medvidi.ai, the fourth AI feature, and the B2B question

Kicker

medvidi.ai is not a product site; it is a single Astro page on a bought SaaS template, hosted on Netlify, unlinked from medvidi.com, with a dead Privacy link, a dead Terms link, a commented-out "Download one-pager" button and a contact form that posts to Netlify Forms. Its Wayback history is the most useful thing in this Inquiry, because it shows the page in two states. In the September 2025 state, the AI Prescriber was labelled "Vision", dated "Est. 2027", and carried an explicit regulatory note reading "The AI Prescriber is not yet FDA-cleared." Sometime after 6 March 2026 — almost certainly on or around the 8 April 2026 launch release — the page was rebuilt. The rebuild deleted the "not yet FDA-cleared" note, deleted the "Verified by BrandPush.co" disclosure that had marked the "As Seen On" press wall as purchased, deleted three of seven executives including the SVP of Product and Engineering whose seat Ilia is interviewing for, deleted the 2027 date, and added an eight-item dated roadmap with before/after operating metrics covering quarters from Q1 2024 onward — that is, a retrospective narrative written in 2026 about work done in 2024. Every operating metric the Inquiry has been leaning on (PDMP ~23 min → one click, in-visit prescribing 40%→55%, end-of-call booking 22%→67%, ADHD return rate 41%→53%) comes from that retrospective block and did not exist in public before roughly April 2026. Meanwhile the "FDA pathway in progress" claim survives, and MEDvidi appears in no FDA database — not 510(k), not De Novo, not PMA, not registration and listing, not the AI-Enabled Medical Device List — while the regulator that actually authorises an AI to touch a prescribing decision in the United States today is a state office (Utah's Office of AI Policy, four authorised pilots, MEDvidi not among them) whose every agreement excludes controlled substances, which is MEDvidi's entire revenue core. On the strategic question: the intent to sell this stack is documented and real (~85%); a live B2B business with an external paying customer is not (~10%); and the best-fitting reading of the artefact as a whole is that its near-term audience is capital, not customers (~55%) — because the estate is 2.5 live brands rather than six, because the "over-built org" argument is arithmetically dead, and above all because the live careers board carries 65 requisitions and not one sales, partnerships, solutions-engineering or customer-success role.

Source legend

---

1. What we found

1.1 What medvidi.ai physically is

One HTML page. Astro v5.13.5, static, on Netlify [AI-LIVE]. Every navigation link in the header is a same-page anchor (#our-agent-stack, #ai-prescriber, #why-medvidi-is-different, #roadmap, #leadership, #partnership) [AI-LIVE]. There is no second page. robots.txt, sitemap.xml and sitemap-index.xml all return the 404 template [AI-LIVE].

The 404 template is the tell. Its <title> reads "Page Not Found - Sassify" — the name of the off-the-shelf Astro SaaS-landing template the site was built from — and it offers a "Contact Support" button pointing at /contact, a route that does not exist [AI-LIVE]. Nobody removed the template's boilerplate.

Four more artefacts of the same kind, all checkable in the DOM [AI-LIVE]:

• The hero's second CTA, "Download one-pager", is wrapped in an HTML comment: <!-- <a href="/" target="_blank" class="btn-outline …">Download one-pager</a> -->. The collateral either does not exist or was withdrawn. The href was / — it never pointed at a file.
• The footer's Privacy and Terms links are both href="#". A site that collects name, email, company and free-text message from health systems and investors publishes no privacy policy and no terms.
• The partnership form is commented in the source as "Partnership Contact Form with Netlify Forms Integration". There is no CRM tag on the page — no HubSpot, no Salesforce, no Marketo, no Pipedrive — and no meeting scheduler. Analytics is GTM GTM-5V3G9DTZ plus GA4 G-1ZZCBJ244S, nothing else.
• The leadership carousel renders seven slide indicators (aria-label="Go to slide 1" … "Go to slide 7") for four published leaders. Seven is the count from the previous version of the page (§1.3). The carousel was not rebuilt, only refilled.

The footer's only badge-shaped mark is a 512×512 PNG that reads "FROM SILICON VALLEY WITH LOVE" with a heart [AI-LIVE]. It is a decorative sticker, not an accreditation, an accelerator or an award. It sits where a SOC 2, HITRUST or HIPAA seal would sit on a site selling to health systems.

medvidi.com does not link to medvidi.ai anywhere. Zero occurrences of the string medvidi.ai in the rendered consumer homepage [AI-LIVE]. The B2B site is reachable only from the press releases that cite it. It is a deck substitute you send to a named person, not a channel.

medvidi.ai carries a registry expiry of 2027-05-14 with NameCheap [WHOIS]. .ai registrations are two-year terms, which puts first registration at roughly May 2025 [REASON] — four months before the site went live. medvidi.com was created 2020-04-08 [WHOIS].

1.2 The archive shows two different sites, and dates the seam

Eight captures exist, and they fall into exactly two groups [AI-CDX] [AI-ARC]:

2025-09-22 (x2)   Vite build.  assets/index-D3aWVQmL.js, vite.svg.
                  Two captures, one digest change between them (same 73,461 bytes of text).
2025-10-06        digest DFGDXKZ2Q5MQ3LKZ63TKDDAYBHPYQV2C   15,354 bytes stored
2025-11-09        same digest                                15,357
2025-12-12        same digest                                15,362
2026-01-21        same digest                                15,358
2026-02-14        same digest                                15,354
2026-03-06        same digest                                15,358
[no capture]
2026-08-19        Astro v5.13.5 build, 82,657 bytes live.  Entirely different page.

Two facts follow, and both matter.

First: the page did not change for five months. From 6 October 2025 to 6 March 2026 the served bytes were identical [AI-ARC]. A company running a live partner motion edits its one sales page more often than never.

