Six Vendors, One Patient: Why Point Solutions Are Failing Infusion Centers

A referral arrives by fax. One page, or four. On it: a patient's demographics, ten required documents, two diagnosis codes, a lab cadence, a pre-medication protocol, a four-step titration ladder, and a line-care order. It is a single clinical instruction, written by one physician, about one person.

It will now be typed into six different systems by four different people, none of whom can see what the others did.

A representative anti-amyloid infusion referral, reconstructed from the standard field set published in manufacturer referral forms. Every colored tag is a different vendor — a different login, a different queue, a different place for the patient to stop moving.

The Stack Nobody Designed

No infusion center set out to build this. It accreted. A referral tool was bought because faxes were getting lost. An intake product was added because registration took too long. An eligibility engine came in because benefits checks were eating a full-time position. A prior-authorization vendor followed, then a chair-scheduling optimizer, then an ePRO platform, then a denials-management layer bolted onto the revenue cycle.

Each purchase was rational. Each vendor was, quite genuinely, better at its slice than whatever it replaced. And the sum of all those good decisions is a workflow in which no single system — and frequently no single person — can answer the question where is this patient right now, and what is she waiting on?

A 2024 analysis of the community oncology technology landscape in The American Journal of Managed Care made the same observation: practices spend heavily on technology and end up with a patchwork of disconnected, siloed systems that require human effort to tie together before any of the value is realized. That last part is the whole problem. The integration layer is a person. Usually several — usually the most experienced nurse or navigator in the building, holding the patient's state in their head and their inbox.

Read the red lines in that grid together and a pattern emerges. Not one of those failure modes lives inside a product. Every one of them lives in the space between products. Which means no amount of improving any individual tool will fix them — and the vendors, quite reasonably, do not consider the gaps their problem.

The Integration Tax Is Paid in Clinician Hours

The cost of fragmentation does not show up on an invoice. It shows up in the schedule of the people you can least afford to waste.

The foundational time-and-motion study here remains Christine Sinsky's 2016 work in the Annals of Internal Medicine, which observed 57 physicians across 430 hours in ambulatory practice. The finding that stuck: physicians spent 27.0% of their time on direct clinical face time with patients, and 49.2% on EHR and desk work — roughly two hours of clerical work for every hour of patient care, plus one to two more hours at home each night. That study predates most of the point solutions now layered on top of the EHR. The stack has only grown since.

Then there is the login tax. Research from clinical-access vendors and published health-system case studies put clinician time lost to authentication alone at more than 45 minutes per shift, with emergency clinicians at one Canadian health system logging in up to 60 times in a single shift. One 19-facility system published annualized savings of 10,000–13,000 hours — worth over a million dollars — purely from consolidating sign-on. That is the price of the seams, measured in a currency infusion centers are already short of.

And the seams do not merely slow things down. They lose people. An analysis of referral behavior in one large health system found that of 103,737 attempts to schedule a referral, only 34.8% resulted in a documented completed appointment. More unsettling still: an ONC-funded study found patient-record match rates as low as 50% between organizations using the same EHR vendor — not different vendors, the same one, defeated by configuration drift and process variation. If two instances of one product cannot reliably agree that two records describe the same human being, the premise that six unrelated products will do so on your behalf is not an engineering assumption. It is a hope.

Full citations in the Sources section at the end of this article.

Delay Is Not an Inconvenience. It Is a Clinical Variable.

Operational friction is usually discussed as a business problem — margin, throughput, staff satisfaction. In infusion, that framing understates it badly, because the thing the friction delays is a drug, and the delay itself has a dose-response curve.

The cleanest evidence comes from a cohort study in Arthritis Care & Research examining infusible medications specifically. When prior authorization was required, the median time from prescription to infusion was 31 days, versus 27 days without. For the roughly one in five patients whose request was initially denied, the median stretched to 50 days. Here is the part worth sitting with: 82% of those denials were ultimately overturned, and the overall approval rate was 96%. Nearly all of that waiting produced no change in what the payer covered. It produced only delay.

What does delay cost? An analysis of 2,241,706 patients in the National Cancer Database, published in JAMA Network Open, modeled five-year mortality against time-to-treatment-initiation. For stage III colon cancer, predicted five-year mortality rose from 38.9% when treatment began within 61–120 days to 47.8% at 181–365 days. For stage I breast cancer, from 9.7% to 15.2% across comparable bands. Administrative delay is not adjacent to the clinical picture. It is part of the clinical picture.

