Since AI is now a common line item in the budget. I'll save you the headache from an overpriced consulting roadmap deck. If you or your vendor can't measure captured revenue, cost savings, or increased capacity on the back of a napkin, you aren't ready.

Vitalik just named the problem we've been solving for 18 months. We've been calling it "doctrine": critics run in a different trust domain than executors, no shared transcript, provenance-signed verdicts, real-signal verification. That's not doctrine. That's adversarial mechanism design for LLM swarms. Anti-collusion runtime is a product category. We're building it.

My LI feed is flooded with posts about the OpenAI and Epic integration. If your moat is connecting to an EHR to a LLM, you're screwed. While I agree, that is something someone can do in a weekend, these posts entirely miss the point. Every enterprise runs on a constellation of systems, and the AI value isn't in any single one, it's in the fusion layer that spans all of them. Practice management (billing, scheduling, claims — can be separate vendor from the EHR, sometimes multiple PMs across specialties) Payroll and HRIS (staffing cost, provider productivity, credentialing, PTO burn) CRM (referral relationships, attorney networks in PI, marketing attribution, patient outreach) ERP and finance (AP/AR, vendor contracts, supply chain, cost accounting) Google Analytics and marketing stack (funnel, campaign attribution, cost per acquired patient) Survey and CX systems (NPS, patient satisfaction, post-visit feedback, complaint intake) Imaging and lab systems (often separate from the EHR, especially at multi-site practices) Telephony and messaging (call center logs, SMS engagement, no-show recovery) Payer portals and clearinghouses (denials, aging, remit data) Document management and e-sign (intake packets, referral letters, HIPAA authorizations) Identity and access (SSO, MFA, role directory, offboarding) Compliance and audit logs (who accessed what, when, why) The questions enterprises actually want AI to answer live in the JOIN across those systems. "Why did our per-visit margin drop last quarter" is not an Epic question. It's a payroll-times-productivity-times-payer-mix-times-denial-rate-times-supply-cost question that touches six systems and requires the AI to know which join is legal and which is fabricated. "Which marketing spend produced the highest-LTV patients" is GA plus CRM plus PM plus payer data plus survey NPS — and the answer changes by service line and by site. "Why is our no-show rate spiking at the Charlotte office on Tuesdays" is scheduling plus telephony plus staffing plus weather plus the specific provider's panel — and no single system holds it.

The number you actually collect is a different number. Nobody in your office knows what the gap is. Not because they're lazy — because the gap only shows up when you compare billed vs. allowed vs. contracted, across every payer, every location, every remit from four clearinghouses. Your AR report shows the balance. It doesn't show the leak. What actually happens: → Aetna pays a shoulder scope at 84% of contract on commercial, 91% on MA. Two remits, two aging buckets, two billers. Nobody catches it. → Cigna downcodes 99214→99213 on 12% of visits at one location, 3% at another. It's a front-desk workflow you inherited in an acquisition three years ago. → UHC pays the assistant surgeon modifier for one partner, silently drops it for another. Eight months running. None of that is in your AR aging. All of it is in your bank. Your CFO doesn't need another dashboard. Your CFO needs one query: For every claim adjudicated in the last 12 months, what did the payer pay vs. what the contract obligated — ranked by dollar impact, with a confidence score on each expected rate. Diligence-grade artifact. Also a Tuesday-morning artifact — if the plumbing exists. It doesn't exist in your PM system. Your PM knows what it billed and what it received. It doesn't know what it should have received, because it never ingested the contract. Billed vs. allowed is a report. Allowed vs. contracted is a recovery. Most practices treat the second gap like the first. It's not. It's money you already earned, sitting in a payer's ledger, waiting for someone to notice. The 89-cent dollar isn't a payer problem. It's a plumbing problem.

Exactly this. Just published the same argument on LinkedIn (focused on Healthcare): lnkd.in/p/eC6djKBV

Old playbook: add sites, centralize, squeeze. Labor is scarce. Reimbursement is not stable. The model breaks. Post-close you inherit mismatched systems, local exceptions, and data that never becomes shared infrastructure. Headcount scales with volume because nothing compounds. Tech-enabled is a subscription list. Tech-multiplied is an operating system that absorbs volume without proportional people. The multiple now tracks the system, not the site count.

Most agent pilots never leave the sandbox. They get a demo on clean data. Then the real workflow shows up and it stalls. The agent drafts a prior-auth packet. Required fields have shifted. No persistent state across the EHR, clearinghouse, and work queue. A human opens three systems and re-enters it. The agent becomes another handoff. Appointment types don’t match. Insurance classes map differently. A payment from six months ago gets treated as current. An auth expires mid-cycle. It doesn’t own the step. It just creates another inbox. Production is the only test that matters.