Second: the current version is not in the archive at all. The rebuild happened between 2026-03-06 and 2026-08-18 (the first Scout fetch) [AI-ARC] [PRIOR]. It can be narrowed: the current hero copy — "Built on 130,000+ visits", "multiplies clinician capacity 10x", "122 million Americans" — is verbatim or near-verbatim from the 8 April 2026 launch release [PR-APR26] [AI-LIVE], and none of those phrases appear in the September 2025 version [AI-ARC]. The most likely date is on or about 8 April 2026 [REASON]. State it as a bracket, not a date: rebuilt between March and August 2026, with copy that first appears in the April release.

This is the load-bearing methodological point of this Finding. The dated 2024–2026 roadmap with before/after operating metrics is not a contemporaneous log. It was published in a rebuild in 2026 and back-dated to quarters running from Q1 2024. Nothing in the public record shows those numbers being asserted at the time the work was done.

1.3 What the rebuild deleted

Diffing the September 2025 text against the live text [AI-ARC] vs [AI-LIVE] produces six deletions. Ranked by how much each would cost the company to be asked about.

(a) The regulatory disclaimer. The September 2025 page had a "Technology & safety" sub-block headed "Regulatory note", reading in full:

"The AI Prescriber is not yet FDA-cleared. MEDvidi will release autonomous capabilities only under applicable regulations and physician oversight." [AI-ARC]

That sentence is gone from the live page [AI-LIVE]. What remains is the phrase "FDA pathway in progress", inside a comparison table cell, plus a meta description advertising an "FDA-pathway AI Prescriber" [AI-LIVE]. So the company went from disclosing that it is not cleared to implying that a clearance process is running, in the same rebuild that moved the product from "Vision" to "launched". Section 1.9 tests whether any such process exists in public.

(b) The 2027 date, and the provenance of "10x". The September 2025 roadmap was three steps, not eight [AI-ARC]:

Past      Manual process        Providers do admin work and follow-ups
2025      AI Clinical Assistant Automates administrative work   1.67x more patient time
Est. 2027 AI Prescriber         Handles follow-up visits        6.67x more patient time
          Combined Impact       10x more patient time per clinician
                                (AI Clinical Assistant + AI Prescriber)

Two things die here. First, the AI Prescriber was the company's own "Est. 2027" item as of September 2025 and was announced as launched on 8 April 2026 — roughly nine to twenty-one months early, depending which end of 2027 was meant [REASON]. Either the roadmap was sandbagged, or what shipped in April is a narrower thing than the 2027 item described. That is the sharpest single question available for the CEO.

Second, "10x" was explicitly the combined 2027-inclusive figure, made of 1.67× from the Clinical Assistant plus 6.67× from the not-yet-built Prescriber [AI-ARC]. The live hero now says the suite "multiplies clinician capacity 10x" in the present tense [AI-LIVE], and the April release says the tool is "enabling clinicians to see up to 10X more patients" [PR-APR26]. The 10× is a 2027 projection re-tensed as a shipped result. This is documented, dated and checkable, and it is exactly the kind of thing a candidate should raise as a measurement question rather than an accusation.

(c) The disclosure that the press wall is purchased. The September 2025 "AS SEEN ON" strip carried the line "AND OVER 500 NEWS SITES — Verified by BrandPush.co" [AI-ARC]. BrandPush is a paid press-release syndication service. The live page keeps the logo wall (Benzinga, Business Insider, MSN) and drops the attribution [AI-LIVE]. The underlying links are unchanged and self-identifying: the AP link is under apnews.com/**press-release**/getnews/ [AI-LIVE] [PR-SEP25].

(d) Three of seven executives. September 2025 published seven [AI-ARC]; August 2026 publishes four [AI-LIVE]:

Sept 2025                                        Aug 2026
Laura Purdy, MD, MBA      — Chief Medical Officer   (gone)
Vasili Razhnou            — Co-founder, CEO         Vasili Razhnou   — Co-founder, CEO
Alex Kulitski             — Co-founder              Alex Kulitski    — Co-founder
David Okonkwo, MD, MBA    — Medical director        (gone)
Oleg Gorbylev             — COO                     (gone)
Konstantin Valiotti       — SVP of Product and Eng. (gone)
Yauhen Zaremba            — VP of Marketing         Yauhen Zaremba   — VP of Marketing
                                                    Aneel Ursani, MD, MBA, FAPA — CMO (new)

Konstantin Valiotti is the seat. He is on medvidi.ai as "SVP of Product and Engineering" in September 2025 [AI-ARC]; he is quoted as "VP of Product, Engineering, and AI" in MEDvidi's own release of 5 February 2026 [PR-FEB26]; he is absent from the live leadership block [AI-LIVE]. That is a much harder vacancy signal than the Comeet requisition timestamps the Inquiry has been using, and it dates the departure to the window February–August 2026 [REASON]. It also confirms the scope: the previous holder already owned Product and Engineering, so the "VP of Product over PM + Eng + Design + Data + PMM" scope in the posting is a restoration, not an expansion. That contradicts findings/company-org-role.md §1.8's hypothesis H3 framing [PRIOR].

(e) The partner-deployment claim. September 2025: "Deployed in isolated VPCs with robust security controls and privacy guardrails. (Implementation varies by partner environment.)" and "Clinically meaningful rationales and comprehensive dashboards for payors, regulators, and internal QA." [AI-ARC]. Both gone. The live equivalent is "HIPAA-aligned by design… all documentation and charting meet regulatory standards" [AI-LIVE] — vaguer, and with no deployment topology at all. The rebuild moved the page away from selling deployable software and toward describing a clinic's internal capability. That cuts against the B2B thesis, not for it, and no prior Finding could see it.

(f) Two speed and outcome claims. "We moved from prototype to production AI in ~3 months" and "AI scribes free ~10 hours/week per provider and boost revenue ~30%" are both gone [AI-ARC] → [AI-LIVE].

1.4 The trust rule for this source, stated once and applied consistently

verification/gaps.md §4 C9 is right that A5 dismissed this site as marketing where inconvenient and used it as data where convenient [PRIOR]. Here is the rule this Finding uses. It is mechanical, so a reader can audit it.

Applied: the roadmap's structure (what was built, in what order, on what surface) is class A and is good evidence — it is corroborated by the portal bundle, the ATS and the consumer site. The roadmap's numbers are class B. The rest is C or D.