The burden is well characterized on the practice side too. The American Medical Association's 2025 physician survey (n=1,000) found physicians completing 40 prior authorizations per week, consuming 13 hours; 95% reported that PA delays care; 79% reported patients abandoning treatment because of it; and 40% now employ staff who do nothing but prior authorization. A separate study of 178 cancer patients in JAMA Network Open found 69% experienced a PA-related delay, 73% of those waited two weeks or more, and 22% never received the recommended treatment at all.

These numbers are usually deployed as an argument about payers, and that argument is worth having. But there is a second reading that matters more to anyone running a center: a large share of that 13 hours is spent assembling information the organization already has — scattered across the intake system, the EHR, the imaging archive, and a shared drive. The payer asked for a cognitive assessment score that exists. Someone has to go find it.

The Anti-Amyloid Era Is About to Break the Seams

Everything above is a description of a system under strain. What follows is the load that is arriving now.

Lecanemab and donanemab — Leqembi and Kisunla — are not simply new infusions. They carry a dependency structure that infusion centers have not previously had to enforce. Both require serial MRI monitoring for amyloid-related imaging abnormalities (ARIA), and critically, those scans are gates, not observations. The scan must be performed, read, and cleared before the next infusion may proceed.

Kisunla's label requires MRI prior to the 2nd, 3rd, 4th and 7th infusions, alongside a titration ladder that changes the dose at each of the first three visits. Leqembi's requirements were tightened in August 2025, when the FDA issued a drug safety communication adding an earlier MRI before the 3rd infusion — a change prompted by six deaths from ARIA and 24 serious ARIA-E cases occurring before the 4th infusion, including four deaths after only the third. The monitoring schedule now spans the 3rd, 5th, 7th and 14th infusions.

Layer on the rest: amyloid pathology confirmation by PET, CSF, or one of the newly cleared blood-based assays; APOE genotyping, where homozygotes carry markedly elevated ARIA risk; CMS coverage available only through a registry pathway, with ALZ-NET requiring baseline submission plus follow-up every six months for two years and then annually; an amyloid PET at roughly 12–18 months to decide whether the patient has cleared and should stop.

This is the point at which the argument stops being about efficiency. A missed MRI in a chemotherapy workflow is a scheduling failure. A missed MRI in an anti-amyloid workflow is an infusion delivered without the safety check the label requires — in a drug class where the FDA has already documented fatalities from the exact complication that scan exists to detect.

The structural picture makes this harder rather than easier. Reporting on lecanemab uptake indicates roughly 60% of patients are infused at sites unrelated to the prescribing neurologist's practice — so the referral loop is split across organizations, not merely across systems. Meanwhile a peer-reviewed Markov model of US infusion capacity through 2032 concluded that projected capacity is insufficient under every regimen scenario tested, and RAND modeling estimated that roughly 2.1 million patients with mild cognitive impairment could progress to dementia while waiting for diagnostic and treatment capacity to catch up.

A center cannot solve national capacity. It can decide whether the capacity it has is spent moving patients forward or reconciling systems.

Why “Best of Breed” Keeps Losing the Argument

It is worth being fair to the point-solution model, because it won for real reasons. Specialized vendors ship faster, understand their niche more deeply, and are usually better than the corresponding module in a monolithic suite. On any single dimension, best-of-breed genuinely is best.

The trouble is that infusion care is not a set of single dimensions. It is a dependency chain, and a chain is evaluated end to end.

The market has begun to reflect this. Bain's 2023 provider IT research found that more than 70% of providers now say they will look to their EHR vendor for new capabilities before meeting with point-solution vendors — a striking inversion of the prior decade's logic, driven by exactly the integration and cost pain described above. A CHIME member survey found 76% of CIOs consider application rationalization critical, while only about one in five has a formal program to do it. KLAS's 2025 operations summit reported consensus around “enterprise-first” standardization, with point solutions admitted only where they demonstrate genuine innovation and fast, provable ROI.

Note what that consensus is not saying. It is not that specialized tools are bad. It is that the burden of proof has moved. A new point solution now has to justify not only its own value but the cost of the seam it creates.

Where AI Actually Helps — and Where It Is Being Oversold

It would be convenient to end with “and then AI solves it.” The honest version is more specific, and the specificity is what makes it credible.

Start with what is oversold. The most-cited prior-authorization AI statistics circulating in this market do not survive verification — a widely repeated set of figures attributed to a major payer-vendor partnership appears in no publication by either party. Claims of “99.8% extraction accuracy” from document-AI vendors are self-reported and unaudited. And the field has a genuine cautionary tale: the widely deployed Epic sepsis prediction model, when externally validated in JAMA Internal Medicine in 2021, showed poor discrimination and calibration and missed roughly two-thirds of sepsis cases — after years of production use across hundreds of hospitals.

