Built for Three: AI-Native Software That Serves Patient, Provider, and Business
By Mona Lisa Healthcare
Three groups of people have a stake in every clinical encounter, and healthcare software has spent thirty years serving them one at a time. The patient wants care that is safe, coordinated, and theirs to carry. The provider wants to practice medicine, not feed a database. The business — the practice, the facility, the operator — needs the encounter to convert into a clean claim that is actually paid. These are not competing interests. They are three views of the same event. Yet the tools built to support that event treat them as a zero-sum contest, and almost every product on the market is an answer to the question of who should lose.
Documentation makes the trade visible. A note thorough enough to satisfy a payer is a note that costs the clinician an hour of their evening. A workflow fast enough to protect the clinician's time is a workflow that ships incomplete claims. A patient portal rich enough to matter is one more surface nobody has time to maintain. Optimize any one corner with conventional software and the other two pay for it.
This paper makes that argument in three moves. First, it names the trade-off and shows, with data, how expensive it is. Second, it explains what "AI-native" means precisely enough to separate it from the chatbot bolted onto a legacy system. Third — and this is the part that matters to a buyer — it shows how the same design choices that relieve the provider are the ones that clean the claim and complete the patient's record, and it closes with the standard of proof you should hold every AI vendor to, including us.
The trade-off triangle: how health software learned to pick winners
Start with the provider, because the evidence is unambiguous. In a time-and-motion study across four specialties, physicians spent 27% of their day on direct clinical face time with patients and 49% on the electronic health record and desk work — nearly two hours of administrative work for every hour of care — plus one to two more hours of documentation at home each night (Annals of Internal Medicine, 2016). A separate analysis of three years of EHR event logs put family physicians at 5.9 hours of EHR time in an 11.4-hour day, including 86 minutes of after-hours "pajama time" (Annals of Family Medicine, 2017). The record-keeping did not augment the work. It became the work.
Now the business. Denials are not a rounding error; they are a structural leak. Insurers on the ACA marketplace denied 19% of in-network claims in 2024, with rates ranging from 3% to 36% across carriers — and consumers appealed fewer than 1% of the roughly 85 million denials, of which two-thirds were upheld (KFF, 2024). A denied claim is rarely a coding typo caught at submission. It is most often an assertion about a clinical fact — the place of service, the medical necessity, the missing element — that was fixed hours earlier at the bedside and cannot be scrubbed back into existence at the clearinghouse.
And the patient sits downstream of both. Their record is scattered across every organization that ever treated them; the clinician deciding their care at 2 a.m. sees a fraction of it; and the note written under time pressure to clear a queue is the same note that will represent them to the next provider. The patient inherits the compromises the other two corners were forced to make.
| Stakeholder | What they actually need | How legacy software shortchanges them | What AI-native aligns |
|---|---|---|---|
| Patient | A complete, portable, safe record and coordinated care | Record fragmented across providers; decisions made on partial data | Nationwide record aggregation, consented sharing, safer after-hours care |
| Provider | To practice medicine, not feed software | ~2 hrs of EHR/desk work per hour of care; nightly "pajama time" | Draft-assist from entered data; review that respects clinical judgment |
| Business | Clean claims that are paid on the first pass | ~1 in 5 claims denied; most never reworked; leaks recur | Denial patterns become rules checked at the note, before submission |
The right-hand column is the whole argument in miniature: the same intervention appears in all three rows, because in an AI-native system it is the same intervention.
What "AI-native" actually means — and why bolted-on AI doesn't qualify
"AI" has become a sticker applied to software that has not changed. A legacy EHR with a chatbot in the corner is a legacy EHR. The distinction that matters to a buyer is not whether a product "has AI," but where the intelligence sits, what it learns from, and who stays in control. Three properties separate AI-native from AI-adjacent.
It sits in the workflow, at the point of decision
Bolted-on AI lives beside the work — a separate screen, a summary you request, a copilot you consult. AI-native intelligence lives inside the work, where the determinative facts are being set: as the note is written, as the code is chosen, as the record is assembled. Placement decides the ceiling. Intelligence applied at the end of a pipeline can only describe what already went wrong; intelligence applied at the beginning can change the outcome.
It learns from your own data, not a generic model's priors
A generic rule library is a checklist. What makes a system a feedback loop is that it learns from the organization's own adjudicated results — the denials your payers send you, every week, naming exactly what your documentation should have said. When a payer refuses a claim for a place-of-service mismatch, that refusal is the highest-quality training signal in the revenue cycle: adjudicated ground truth, payer-specific and facility-specific. Most organizations route it to a rework queue and discard the lesson. An AI-native system routes it back to the point of care as a rule, checked on the next visit. Your data teaches your software.