Practical consequence for the room. Do not say "MEDvidi's PDMP integration saved 23 minutes a check." Say: "Your own roadmap says the PDMP check was ~23 minutes and is now one click. If that is per new patient, it is about two-thirds of the 30-hours-a-month claim on its own — bigger than the scribe. Is that the right read?" The first sentence can be wrong. The second cannot.

1.5 The operating metrics, one at a time

All from the live roadmap [AI-LIVE], all class B. For each: what is measured, over what population, and what it would imply if true.

Two of these are worth more than the rest.

M8 is the number the study needed and nobody joined up. verification/gaps.md §1 spotted that A5 held it and A3 held the 3.43-visits-per-patient figure and neither combined them [PRIOR]. Stated plainly: if the ADHD return rate is 53%, then just under half of ADHD patients — the largest and highest-margin segment — never return after their first visit. That makes 3.43 substantially a churn number, not a days-supply number. It also makes M7 (end-of-call booking, 22%→67%) the highest-leverage shipped thing on the whole roadmap, because it is the mechanism that moves M8.

M15 is internally consistent, which is a point in the source's favour.

1,500 hours/month ÷ 50 providers          = 30 h/provider/month   [REASON]
Company's independent claim, Apr 2026:      "30+ hours per provider each month"  [PR-APR26]
187 workdays: 1,500 h ÷ 8 h                = 187.5 workdays        [REASON]

The arithmetic closes. That does not make it measured, but it means the page is not innumerate, and it means the "8 hours per week" headline [AI-LIVE] is the same claim in different units (30 h/month ÷ 4.33 weeks = 6.9 h/week [REASON] — so "8 hours per week" is roughly 15% richer than "30+ hours per month", the only internal inconsistency found).

One class-C statistic is distorted, and the archive proves it. September 2025 read: "Providers spend ~16 hours per week on paperwork — charting, PDMP checks, prescribing — consuming 25% of total healthcare spend" [AI-ARC]. That is the well-known US finding that administrative activity accounts for roughly a quarter of health expenditure. The live page reads: "AI assistant helps manage ongoing care, which saves providers 30+ hours monthly to evaluate and support new patients and helps reduce total spend by 25%" [AI-LIVE]. A statistic about how much of national health spend is administrative has been rewritten into a claim that the product cuts total spend by a quarter. Same number, opposite meaning. Flag it as a copy defect, not as fraud — but it is the reason class C is "ignore".

Similarly, the population framing moved: "32 million Americans face barriers to treatment" on 5 February 2026 [PR-FEB26] became "122 million Americans already lack mental health access" by April [PR-APR26] [AI-LIVE]. Both are real US figures measuring different things (the larger is the HRSA count of people living in a mental-health professional shortage area). The site now uses the bigger one.

And the anchor claim itself moved: "85% of visits are follow-ups" (Sept 2025) [AI-ARC] → "80% of psychiatric visits are routine renewals" (live) [AI-LIVE]. The live page contains both — 80% in the desktop table, 85% in the mobile table, on the same page, today [AI-LIVE]. Nobody finished the edit.

1.6 What the roadmap actually shows, once the numbers are set aside

Read as class A — sequence and surface rather than magnitude — the roadmap is genuinely informative, and it says something the job posting does not.

Q1 2024      Automated intake                  identity verification, pharmacy auto-attach
Q3–Q4 2024   Prescribing from the EHR          DoseSpot inside the EMR, multi-state
2025         National coverage                 multi-licence provider strategy
Early 2025   PDMP in EHR + treatment continuity  one-click PDMP; end-of-call booking
Q2 2025      Personalized ADHD booking         two flows: new vs transfer
Q2 2025→2026 AI charting             [AI]      limited beta -> "not yet at full fleet coverage"
Q3 2025      AI chart reviewer        [AI]     2-3% -> 100% coverage
Late 2025    AI Receptionist launch   [AI]     SMS + email; "voice channel in development"
Early 2026   Algorithmic availability maximisation   schedule-gap compression

Five of nine items are not AI. They are integrations, workflow placement and funnel design. The four tagged [AI] arrive in the last five quarters. A5's contrarian claim — that MEDvidi's biggest wins came from integrations, not models — survives this Finding and is strengthened by it, because the sequence is now visible independent of the magnitudes [PRIOR].

Two items deserve naming because no other Finding has them:

• "Early 2026 — Algorithmic availability maximisation." "Provider schedules algorithmically compressed to maximise availability — fragmented gaps closed automatically, unlocking new patient capacity from existing schedules without additional hires." [AI-LIVE] This is a fifth shipped capability, it is a scheduling optimiser rather than a model, it is the most recent thing on the roadmap, and it appears in neither the job posting nor any of the seven Findings. In a business where clinicians are paid per completed visit and calendar density is the margin, it may be the highest-ROI item on the list.
• "Voice channel in development" as of Late 2025 [AI-LIVE], against the April 2026 release's "An AI Receptionist handles rescheduling via SMS and voice" [PR-APR26]. A5 parked whether voice is live; the two company sources still disagree, and the roadmap entry is the older of the two. Treat voice as pilot-or-recent, not steady state.

1.7 The fourth AI feature: what the AI Prescribing Assistant is claimed to do

Verbatim from the launch release [PR-APR26]:

"MEDvidi… launches its AI Prescribing Assistant, which helps clinicians across the US manage routine medication renewals for patients with ADHD, anxiety, and depression. While the system automates workflow, all prescribing decisions remain under the control of licensed physicians. Built on data from 130,000+ psychiatric visits, the tool is already cutting 30+ hours of administrative work per provider each month and enabling clinicians to see up to 10X more patients."

"The AI Prescribing Assistant works as a clinical verification layer, grounded in evidence-based guidelines and MEDvidi's proprietary dataset of thousands of historical visits. Crucially, the AI does not prescribe independently; every decision is reviewed and approved by a licensed physician."

"It reviews patient responses to treatment, checks adherence to clinical guidelines, ensures documentation meets regulatory standards, and flags potential safety or compliance considerations."