Now the part that holds up. The strongest evidence for technology-mediated attention in this population comes from Ethan Basch's randomized trial in JAMA in 2017, where patients on chemotherapy for metastatic solid tumors who self-reported symptoms through a web interface showed a median overall survival benefit of roughly five months, with nurses acting on automated alerts 77% of the time. Intellectual honesty requires the follow-up: the larger PRO-TECT cluster-randomized trial across 52 practices did not replicate the survival benefit, though it did show meaningful reductions in emergency department visits and hospital admissions (1.48 vs. 1.81 per patient-year). Anyone citing the first result without the second is selling something.

So what is the defensible claim? Not that AI diagnoses better than clinicians. It is narrower and more useful: the failures described in this essay are overwhelmingly failures of information movement, not failures of clinical judgment — and information movement is precisely what this technology is good at.

• Reading the fax. A referral arrives as an image. Extracting demographics, diagnosis codes, required documents, and the titration schedule into structured fields removes the first and most error-prone re-keying step in the entire chain.
• Assembling the authorization packet. Much of the 13 hours a week physicians spend on PA is spent locating documents the organization already holds. A system that can see across intake, EHR, imaging, and labs can assemble that packet rather than asking a human to hunt for it.
• Enforcing the dependency graph. This is the one that matters most in the anti-amyloid era. If the protocol says the 3rd infusion requires a cleared MRI, that should be a machine-enforced precondition on the appointment — not a note in a chart that a nurse has to remember to check.
• Watching the patient between infusions. The Basch and PRO-TECT results converge on this much: structured, proactive symptom capture between visits reduces acute utilization. That is a real, replicated finding.
• Predicting the denial before it happens. Optum's denials analysis found roughly 84% of denials were potentially avoidable. Most trace to something missing or inconsistent upstream — which is to say, visible at the time, to a system that can see the whole record.

There is also a strategic argument that infusion leaders should not miss. Payers are already deploying AI at scale in utilization review — a development documented in Health Affairs in 2025 and one the AMA has directly linked to rising denial volume. Providers operating a fragmented stack are bringing manual processes to an automated fight. Whatever one thinks of AI in healthcare generally, the asymmetry in this particular negotiation is already real.

Four Principles for an Integrated Infusion Platform

If a center were to rebuild this deliberately rather than by accretion, what would it insist on? We build in this space, so we hold these views with an obvious interest — but they are stated as testable claims, and they are the criteria we would apply to any vendor, including ourselves.

None of these principles requires believing that integrated platforms are inherently superior to specialized tools. They require only accepting that the seams are where infusion care actually fails, and that whoever is accountable for the whole journey has to be able to see the whole journey.

The Patient Is the Only Thing That Has to Cross Every Boundary

Return to the fax at the top of this essay. Six destinations, ten documents, a titration ladder, and four MRI gates spread across a year and a half. Every system that referral touches will do its job well. The intake platform will capture the demographics accurately. The authorization vendor will work the payer portal efficiently. The scheduler will fill the chair. The billing system will submit a clean claim.

And a patient with early Alzheimer's disease — for whom the window of eligibility is measured in months of cognitive decline — may still wait an extra six weeks because a cognitive assessment score that existed the whole time lived in a system the person assembling the authorization could not see.

That is not a software defect. Every product performed to spec. It is an architecture defect, and architecture is the one thing no individual vendor in a fragmented stack is accountable for.

Infusion centers have spent a decade buying the best tool for each step. The next decade will belong to the ones who ask a harder question: not which tool is best for this step, but what does this patient need next, and does anything in this building actually know?