It keeps the human in control and is honest about its limits
An AI-native system in medicine earns trust by refusing to overstep. It evaluates rather than fabricates; it cites its evidence; it leaves the clinical record the clinician's to sign; and it measures what it can prove rather than asserting what it cannot. AI that quietly makes clinical decisions, or that markets outcomes it has never measured, is not more advanced — it is less trustworthy.
One artifact, three beneficiaries
Here is the mechanism that dissolves the triangle, stated as plainly as we can. The clinical note is not three documents. It is one. The version that is easiest for the clinician to produce, the version that satisfies the payer, and the version that best serves the patient are the same well-formed record — and conventional software treats them as three because it improves each in isolation. AI-native software improves the artifact once, and all three stakeholders inherit the result.
This is why the trade-off is an artifact and not a law. The clinician's evening and the organization's denial rate are not in tension when the intervention that fixes one is the intervention that fixes the other. A note that is missing the chief complaint is harder to defend clinically, more likely to be denied, and less useful to the next provider — three failures with a single root. Repair the root at the point of documentation and the three benefits arrive together, from one action, at no cost to each other.
The provider: giving clinicians their attention back
The documentation burden is the clearest place where Mona Lisa refuses the trade-off. Lisa AI, the platform's intelligence layer, is built on an inversion of the usual approach: it does not write the clinician's clinical judgment for them — it relieves the mechanical weight of recording it. The model evaluates what the clinician wrote against the organization's rules, cites its evidence, and proposes ready-to-insert corrections. Nothing enters the record that a human did not write or explicitly accept. A false flag costs a clinician two seconds to dismiss; a fabricated clinical fact would cost far more, so the system is designed to make the first kind of error and never the second.
The after-hours case: the blank page, removed
Consider the hardest documentation moment in skilled nursing: nights and weekends, when a resident's condition changes and no physician is in the building. A floor attendant brings a tablet to the bedside and rings a provider by telehealth. The provider assesses the resident, makes a decision, enters the billing codes — and moves to the next virtual room, with no time to write. Historically, that encounter reaches billing as a bare code with no story behind it.
Lisa AI closes that gap without crossing the line into practicing medicine. When the chart goes idle, it drafts a provisional note from the data the clinician already entered — the codes and the structured record — so that end-of-shift completion becomes a task of supplementing and attesting rather than facing a blank page. The draft is never the signed note; the clinician must complete and attest it; and the same evaluative review then checks it before sign-off. Generation is bounded, subordinate to human attestation, and always followed by evaluation. The clinician gets their evening back; the record is complete; the encounter reaches billing as a real note. Same intervention, three beneficiaries.
Mandatory review, advisory findings — why we do not block
Mona Lisa built hard-stop enforcement that prevented a clinician from completing a visit until findings were resolved, ran it, and removed it six days later. The shipped design makes the review mandatory and the fix advisory: every note is checked, no finding blocks the clinician, and every decision — applied, addressed, dismissed, ignored — is recorded against the visit. The reasoning is a statement of respect for the profession: a rule is general and the patient is specific; a system that blocks in the inevitable exception forces false compliance and corrupts the very signal it was meant to protect. Authority should sit where accountability sits — with the clinician who signs the note. Software that respects that boundary is software clinicians will actually use, which is the precondition for every other benefit in this paper.
The business: cleaner claims, fewer denials
The same note that spares the clinician is the one that protects the revenue. Mona Lisa's approach to denials is not a better scrubber; it is a shift in where the intelligence sits. Because the platform runs the billing pipeline and receives the remittances, it can do what a downstream tool cannot: turn each payer's adjudicated denials into rules that are checked at the point of documentation, on the next visit, for that facility and that payer. The remittance teaches the note.