And from the live site's comparison table [AI-LIVE]:

Note the collision inside the company's own copy. The release says "every decision is reviewed and approved by a licensed physician"; the table says "Automated (80% of visits)" and "Instant (self-service follow-ups)" and "scales without adding doctors". Both cannot describe the same workflow at the same volume. If a physician reviews and approves every renewal, the physician's time is still in the loop and the capacity claim shrinks to the review-time saving. If the renewals are self-service and instant, the physician is not reviewing each one. The economic claim and the compliance claim are load-bearing against each other, and which one is true determines both the regulatory class (§1.8) and whether the feature is worth building (§1.11).

1.8 Where the Cures Act CDS exemption stops and FDA device regulation starts

Statute: §3060(a) of the 21st Century Cures Act added FD&C Act §520(o)(1)(E), which excludes software from the device definition only if all four criteria hold [FDA-CDS]:

1. It does not acquire, process or analyse a medical image, an IVD signal, or a signal from a signal-acquisition system.
2. It displays, analyses or prints medical information about a patient, or other medical information such as peer-reviewed studies and practice guidelines.
3. It supports or provides recommendations to a health care professional about prevention, diagnosis or treatment.
4. It enables the HCP to independently review the basis for those recommendations, such that it is not the intent that the HCP rely primarily on them.

FDA's implementing guidance was reissued 6 January 2026 (re-published 29 January 2026), superseding the 28 September 2022 final guidance [FDA-CDS]. Four changes matter here:

• Single recommendations are now tolerated. The 2022 guidance effectively required a list of options; the 2026 guidance exercises enforcement discretion for a single clinically appropriate recommendation, giving as its own example software that recommends "a specific FDA-approved drug for an HCP to consider based on the patient's symptoms and medical history" [FDA-CDS]. That example is close to MEDvidi's described behaviour, and it is the most favourable regulatory development available to them.
• The source of the recommendation is constrained. Recommendations must be grounded in "well-understood and accepted sources, such as clinical guidelines and peer-reviewed literature" [FDA-CDS]. MEDvidi markets the opposite as its moat: "grounded in evidence-based guidelines and MEDvidi's proprietary dataset of thousands of historical visits" [PR-APR26], and the site calls the dataset "domain-specific AI that generic models cannot replicate" [AI-LIVE]. To the extent a recommendation is driven by an unpublished internal corpus rather than by published guidance, criterion 4 gets harder, not easier, because the clinician cannot independently review a basis they cannot see.
• Time-critical output breaks criterion 4. The 2026 guidance relocates the time-critical restriction from criterion 3 to criterion 4, on the theory that urgency prevents independent review [FDA-CDS]. "Instant, self-service" [AI-LIVE] is a design that compresses review time toward zero. Same logic, and it is the company's own word.
• Generative AI and patient-facing tools are simply not addressed. The guidance contains no AI-specific language, and consumer-facing decision support, symptom checkers and health chatbots are excluded from it entirely [FDA-CDS]. So the AI Receptionist — a patient-facing agent on SMS, chat and voice — sits in a regulatory void, not in an exemption. FDA has signalled a device-software policy revision in FY26 [FDA-CDS].

The bottom line a VP of Product should be able to state in one breath. Under the January 2026 guidance, an AI that hands a clinician a reviewable renewal recommendation, sourced from guidelines, with the clinician genuinely deciding, is very likely non-device CDS — which needs no FDA submission at all. An AI that issues the renewal because reviewing each one would destroy the capacity claim is not CDS, and then the question is which device pathway applies. The two versions of MEDvidi's own copy sit on opposite sides of that line.

1.9 The "FDA pathway in progress" claim, tested

Every public FDA device register, queried 2026-08-19 [FDA-DB]:

openFDA device/510k                search "medvidi"   -> No matches found
openFDA device/pma                 search "medvidi"   -> No matches found
openFDA device/registrationlisting search "medvidi"   -> No matches found
openFDA device/classification      search "medvidi"   -> No matches found
openFDA device/recall              search "medvidi"   -> No matches found
openFDA device/enforcement         search "medvidi"   -> No matches found
openFDA device/udi                 search "medvidi"   -> No matches found
FDA AI-Enabled Medical Device List (393,701 bytes)    -> 0 occurrences of "medvidi"
openFDA registrationlisting        "razhnou"          -> No matches found
openFDA registrationlisting        "kulitski"         -> No matches found
openFDA registrationlisting        "medvidi inc"      -> No matches found

Sibling-brand names were checked to avoid a false negative. "ezcare" returns two 510(k)s and a registration — all unrelated: a dental handpiece cleaner (K113674, Handpiece Headquarters) and the Versatile 1 EZCARE wound vacuum (K061919, Blue Sky Medical, now Smith & Nephew) [FDA-DB]. There is no MEDvidi-linked device record under any name checked.

Separately, SEC EDGAR full-text search returns zero hits for "MEDvidi" across all filings [FDA-DB]. No Form D, no S-1, no mention in anyone else's filing.

How to state this without overreaching. Three things are true simultaneously and all three must be said together:

1. There is no public FDA record of any kind for MEDvidi. That is a fact about the registers, checked across seven databases and the curated AI list, and it is stated as a fact.
2. Absence of a record is not proof of absence of a process. FDA's Q-Submission programme — pre-submissions, informational meetings, breakthrough-device requests — is confidential and does not appear in any public database until a marketing authorisation issues. A company that filed a Pre-Sub last month would look exactly like this. So "in progress" is unfalsifiable from outside, which is precisely why it is a comfortable thing to print.
3. The bigger problem is that the claim may be category-incoherent. If the product is non-device CDS under the January 2026 guidance — which is what the "clinical verification layer, physician approves everything" framing asserts — then there is no FDA pathway to be in progress, because non-device CDS requires no FDA submission. "FDA pathway in progress" is therefore either an admission that the company believes its product is a device (contradicting its own safety framing), or it is a phrase doing marketing work. Both readings are worth a question; neither is worth an accusation.