Sources & Further Reading

• Golla V, Zuckoff A, MacPhail L, Tam J, Rariy C. “Tangle of Tech: How Community Oncology Practices Can Successfully Navigate the Health Technology Landscape.” The American Journal of Managed Care, October 2024 — source of the “patchwork of disconnected and siloed solutions” framing.
• Sinsky C, Colligan L, Li L, et al. “Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties.” Annals of Internal Medicine, 2016;165(11):753–760 — 27.0% direct clinical face time vs. 49.2% EHR and desk work.
• “Treatment Delays Associated With Prior Authorization for Infusible Medications: A Cohort Study.” Arthritis Care & Research, 2020;72(11) — median 31 vs. 27 days; 50 days when initially denied; 82% of denials overturned; 96% overall approval rate.
• Cone EB, Marchese M, Paciotti M, et al. “Assessment of Time-to-Treatment Initiation and Survival in a Cohort of Patients With Common Cancers.” JAMA Network Open, 2020;3(12):e2030072 — n=2,241,706; stage III colon five-year mortality 38.9% → 47.8% by time-to-treatment band.
• Chino F, et al. “Patient Experiences With Prior Authorization in Cancer Care.” JAMA Network Open, October 2023 — n=178; 69% reported delay, 22% never received the recommended treatment.
• American Medical Association. 2025 AMA Prior Authorization Physician Survey (n=1,000) — 40 PAs/physician/week, 13 hours, 95% report care delays, 79% report treatment abandonment, 40% employ dedicated PA staff.
• “Closing the Referral Loop” referral-completion analysis, 2018 — of 103,737 referral scheduling attempts, 36,072 (34.8%) resulted in a documented completed appointment.
• ONC-funded patient-matching research and GAO-2019 findings on patient identification — match rates as low as 50% between organizations sharing the same EHR vendor.
• CAQH. 2025 CAQH Index (2024 data) — $258B in avoided administrative costs; $21B in remaining automation opportunity; prior-year Index reports an average 70 minutes saved per patient visit under fully automated workflows.
• U.S. Food and Drug Administration. Drug Safety Communication: FDA to recommend additional, earlier MRI monitoring for patients with Alzheimer's disease taking Leqembi, August 28, 2025 — adds MRI before the 3rd infusion following six ARIA deaths and 24 serious ARIA-E cases occurring before the 4th infusion.
• KISUNLA (donanemab-azbt) FDA prescribing information, label updated July 2025 — modified titration (350/700/1050/1400 mg) and MRI prior to the 2nd, 3rd, 4th and 7th infusions.
• Alzheimer's Network for Treatment and Diagnostics (ALZ-NET) — data-element and submission schedule documentation: baseline registration, follow-up every 6 months for 2 years, then annually, with imaging submitted per scan.
• “Is US infusion capacity sufficient to meet demand for anti-amyloid therapies?” Markov capacity model, 2023–2032 — concludes projected capacity is insufficient under all tested regimen scenarios.
• RAND Corporation dementia-capacity modeling — projected wait times across diagnosis and treatment stages; estimate that ~2.1 million patients with MCI could progress to dementia while waiting.
• Bain & Company. 2023 Healthcare Provider IT Report: Doubling Down on Innovation — more than 70% of providers will look to their EHR vendor before meeting point-solution vendors.
• KLAS Research. Healthcare Operations Summit 2025 — consensus toward “enterprise-first” platform standardization; CHIME member survey on application rationalization (76% call it critical, ~1 in 5 has a formal program).
• Basch E, Deal AM, Dueck AC, et al. “Overall Survival Results of a Trial Assessing Patient-Reported Outcomes for Symptom Monitoring During Routine Cancer Treatment.” JAMA, 2017;318(2):197–198 — the positive ePRO survival result.
• Basch E, et al. PRO-TECT cluster-randomized trial (52 practices, 1,191 patients) — survival benefit did not replicate (HR 0.99); ED/hospital admissions 1.48 vs. 1.81 per patient-year (p=.006). Cite alongside the 2017 result, never instead of it.
• Wong A, Otles E, Donnelly JP, et al. “External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients.” JAMA Internal Medicine, 2021;181(8):1065–1070 — poor discrimination and calibration; missed roughly two-thirds of sepsis cases.
• Optum. 2024 Denials Index (124M remits, 1,400+ hospitals) — ~15% denial rate; 84% of denials characterized as potentially avoidable. See also Premier Inc. (2025) on initial denial rates and overturn rates, and Kodiak Solutions on net revenue lost to final denials.
• “The AI Arms Race In Health Insurance Utilization Review.” Health Affairs, 2025; and American Medical Association reporting on AI-driven prior-authorization denials.
• NSI Nursing Solutions. 2025 National Health Care Retention & RN Staffing Report (450 hospitals, 37 states) — 16.4% staff RN turnover; $61,110 average cost to replace one RN.

A note on sourcing: several statistics commonly circulated in this market — including widely quoted prior-authorization AI performance figures, referral-leakage percentages, and infusion chair-utilization benchmarks — could not be traced to a primary source during research for this article, and have been deliberately omitted rather than repeated. Where a figure in this article is vendor-reported rather than peer-reviewed, we have said so in the text.