You can't optimize your way to revenue that was never captured in the first place. That's not a theoretical argument — it's the thesis Forbes published on July 8, 2026, written by Ambience Healthcare's Chief Revenue & Value Officer Mike Valli. The revenue cycle was built around an expertise constraint that AI has now eliminated: you couldn't put a coder in the exam room during a visit. So physicians focused on care, documented afterward, and passed incomplete notes to back-office coders churning through backlogs without clinical context. Revenue leaked at the source. And the industry's response — layering AI onto downstream cleanup — optimizes a broken process instead of fixing it.
Meanwhile, patient records now average 60 documents and 50,000 words per encounter (AKASA, July 8, 2026), the coding workforce is aging out, and payers are deploying AI aggressively to optimize denials and audits. Practices still running traditional back-office coding aren't just inefficient — they're structurally incapable of capturing the revenue they've already earned.
The Source Problem: Why Downstream AI Can't Fix Upstream Revenue Loss
The revenue cycle's foundational design flaw is timing. A physician sees a patient. The visit generates clinical information — symptoms, examination findings, diagnostic reasoning, treatment decisions. The physician documents what happened, usually after the visit, usually under time pressure, usually with incomplete detail.
That documentation then moves to back-office coders. These coders work through queues of charts they weren't present for, in clinical contexts they didn't observe, applying codes to notes written for clinical purposes rather than billing accuracy. They're skilled professionals doing impossible work — translating secondhand clinical narratives into precise billing codes without the firsthand context that makes precision possible.
Forbes frames this clearly: the industry's AI response has largely followed the same broken path. Rather than fixing the source, most healthcare AI layers automation onto downstream cleanup steps — processing denial queues faster, automating appeal letter generation, scrubbing claims that were miscoded at the source. It's faster cleanup of the same structural mess.
"You can't optimize your way to revenue that was never captured in the first place." — Mike Valli, Chief Revenue & Value Officer, Ambience Healthcare (Forbes, July 8, 2026)
The distinction matters because downstream AI has a hard ceiling: it can only recover revenue from claims that were submitted. Revenue lost to undercoding, missed charges, incomplete documentation, or uncaptured complexity at the point of care never enters the system. No amount of denial management, appeal automation, or claim scrubbing can recover revenue that was never billed.
The Scale of the Problem: 60 Documents, 50,000 Words, No Coders
AKASA CEO Malinka Walaliyadde's July 8, 2026 interview with Healthcare IT Today quantifies why human coding can't keep pace:
Patient records now average 60 documents and 50,000 words. That's the equivalent of a short novel — for every single patient encounter. No human coder can read, comprehend, and accurately code a 50,000-word record at the speed practices need to maintain cash flow. The math doesn't work. It never did at this scale, and records are only getting longer.
Compounding the volume problem: the coding workforce is aging out. Walaliyadde reports that providers can't find replacement coders fast enough to maintain staffing levels. This isn't a recruiting challenge that better job postings solve — it's a demographic shift. The pipeline of trained medical coders is shrinking while the complexity and volume of records keeps expanding.
AKASA's response is institution-tuned LLMs — large language models customized for each institution's specific coding needs, complexity patterns, and payer requirements. Every hospital billing interaction is reviewed by an AI "copilot," with human review of results. Walaliyadde expects progression from copilot to autopilot for certain work types — moving from AI-assisted coding to AI-autonomous coding for straightforward cases while humans focus on complex edge cases.
The Evidence: What AI Coding Actually Delivers
Zedtreeo's July 8, 2026 analysis — "AI in Medical Billing 2026: What Actually Changed" — compiles third-party documented outcomes that separate marketing claims from verified results:
| Organization | AI Solution | Documented Result |
|---|---|---|
| Inova Health System | Nym Autonomous Coding | $1.3M annual savings, 50% DNFB reduction |
| Cleveland Clinic | Autonomous coding | 100 documents coded in 1.5 minutes |
| UCSF Health | H2O Document AI | 25,000 staff hours saved/year (1.4M faxes automated) |
| Moffitt Cancer Center | AI denial prevention | Denied revenue reduced 16% → 5% |
| 5-physician practice | DrCatalyst | $705K annual impact, AR days 52 → 19 |
The aggregate benchmarks are equally concrete:
- Clean claim rates: 84-88% → 94-98% with AI
- Denial rate reduction: 30-50%
- Days-in-AR reduction: 15-25 days faster
- Cost-to-collect reduction: 15-30% lower
- Average ROI: 451% (5-8x return)
These aren't projections. They're documented outcomes from named organizations with specific, measurable results. Cleveland Clinic coding 100 documents in 1.5 minutes isn't a vendor claim — it's a published operational metric. Inova's $1.3M in annual savings is auditable. The 5-physician practice cutting AR days from 52 to 19 is a scale-appropriate proof point for medical practices considering AI adoption.