The honest interview line: "Your site says FDA pathway in progress and there is nothing public — which is expected, since Q-Subs are confidential. What I want to understand is the intended classification. If the Prescribing Assistant is non-device CDS under the January 2026 guidance, there is no submission to make. If it isn't, the grounding-in-proprietary-data language is the thing I'd change first, because that is the criterion-4 problem."

Note also the timing asymmetry: the September 2025 page said "not yet FDA-cleared" [AI-ARC] — which implies believing a clearance is needed. The live page dropped that sentence in the same rebuild that launched the product [AI-LIVE]. That happened after the 6 January 2026 guidance made the non-device route easier [FDA-CDS]. That is a plausible and benign explanation and should be offered as one.

1.10 The regulator that actually gates this today is a state, not FDA

The most useful thing found in this pass. As of August 2026 the only US jurisdiction with a public, named, contractual authorisation regime for AI-issued prescription renewals is Utah's Office of Artificial Intelligence Policy, operating "regulatory mitigation agreements" under statutory authority [UT-OAIP].

The complete list of authorised pilots, as published on 2026-05-18 [UT-OAIP]:

Legion Health Agreement     AI psychiatry refills (non-controlled)
Doctronic Agreement         Oct 2025 - Oct 2026, renewable
Dentacor Agreement
ElizaChat Agreement

MEDvidi is not on it.

The Legion Health terms establish, in a public instrument, exactly what an AI may touch in a prescribing decision [UT-OAIP]:

• May do: renew an existing prescription originally ordered by a licensed physician.
• May not: "initiate new treatments, adjust dosages, or prescribe higher-risk medications"; may not diagnose.
• Formulary: "a unique formulary for each company pilot, ranging from a list of about 20 to about 200 approved medications… and do not include any controlled substances."
• Phasing: first 250 requests reviewed by a clinician before completion; next 1,000 under intensive retrospective review; then monthly randomised audit. Reported agreement-rate threshold above 98% to advance.
• Only in phase two may the AI "submit the prescription renewal directly to the pharmacist", with the pharmacist retaining escalation authority and a direct line to a licensed professional at the prescribing company.
• Escalation triggers: suicidality or self-harm, mania or hypomania, pregnancy status change, severe adverse effects, worsening symptoms, identity or prescription mismatch; patients may demand human review at any time.
• Refill cap: "a limited number of refills per prescribed medication without a subsequent in person or telehealth visit."
• Liability: a named licensed physician on every AI-generated renewal, plus a requirement to "maintain medical malpractice insurance that covers AI liabilities."
• Term: 12 months, temporary, conditional, with monthly reporting to OAIP.

This is the shape of the answer to "what may an AI touch". It is a scope-of-practice and state-licensure question first, and an FDA question second. The FDA question is about whether the software is a device. The Utah question is about whether the prescription is lawfully issued — and that is the one that can close a clinic.

Two consequences.

One: a direct competitor has the instrument MEDvidi does not. Legion Health is an AI-native psychiatry clinic doing the identical thing, with a named public agreement, a published formulary boundary and a phased-validation protocol. That is a defensible, auditable regulatory position and a genuine sales asset. MEDvidi's equivalent asset is the phrase "FDA pathway in progress" on a marketing page.

Two: the formulary exclusion is fatal to the revenue core. Every OAIP pilot excludes controlled substances [UT-OAIP]. MEDvidi's information architecture is organised around Schedule II–V branded drugs, and its own patient FAQ states that established patients may receive urgent refills without a visit "excluding controlled substances" [PRIOR]. So the AI Prescribing Assistant, in any form a regulator has so far blessed anywhere in the United States, can only touch the non-controlled remainder: SSRIs, SNRIs, bupropion, trazodone, atomoxetine, guanfacine, clonidine, and the GLP-1 line.

1.11 What the fourth feature is actually worth, once the exclusion is applied

A5 modelled the cannibalisation risk on the assumption that the automation applies to all follow-ups [PRIOR]. With §1.10 the model has to be re-run on the addressable slice.

Inputs (all published, all class A or company-stated)
  Annual visits, 2025                        120,000       [PR-APR26]
  Follow-up share of visits                  ~80%          [AI-LIVE], company's own premise
  Follow-up price                            $159          [PRIOR] [WEB]
  ADHD / controlled-substance share of the book: NOT PUBLISHED — the pivotal unknown

Scenario A - controlled substances are 60% of the follow-up book   [REASON, assumption]
  Follow-ups                    120,000 x 0.80        =  96,000
  Non-controlled follow-ups     96,000 x 0.40         =  38,400
  Automatable at OAIP-style scope                     =  38,400 visits/yr
  Revenue exposed if repriced to zero  38,400 x $159  =  $6.11M   (~23% of $27M)
  Clinician cost avoided at $37.50/visit              =  $1.44M
  Net if simply removed                               =  -$4.67M

Scenario B - controlled substances are 80% of the follow-up book   [REASON, assumption]
  Non-controlled follow-ups     96,000 x 0.20         =  19,200
  Revenue exposed                                     =  $3.05M   (~11% of $27M)
  Clinician cost avoided                              =  $0.72M
  Net if simply removed                               =  -$2.33M

[REASON]. The controlled-substance share is not published and is the single number that sizes this feature. Ask for it.

The conclusion holds in both scenarios and is the useful one: the AI Prescribing Assistant is not a productivity feature, it is a pricing decision. It is worth building only if the asynchronous renewal is repriced — as a membership, as an async-visit fee, or as software sold to somebody else — and only across the non-controlled book, which is the smaller and lower-margin half of the business. A5's conclusion was right; the magnitude was wrong, because A5 applied it to the whole follow-up line [PRIOR].

And note what falls out of this: the feature that would help MEDvidi most is the one it cannot legally build, and the feature it can build competes with its own $159 fee in the segment where it has least pricing power. That asymmetry is the real content of "AI Clinic Track".

