare an application or push a scheduling request. When a physician has to review something, the request can arrive with the relevant pieces already attached. She gets a way to act without falling off the people responsible for her care.

That only works if the plan and her circumstances remain coherent wherever the system retrieves them. Active recommendation, not the one already replaced. Open barrier, not the one already handled. Facts with dates and sources, still related to one another. Otherwise a scrap of conversation, an old medication list, and a lone mention of money get blended into advice that belongs to no moment in her actual history. It can sound kind. It can also be wrong in a way a dated field would have made obvious.

Structure should make the context reviewable. It should also make the holes visible. If the current plan is uncertain, the assistant should say so and help her take the question back to the team. Changing treatment is not a chatbot decision. Helping her move includes helping her see when a decision still has to be made. Helping the clinician includes not inviting them to surrender the act of deciding what the visit was.

Her actions have to reenter the record. She books the appointment herself. She learns the program closed. She reports that the medication is finally affordable. Those updates should come back with their source marked, so the team can reconcile them and see what is left. Otherwise her initiative becomes another private version of care that the next person has to stumble over. Her words are still words. They become usable when they can change the same relationships the clinic has been keeping.

Portability starts to mean something here. She can take the relevant history to another clinician or an authorized service she chooses, including the work already done and the help still needed. What began in the first encounter can follow her into a decision she makes herself. A navigator is still there when the process gets ugly. Someone who needs a phone call or a person in a chair should be able to continue from the same plan. Portability that moves only the note moves the least durable part, and sometimes the most distorted one.

It also means taking her aims seriously. She may want to understand an instruction before she acts, compare ways of getting to an appointment, or ask whether the next proposed step fits her life. The system should help her take part in those choices and make her preferences known. Her place in the record grows by what she adds and by what she decides.

That is why the physician and the patient still have to sit at the beginning. The conversation reveals what needs attention and creates enough trust that help can be accepted. Asking the physician to complete every later application would spend the time the technology just returned. Asking the physician to stop thinking because the draft sounds finished would spend something harder to recover. A clean handoff lets the relationship keep its clinical and human value while the wider team does what it knows how to do. Ambient AI earned its place by protecting that relationship. It will deserve a larger one when the information created there can leave the room without losing its shape, and without acquiring a shape nobody in the room gave it.

That is the direction we are building toward at Serelora. The next generation of this software should produce a computable clinical encounter. Keep the richness of the conversation. Make the information usable at scale, including the practical help that decides whether a care plan is possible. Hold the facts still enough, through each contraction and expansion, that a later reader can tell a report from a conclusion. When clinical work and navigation run together, the team needs to know what is underway, who owns the next step, and what still stands between the patient and that step. Orders, codes, and prior authorization belong in that encounter. They are not the encounter.

Then measure something less flattering. A completed note says documentation ended. Whether she obtained the medication, reached the appointment, or received the support she asked for says whether the work did. Whether the signed note still matches the visit says whether the documentation can be trusted as evidence of either. Follow those events and a clinic can begin to see which interventions helped, where access failed in the same way twice, and what its patients still do not have. The industry already knows how to count minutes returned to the physician. It is less eager to count whether the plan survived contact with her month, or whether the record survived contact with the model.

Outcomes and equity are the point. Operational measures tell you whether the machinery is moving. To know whether it is serving its purpose you have to follow what happens after the tasks show complete. A shorter note can give a physician room to listen. Whether that room produced a workable plan, timely care, or better health is a different kind of evidence. An eloquent chart note is not proof that care occurred. It is proof that someone was heard long enough to generate one, or that a system was fluent enough to sound as if they were.

Equity asks who received the improvement. Average wait times can fall while the same people still cannot find a ride, an interpreter, or a way to pay. A digital path can be smooth for patients who can work it alone and leave everyone else waiting on help that never comes. Speech systems already err unevenly across voices. If those errors enter the note and then the codes, inequity acquires a documentation trail that looks objective. See the difference. Follow the open barriers. Ask whether the help improved access and outcomes for the people who needed it. If ambient AI makes the clinic faster only for patients who were already going to follow through, the workflow problem was solved for the wrong room.

That is the standard I want applied to Serelora. The patient should be better able to understand and pursue care. The team should be better able to help her obtain it. The record should let us see what followed, and it should still be hers. Efficiency matters when it changes her life. Otherwise it is a nicer afternoon for the office.

Research should feel the difference. An apparent failure of treatment may be a failure of access, or of use, or of a plan that was never possible, or of a note that drifted until the study cohort no longer described the visits that produced it. Keep the circumstances, and keep the transformations from inventing new ones, and those questions become investigable. Causation still needs a real study. The same discipline that helps one patient can keep a population analysis from discovering nonadherence where it should have found a price.

For her, it is simpler. She told a physician something difficult because, for a moment, she thought explaining it might matter. The physician listened and understood how it touched her care. Everything built after that ought to give the conversation a decent chance of changing the next week, and no extra chance of rewriting what she said.

When she comes back, I want the chart to show what the team did with what she handed them. I want the person who opens it to know whether help reached her and what is still outstanding. I want that person to be able to tell her sentence from the system’s. She already did the hard part. She explained why the plan was difficult. The next conversation should have the benefit of what we did after we heard her.