The Touchless Claims Frontier
Athenahealth's 2026 RCM Trends report describes the trajectory: the industry is moving toward "touchless" claims — prior authorizations processed instantly, straightforward claims resolved in minutes, end-to-end without human intervention.
Touchless claims aren't just faster processing. They represent a fundamentally different architecture. In a touchless workflow:
- AI codes at the point of care — capturing the full clinical context during or immediately after the visit
- Pre-submission validation — AI checks coding accuracy, documentation completeness, payer-specific requirements, and prior authorization status before the claim is generated
- Automated submission — clean claims route directly to payers without human review for straightforward cases
- Exception routing — complex cases, unusual payer requirements, and edge cases flag for human review instead of failing silently
The clean claim rate improvement — from 84-88% to 94-98% — is the metric that proves the architecture works. When 94-98% of claims process without rework, the entire downstream infrastructure shrinks: fewer denials to manage, fewer appeals to write, fewer follow-up calls to make, less AR aging to chase. The denial management workload drops not because denials are handled faster, but because they don't happen.
The Payer AI Imbalance: Why Waiting Is Getting More Expensive
Forbes makes a critical strategic point that most practices underestimate: payers are deploying AI aggressively to optimize denials and audits. While providers debate AI adoption timelines, payers are already using AI to identify patterns in claims data, automate coverage determination denials, and optimize audit targeting.
This creates an asymmetric dynamic. Payers with AI denial optimization face providers with manual coding and paper-based appeals. The gap widens every quarter. Practices that wait for "proven" AI adoption are actually falling behind — they're competing against payer AI with human processes that can't match the speed, pattern recognition, or consistency of automated systems.
Healthcare IT Today's July 7 report on voice AI agents in RCM underscores the scale: SuperDial has conducted over 7 million voice AI calls for RCM clients. One customer cleared a 70,000-claim backlog in weeks. The vision: provider AI agents communicating directly with payer AI agents, with standards being developed for agent-to-agent payer-provider interactions.
This is where point-of-care AI creates the most strategic value. When coding is accurate at the source, AI-driven claims submission is cleaner, insurance verification is automated upstream, and the entire claims lifecycle requires less intervention. The practice's AI operates from a position of data strength — complete, accurate claims — rather than compensating for upstream data weakness.
Institution-Tuned vs. Generic: Why One-Size AI Fails in Coding
AKASA's approach highlights a critical architectural distinction: institution-tuned LLMs unlock capabilities that generic AI can't match. A general-purpose language model trained on internet text understands medical terminology. An institution-tuned model understands your specific payer mix, your specialty's coding patterns, your EHR documentation structure, and your historical denial triggers.
The difference matters in practice:
- Generic AI suggests CPT codes based on general medical knowledge — often correct at a surface level, frequently wrong on specificity, modifiers, or payer-specific requirements
- Institution-tuned AI codes based on your practice's specific patterns — understanding that Payer A requires modifier 59 for a procedure combination that Payer B processes without it, or that your documentation style for E/M leveling consistently maps to specific code selections
For medical billing, this distinction determines whether AI reduces coding errors or creates new ones. Generic AI that codes 90% correctly still generates a 10% error rate — which at scale means hundreds of miscoded claims per month, each requiring rework. Institution-tuned AI that achieves 97-99% accuracy eliminates most of that rework entirely.
The Five-Layer Point-of-Care Revenue Capture Stack
Combining the Forbes source-of-care thesis with AKASA's institution-tuned approach and Zedtreeo's outcome benchmarks, the architecture for AI-driven point-of-care revenue capture has five layers:
1. Ambient Clinical Documentation
AI listens during the encounter, generating structured clinical notes in real time. The physician focuses on the patient, not the keyboard. Documentation captures the full clinical context — not a post-visit reconstruction from memory. This eliminates the documentation quality problem that cascades into coding errors downstream.
2. Real-Time Coding
AI applies CPT, ICD-10, and modifier codes based on the documented encounter — while the clinical context is still fresh and complete. Institution-tuned models match codes to payer-specific requirements, specialty-specific patterns, and the practice's historical accuracy data. Coding happens in minutes, not hours or days.