1.12 The rest of the estate: six brands in the config, two and a half in the market

A1 read six tenant keys out of the portal bundle — ezcareclinic (default), mmjdoctor, esacare, medvidi, mangoclinic, mdberry [PRIOR]. The multi-tenancy is real in the code. The market reality is thinner. All checked 2026-08-19 [SIB]:

Details that matter:

• ezcareclinic.com 301s to ezcareclinic.io, which now serves <title>MEDvidi | Mental Health Clinic USA [SIB]. All its social links point to MEDvidi's accounts (facebook.com/medvidicom, tiktok.com/@medvidi_com, linkedin.com/company/medvidi), and its /team/ pages are the same nine physicians as medvidi.com [SIB]. This is not a sibling brand. It is MEDvidi's own decommissioned SEO property, find-and-replaced from "EZCare" to "MEDvidi" and abandoned.
• mangoclinic.com returns 301 → https://medvidi.com/ [SIB]. A brand consolidation no Finding recorded.
• mmjdoctor.com is a WordPress blog with two posts, dated 24 and 30 March 2023, on psychotherapy and "How to Get Xanax Prescribed to You Online" [SIB]. No booking, no pricing, no clinic.
• mdberry.com is 17KB of unedited hospital-template boilerplate — "When you arrive, unless you're told otherwise, please go directly to the Admitting Department, located in the main hospital lobby" [SIB]. There is no clinic behind it.
• esacare.com is the only genuinely separate live business: ESA letters at $99/year Basic (1 physician visit, 1 letter) and $139/year Plus (1 visit, 2 letters), both shown struck through from $149 and $189, with a 10%-off email-capture popup [SIB]. It runs on nginx 1.19.6 — a 2020 build — which is its own small statement about how much attention the asset gets.

This materially weakens the "six-brand multi-tenant platform proves B2B readiness" argument that the completeness critique leans on [PRIOR]. The configuration model is real and is good engineering. But it has been exercised against two and a half live consumer brands, all owned by the same company, none of them an external tenant. Multi-tenancy that has never served a tenant it does not own is a design, not a proof.

1.13 Therapy, reconciled — contradiction C6 closed

verification/gaps.md §4 C6 records an unresolved conflict: A1, A2 and A6 concluded MEDvidi exited therapy; A7 recorded that sibling EZCare still sells 60-minute therapy sessions with a personal licensed therapist, in 20 states including New Jersey and Oklahoma [PRIOR].

The live evidence settles it, and neither side was right. ezcareclinic.io, today, says verbatim [SIB]:

"MEDvidi is an online mental health treatment center that provides our patients with professional ongoing therapy and medication management." "Step 2 – Talk to your personal licensed therapist…" "Follow-up therapy sessions last for 60 minutes each. Follow-ups with a prescriber last for 15 minutes each." "All MEDvidi medical providers are licensed therapists… PsyD, a PHD, or… Licensed Clinical Social Worker, Licensed Mental Health Counselor, Licensed Professional Counselor…" "…prescribing providers who are licensed to serve residents of Texas (TX), Florida (FL), California (CA), New York (NY), Pennsylvania (PA), Ohio (OH), North Carolina (NC), Colorado (CO), Illinois (IL), Indiana (IN), Virginia (VA), Arizona (AZ), Michigan (MI), New Jersey (NJ), Wisconsin (WI), Missouri (MO), New Mexico (NM), Washington (WA), Oklahoma (OK), Oregon (OR)."

So: the company did exit therapy — and left a live, MEDvidi-branded page promising every patient a personal licensed therapist and 60-minute therapy sessions, in a 20-state list that contradicts the 35/36 states it advertises elsewhere. It is not a second business line. It is content debt with legal teeth. A public page in the company's own name, describing a service it does not provide and a state footprint it does not have, is precisely the Section 5 deception surface A4 spends its Finding worrying about [PRIOR]. Its nav also lists "Chronic Fatigue Syndrome" — a condition absent from medvidi.com and adjacent to the narc service code A1 found live in the portal [SIB] [PRIOR].

Cost to fix: one afternoon. It is the cheapest credible thing a new VP of Product could point at in week one.

1.14 The publicity programme, and who it is for

Four wire releases in eight months, on four different paid distribution networks [PR-SEP25] [PR-FEB26] [PR-APR26] [AI-LIVE]:

2025-09-01  "Introduces AI-Powered Healthcare Solutions to Revolutionize Telemedicine"
            GetNews -> AP News press-release channel, Benzinga.  BrandPush-verified. [PR-SEP25]
2026-01-15  "Launches AI-Powered Healthcare Solutions to Transform Telemedicine"
            MarketersMedia -> FinancialContent, Business Insider markets channel.
2026-02-05  "Announces AI Clinical Assistant to Address the Shortage of Mental Health
            Professionals"  ACCESS Newswire -> Yahoo Finance, MSN, NewsBreak.  [PR-FEB26]
2026-04-08  "Launches AI Prescribing Assistant to Tackle America's Mental Health Access
            Crisis"  GLOBE NEWSWIRE -> Yahoo Finance, Manila Times, AI Journal.  [PR-APR26]

Every one is a product/AI announcement. None is patient acquisition. Wire syndication of this kind does not reach patients searching for "adderall prescription online" — it reaches people who read financial wires. For a founder-owned, unfunded company [PRIOR], that is a deliberate spend aimed at a specific audience.

Caution for the room: "MEDVi" is a different company (telehealth ED and GLP-1) and shows up in the same searches [PR-APR26] context. That name collision also partly explains the contaminated BBB narratives A5 flagged [PRIOR]. Do not quote a MEDVi number as MEDvidi's.

---

2. Capability / object table

---

3. Reconciliation notes

R1 — C9 (can medvidi.ai be trusted?) is now answerable. §1.4 sets a five-class rule and §1.2 gives the reason: the metrics block is retrospective, published in a 2026 rebuild, about quarters from 2024. Class B — cite as the company's claim, never as a measurement. A5's substantive conclusion (integrations beat models) survives, because it rests on the roadmap's sequence, which is class A and independently corroborated [PRIOR].

R2 — C6 (therapy) is closed, against both sides. §1.13. MEDvidi did exit therapy; the "sibling still sells it" evidence is stale copy on MEDvidi's own decommissioned domain, not a live business line. A6's positioning argument stands; A7's contradiction dissolves into a content-debt finding with an FTC edge [PRIOR].