This is not a theory. Across a contracted network of seventeen skilled nursing facilities, an analysis of the platform's own adjudicated claims (November 2025–May 2026) found that nearly one in three adjudicated claims was denied — and that the single largest correctable cause was a two-digit place-of-service code, selected in a dropdown by a clinician with no way of knowing how a given payer would treat it.
| Place of service | Adjudicated | Pay rate | What it means |
|---|---|---|---|
| POS 02 — Telehealth (non-home) | 104 | 92% | The highest-paying setting — worth documenting well |
| POS 31 — Skilled Nursing Facility | 287 | 79% | Pays well on clinically similar encounters |
| POS 32 — Nursing Facility | 243 | 50% | Half denied — often the same care, wrong code |
| POS 13 — Assisted Living Facility | 17 | 0% | Paid zero times — a code that simply does not reimburse here |
Revenue that shows up, not just losses avoided
Denial prevention is loss avoidance, measured against a counterfactual — genuinely valuable and genuinely hard to prove. So Mona Lisa built the other half too. Where most compliance tooling hunts for overcoding because it creates audit exposure, Lisa AI also flags undercoding: a note whose language describes a materially more complex encounter — the patient was sent to the emergency department, services were escalated — but is billed at a routine level. That is work the clinician performed and documented and then billed for less of than they did. Correcting it is revenue that shows up in the remittance, attributable and immediate, requiring no one to believe a counterfactual. For an organization skeptical of AI claims generally, it is the part of the case that arrives with its own evidence.
The patient: a complete record and safer care
The third corner is the one the industry most often forgets, because the patient does not sign the purchase order. Mona Lisa's answer is Health Compass, a patient-controlled record hub built on the emerging national interoperability standard. Under the Trusted Exchange Framework (TEFCA), an identity-proofed individual can retrieve their records nationwide; Health Compass operationalizes exactly that — identity verified to a federal assurance standard, records pulled across the country and normalized into one longitudinal record the patient controls.
This is not the same as the organization-to-organization data exchange other systems provide. It puts the individual in control of assembling their complete record and choosing, field by field, what to share with family, caregivers, and other providers — with consent that is granular, revocable, and logged. And it feeds the clinical workflow directly: at admission, a fuller medication and problem list; at the point of an after-hours decision, a chart the covering clinician can actually rely on.
Because Sherpa Care is a full clinical platform and not only a system of record, it can also carry data classes a documentation system cannot — a pharmacogenomic result that makes every future prescription safer for a polypharmacy resident, a stream of wearable and wellness data feeding the intelligence layer — each held in the patient's longitudinal record and governed by the same consent architecture. The through-line is constant: the patient's record is theirs, it is complete, and the completeness makes the care safer and the claim cleaner at the same time.
How the Mona Lisa team builds AI-native software
The capabilities in this paper are the product of a set of engineering convictions, several of them learned by getting the first version wrong. They are worth stating plainly, because they are what distinguish a team that builds AI-native software responsibly from one that ships a model and hopes.
| Design principle | What it means, and why it matters to you |
|---|---|
| Evaluate, do not fabricate | The model checks the clinician's work and cites evidence; it does not invent clinical facts. A false flag costs two seconds. Generation, where used, is bounded to rendering data the clinician entered, and is always completed and attested by a human. |
| Learn from your own data | Rules derive from your adjudicated denials, not a generic library. Your payers and your facilities shape what the system checks — which is what turns a checklist into a feedback loop. |
| Rules are data, owned by you | A compliance lead changes what the AI checks by editing a spreadsheet — no engineer, no release. The loop from a new denial to a new safeguard is measured in days, not release cycles. Your compliance policy stays yours. |
| Wrap the model in code that does not trust it | Deterministic logic handles redaction, context assembly, and filtering on both sides of the model. Severity is withheld from the model so it cannot bias toward flagging. The model is a component, not the system. |
| Authority where accountability sits | The clinician signs the note and carries the liability, so the clinician decides. Mandatory review, advisory findings. Software that respects that is software clinicians adopt. |
| Standards-native, not standards-adjacent | HL7 and FHIR interoperability and TEFCA-based patient access are built in, not sold as add-ons — the foundation for a record that follows the patient and a platform that complements the systems you already run. |
| Measure what you can prove | Every AI action is metered and every review is reproducible. We report mechanisms we can substantiate and decline to publish outcomes we have not measured. Honesty is an engineering discipline here, not a marketing posture. |
Read together, these are not seven features. They are the profile of a team that treats a medical record as something to be protected, not automated away — and that is willing to remove its own hard-stop enforcement six days after shipping it because the data said the softer design was the better one. That willingness to be corrected by evidence is the capability behind all the others.
The honesty standard: what to demand from any AI vendor
Documentation AI is routinely sold on denial-rate reductions and hours-saved figures produced by comparing a period before deployment to a period after — in an environment where staffing, payer mix, contracts, and patient acuity all changed at once. Those are not measurements. They are before-and-after anecdotes with a percentage sign, and the category is being poorly served by them.