3. Pre-Submission Validation
Before the claim is generated, AI validates coding accuracy against payer rules, checks documentation completeness, verifies eligibility status, confirms prior authorization alignment, and flags potential denial triggers. Claims that don't meet clean-claim criteria route to human review. Claims that pass go directly to submission.
4. Touchless Submission
Clean claims — the 94-98% that pass validation — submit automatically to payers without human intervention. The claim lifecycle from encounter to submission compresses from days to hours. Cash flow accelerates because the billing cycle starts at the point of care rather than waiting for back-office processing queues.
5. Exception Intelligence
The 2-6% of claims that don't pass validation don't fail silently. AI categorizes the exception — documentation gap, payer rule mismatch, coding complexity, authorization issue — and routes it to the appropriate human reviewer with specific instructions on what needs attention. This is the hybrid model: AI handles volume, humans handle judgment.
What This Means for Specialty Practices
The source-of-care thesis applies with particular force to specialty practices, where coding complexity is highest and the gap between clinical work and accurate billing is widest.
ENT practices face coding complexity around surgical procedures, imaging interpretations, and multi-code visits that require precise modifier assignment. Point-of-care AI that captures the full surgical encounter — procedure details, complications, bilateral/unilateral distinctions — codes accurately at the source rather than relying on operative note interpretation days later.
Dermatology practices deal with high-volume E/M visits interspersed with procedures — biopsies, excisions, destructions — where each service requires specific documentation to support medical necessity. AI that validates documentation completeness at the point of care prevents the downstream discovery that a biopsy note lacked the clinical indication language the payer requires.
Multi-provider practices benefit from institutional tuning that normalizes coding patterns across providers. When Provider A consistently undercodes E/M levels and Provider B overcodes, AI trained on the practice's data identifies and corrects both patterns — capturing missed revenue from Provider A while preventing audit exposure from Provider B.
The Cost-to-Collect Math
Zedtreeo's data puts the financial case in straightforward terms. AI-driven practices see cost-to-collect reductions of 15-30%. For a practice spending 8-12% of revenue on collections, a 30% reduction is meaningful margin improvement without any revenue growth required.
But the bigger number is the revenue that point-of-care coding captures that traditional coding misses. Undercoding, missed charges, incomplete documentation, uncaptured complexity — the revenue that was never billed. This is Forbes' core point: you can't optimize your way to revenue that was never captured. Downstream AI recovers denied revenue. Point-of-care AI captures revenue that was never billed in the first place.
The DrCatalyst case study illustrates the combined impact: a 5-physician practice generated $705K in total annual impact while cutting AR days from 52 to 19. That's $141K per physician per year in recovered and captured revenue — from a practice small enough that every dollar of overhead matters.
What to Evaluate in July 2026
The convergence of Forbes' source-of-care thesis, AKASA's institution-tuned AI capabilities, and Zedtreeo's third-party outcome data creates a clear evaluation framework for practices considering AI adoption:
1. Does the AI code at the point of care or downstream? If coding still happens in a back-office queue, the AI is optimizing the broken process — not fixing it. Point-of-care coding is the architectural requirement that determines whether revenue gets captured or lost.
2. Is the AI tuned to your institution's patterns? Generic coding AI creates new error categories. Institution-tuned AI that understands your payer mix, specialty coding patterns, and documentation structure delivers the 97-99% accuracy required for touchless claims.
3. Can you see the clean claim rate improvement? The benchmark is clear: 84-88% → 94-98%. If your AI vendor can't demonstrate this improvement with your data, the ROI case doesn't hold. Ask for measured clean claim rates, not projected ones.
4. What happens when AI is wrong? The best systems don't hide errors — they flag uncertainty and route exceptions to human review. Ask about the exception handling architecture, not just the accuracy rate. A 98% accuracy rate with silent failure on the other 2% is worse than a 95% rate with intelligent exception routing.
5. Does the AI address the workforce gap? With the coding workforce aging out, the question isn't whether to adopt AI coding — it's when. Practices that build AI coding infrastructure now operate from a position of strength. Practices that wait until their last experienced coder retires adopt AI from a position of crisis.
The revenue cycle is broken at the source. Forbes confirmed it. AKASA quantified it. Zedtreeo documented the fix. AI at the point of care captures revenue that downstream optimization can never reach — and with a 451% average ROI, the financial case closed before this article was published.