R3 — the "over-built org" leg of the B2B thesis is dead and should not be used. verification/gaps.md §6 argues the B2B thesis partly from "53 people against $20–27M, five to seven times below the nearest comparable on revenue per head" [PRIOR]. verification/quantitative.md §2.3 had already destroyed that number: A3 divided $20.3M by whole-company headcount (200–300) and then attributed the ratio to the 55-person product-and-engineering org. Corrected, it is $369k–$545k per product+eng head against LifeMD's $499k per employee — at or above the comparable, not 5–7× below — and LifeMD's 389 excludes its affiliated medical group, so the comparison is not like-for-like anyway [PRIOR]. Do not say the org is oversized. The surviving and better version of the argument is the one quantitative.md itself supplies: a clinic doing 180,000 visits a year does not need thirty engineers for a booking flow. That is a question about what they are building, not about affordability.

R4 — C13 (profitability) is settled and dated. The "$20M ARR and profitability" claim is in the CEO's bio, which is unchanged since September 2025 [AI-ARC] [AI-LIVE]. It is a Sept-2025 statement still sitting on a page rebuilt in 2026. The April 2026 release independently states $27M ARR and 100% YoY growth [PR-APR26]. Those cohere:

$27.0M in Apr 2026 at 100% YoY  ->  ~$13.5M in Apr 2025          [REASON]
$20M in Sep 2025 sits almost exactly on that line                [REASON]
Extrapolating the same rate to Aug 2026 gives roughly $34-38M    [REASON, do not quote]

The consistency is a genuine point in the sources' favour and is worth knowing. The extrapolation is not; it assumes a growth rate holds, and the last dated figure is four months old. Quote $27M as of April 2026, note $20M is a stale Sept 2025 bio line, and say you would expect the current number to be higher.

R5 — state count. Three live company numbers: the roadmap says 33 [AI-LIVE], the April release says 36 [PR-APR26], the consumer FAQ has said 35 [PRIOR], and ezcareclinic.io lists 20 [SIB]. The roadmap's 33 is labelled "2025", so it is a different vintage, not a contradiction. Use 36 as of April 2026, and note the 20-state page is stale.

R6 — visits. 120,000 annual (2025) [PR-APR26]; "10,000+ real patient visits per month" [PR-APR26]; "15,000+ monthly patient encounters" [AI-LIVE]; "130,000+ psychiatric visits" as the training corpus [PR-APR26] [AI-LIVE]. These are three different objects: a 2025 actual, a monthly run-rate, and a cumulative dataset. verification/quantitative.md §2.2 is right that 15,000 encounters ≠ 15,000 visits [PRIOR]. Safe form: 120,000 visits in 2025, running at 15,000+ encounters a month by mid-2026, on a cumulative base of 130,000+ visits.

R7 — the AI feature count. The posting names three [PRIOR]. The AI Prescribing Assistant is a fourth [PR-APR26]. Algorithmic availability maximisation is a fifth (§1.6), is the most recent, is not a model, and appears in no Finding and no posting [AI-LIVE]. If the interview turns on "what have you noticed", this is the answer.

R8 — where A5 was right and where it needs correcting. Right: integrations over models; the per-visit compensation argument; the cannibalisation direction. Needs correcting: the cannibalisation magnitude, which assumed the whole follow-up line rather than the non-controlled slice (§1.11), and the treatment of medvidi.ai as simultaneously untrustworthy and authoritative (§1.4) [PRIOR].

---

4. Open Questions / Parked

1. What share of the follow-up book is controlled substances? The single number that sizes the AI Prescribing Assistant (§1.11). Not published anywhere. Settled by: the CEO.
2. What exactly did the AI Prescribing Assistant ship on 8 April 2026? The roadmap said "Est. 2027"; the release says launched. Is it the autonomous renewal of the 2027 vision, or a review queue with a recommendation? Settled by: the CEO, or a clinician demo.
3. Is there a Q-Submission? Unknowable from outside — Q-Subs are confidential until an authorisation issues (§1.9). Settled by: the CEO. Ask what classification, not whether.
4. Has MEDvidi applied to Utah OAIP, or to any state equivalent? Absent from the May 2026 list [UT-OAIP]; applications in flight are not published. Settled by: the CEO, or a later OAIP list.
5. Is there a real B2B pipeline behind partnerships@medvidi.ai? No CRM, no scheduler, no sales headcount (§1.1, §1.14, §6). Settled by: the CEO.
6. Was there ever a one-pager? The button is commented out with href="/", which suggests it never pointed at a file [AI-LIVE]. Settled by: asking for it.
7. When exactly was the site rebuilt? Bracketed to 2026-03-06 → 2026-08-18, most likely ~2026-04-08 (§1.2). Settled by: Netlify deploy history, or a capture nobody has.
8. Why did Valiotti leave, and what shipped under him? Public professional facts only: on the site Sept 2025, quoted Feb 2026, absent Aug 2026 (§1.3d). Out of scope to pursue further. Settled by: the CEO, if he volunteers it.
9. What is the PDMP check's frequency? Per visit, per new patient, per controlled- substance visit? The 23-minute figure scales completely differently under each (M6). Settled by: the CEO. A5 parked the same question [PRIOR].
10. What is "RR1" measured over? 30/60/90 days? Same-provider or any-provider? Determines whether 53% is good or alarming (M8). Settled by: the CEO.
11. Is esacare.com still transacting, and on whose entity? It has its own pricing, its own phone number and a 2020-vintage web server [SIB]. Settled by: the CEO, or a Secretary of State search.
12. Does a Form D exist under a holding-company name? EDGAR is clean for "MEDvidi" [FDA-DB]; a differently-named entity would not surface. Settled by: an entity search once the parent's legal name is known.
13. Is the "AI Clinic Track" the B2B track? The posting names two tracks; §1.6 and §6 suggest AI Clinic is the sellable half. Three incompatible hypotheses already exist in the Inquiry (C10) [PRIOR]. Settled by: the CEO — and it is the question that decides what the job is.