Mona Lisa's position is that a vendor should tell you what it can prove and what it cannot, before you sign. We can substantiate that every AI call is metered, that every review is reproducible against the exact rules and evidence it used, and that the rules derive from real adjudicated denial data. We do not claim a denial-rate reduction attributable to the AI, because the record-level join between a finding a clinician accepted and a claim paid eleven weeks later is genuinely hard to build — and we would rather say so than publish a number we cannot defend.
Five questions to ask every AI vendor (including us)
- Where does your intelligence sit — at the point of documentation, or at the end of the pipeline?
- Does it learn from my adjudicated data, or is it a generic library I cannot change?
- Can my compliance team change what it checks without an engineer and a release?
- Does it ever write clinical content, and if so, who completes and attests it?
- Show me the record-level join between your AI's findings and my claim outcomes — not a before-and-after.
The difference worth weighing is whether a vendor can answer the fifth question honestly, and whether they volunteer the answer before the contract or after. A partner willing to hold itself to that standard is a partner whose other claims you can trust.
The platform at a glance
The three products are one platform, designed so that each stakeholder's benefit reinforces the others rather than competing with them.
| Product | What it is | Who it serves first |
|---|---|---|
| Sherpa Care | The AI-native clinical, operational, and revenue-cycle platform — documentation, coding, billing, and reporting in one system. | Provider and business: less documentation weight, cleaner claims, real-time visibility. |
| Health Compass | The patient-controlled record hub: nationwide record aggregation under TEFCA, unified with genetics and wellness data, with granular consent and sharing. | Patient: a complete, portable record they control — and safer, better-informed care. |
| Lisa AI | The intelligence layer across both: documentation review, denial-pattern rules learned from your own data, undercoding capture, and after-hours draft-assist. | All three: the mechanism that makes one artifact serve patient, provider, and business at once. |
Crucially, the platform is built to complement the systems an organization already runs rather than to demand a rip-and-replace. Standards-native interoperability means Sherpa Care and Health Compass can sit alongside an existing EHR, filling the seams — the incomplete record, the leaking claim, the after-hours gap — that the incumbent system leaves open.
Conclusion
The trade-off among patient, provider, and business has been treated as the natural condition of healthcare software for so long that most buyers have stopped noticing they are being asked to choose. They accept the clinician's lost evenings as the cost of a defensible note, the denial rate as the cost of moving fast, the fragmented record as the cost of doing business. None of it is natural. All of it is a consequence of where the intelligence was placed.
Move the intelligence into the workflow, teach it from the organization's own data, and keep the clinician in control, and the three interests stop competing. The note that takes the clinician less effort is the clean claim is the complete patient record. That is not a slogan; it is an architectural fact, and it is the one Mona Lisa Healthcare was built on. Built for three — because the encounter always was.
We would rather show you the mechanism and the standard of proof than hand you a percentage we made up. Ask us the fifth question.
References
- Sinsky C, et al. "Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties." Annals of Internal Medicine, 2016. https://www.acpjournals.org/doi/10.7326/M16-0961
- Arndt BG, et al. "Tethered to the EHR: Primary Care Physician Workload Assessment." Annals of Family Medicine, 2017. https://www.annfammed.org/content/15/5/419
- KFF. "Claims Denials and Appeals in ACA Marketplace Plans in 2024." 2024. https://www.kff.org/patient-consumer-protections/claims-denials-and-appeals-in-aca-marketplace-plans-in-2024/
- Mona Lisa Healthcare. "Closing the Loop Between the Remittance and the Note." Lisa AI documentation-review white paper; adjudicated claims, Nov 2025–May 2026, 17-facility skilled nursing network.
- CMS. Nursing Home Five-Star Quality Rating System — Technical Users' Guide (Quality Measures domain). https://www.cms.gov/medicare/provider-enrollment-and-certification/certificationandcomplianc/downloads/usersguide.pdf
About Mona Lisa Healthcare. Mona Lisa Healthcare builds Sherpa Care, a standards-native platform unifying clinical delivery, operations, and the revenue cycle for decentralized, multi-service-line care; Health Compass, a patient-controlled record hub aligned with nationwide interoperability; and Lisa AI, the intelligence layer across both. Claims figures cited from the platform's own adjudicated billing data are presented as the diagnostic baseline that shaped the product, not as outcomes attributable to the AI. External statistics are cited to their published sources above.