---

5. What this does NOT cover

• Any authenticated surface. Nothing was logged into; no account created; no form submitted; the partnership form was read, never posted to. No credential attempt, no host enumeration.
• Non-public FDA activity. Q-Submissions, pre-submission meetings, breakthrough-device requests and their contents are confidential by design. This Finding reports the public registers and says so explicitly (§1.9).
• Other states' AI-prescribing regimes. Utah is treated here as the only public, named, contractual regime found. Texas, Arizona and others have AI-in-healthcare statutes of a different kind; those were not surveyed. Do not claim Utah is the only jurisdiction with any rule — claim it is the only one with a published, company-named authorisation register.
• Whether the operating metrics are true. They are unverifiable from outside. §1.4 sets the rule; it does not adjudicate.
• Any private-individual facts. Only public professional roles as the company itself publishes them, and only where they bear on the org (§1.3d).
• The consumer funnel, ad creative, pharmacy-routing feasibility, and the org design. Those are other Scouts' gaps in this same pass; this Finding stays off them.
• Whether the B2B thesis is strategically wise. §6 estimates whether it is happening. Whether MEDvidi should arm the competitors who will need this stack is a judgment for the candidate, and it is flagged, not answered.

---

6. What this means for a VP of Product

The verdict, with numbers I will defend

Split the strategic question into three claims, because they have very different answers and collapsing them is how a candidate gets caught.

Claim A — MEDvidi intends to sell its clinical AI stack to third parties. Confidence: ~85%.

Documented, not inferred. A dedicated domain registered ~May 2025 [WHOIS]. A partner-typed form offering Health System / Technology Partner / Provider Group / Investor [AI-LIVE]. A partnerships@medvidi.ai alias. A September 2025 CTA that read "Co-develop or deploy an agent across your network and measure results within weeks" and a "Technology & safety" block promising isolated VPCs with "implementation varies by partner environment" [AI-ARC] — someone had thought about deploying into somebody else's estate. A second named product track in the job posting. And a real multi-tenant configuration model underneath [PRIOR]. The intent is not in question.

Claim B — MEDvidi has a live B2B business with at least one external paying customer today. Confidence: ~10%.

Eleven independent absences, every one checkable:

1. No sales. 65 live requisitions; not one sales, business-development, account-executive, solutions-engineering, partnerships or customer-success role. Zero keyword matches for any of those terms on the whole careers index [JOBS-LIVE]. Companies that intend to sell enterprise software in the next four quarters hire a seller before they hire a VP of Product.
2. No named customer. The "As Seen On" wall is press, and it was disclosed as purchased until the disclosure was deleted [AI-ARC].
3. No case study. The only "customer" quantified is "a single 50-provider team" — which is MEDvidi's own bench [AI-LIVE] [PRIOR].
4. No external testimonial. The one quote, Dr. Henry Bradford MD, is on MEDvidi's own /team/ page [AI-LIVE] [PRIOR].
5. No pricing, no packaging, no tiers [AI-LIVE].
6. No collateral. The one-pager button is commented out [AI-LIVE].
7. No legal surface. Privacy and Terms are href="#" on a page collecting business contact data [AI-LIVE].
8. No pipeline plumbing. Netlify Forms; no CRM, no scheduler [AI-LIVE].
9. No discoverability. medvidi.com never links to it; no sitemap, no robots.txt, /contact 404s [AI-LIVE].
10. No motion. The page did not change for five months [AI-ARC].
11. No proven tenancy. Six configured brands, two and a half live, all self-owned (§1.12).

Claim C — the near-term audience for medvidi.ai is capital rather than customers. Confidence: ~55%.

This is the reading that fits the artefact best, and it is the one no Finding proposed. Consider what actually changed between September 2025 and August 2026: four paid wire releases on four different financial-news networks in eight months, none of them patient acquisition (§1.14); a leadership page with an ARR-and-profitability claim in the CEO's bio; an "Investor" option on the only form; the deletion of a regulatory disclaimer and a paid-PR disclosure; an "FDA pathway" phrase that is unfalsifiable from outside; a partner-deployment section replaced with softer, clinic-internal language; a 2027 roadmap item pulled forward into a launch; and the hiring of a VP of Product over 53 people with an explicit AI track. That is the profile of a founder-owned, cash-generative company constructing an institutional narrative — for a first priced round, a strategic investor, or a sale — rather than one standing up a go-to-market. EDGAR is clean, so no Reg D offering has been noticed under this name [FDA-DB]; that is consistent with "not yet", not with "never".

Why only 55%: the alternative benign reading is equally live — that medvidi.ai is simply a recruiting-and-credibility asset for a company hiring senior product and engineering people in five countries, and the "Investor" option is a form field someone added in ten seconds. That reading explains everything except the four wire releases.

What would confirm or kill each, in public, cheaply

The last row is worth saying out loud: the "not yet FDA-cleared" disclaimer has already been quietly deleted once. Watch what the copy does next; it is the cheapest running signal available.

The argument this changes

The Inquiry's strongest available thesis has been that MEDvidi is building an operating system for regulated prescribing clinics with its own clinic as customer zero. This pass does not overturn that. It re-times it, and re-times it in the candidate's favour.

The OS is half-built and unsold. The building is real: multi-tenant configuration, PDMP inside the EMR, e-prescribing inside the EMR, 100% chart review, an identity-verified intake, a scheduling optimiser. The selling has not started: no seller, no price, no customer, no collateral, no legal surface. And the one genuinely sellable, genuinely differentiated capability — an AI that can safely touch a prescribing decision — is precisely the one that cannot legally be applied to MEDvidi's own revenue core, because every public authorisation regime in the United States excludes controlled substances (§1.10). Meanwhile a competitor has the public instrument MEDvidi does not (§1.10), and MEDvidi has a phrase.

That is a coherent and genuinely interesting position to walk into, and it produces a sharper agenda than "fix the funnel":

1. Decide which business the 53 people are for, out loud, in Q3. Not as a strategy offsite — as a resourcing split with names on it. If the answer is the clinic, the AI Clinic Track should be measured in clinician hours and visit density, and medvidi.ai should be maintained or retired rather than left with dead legal links. If the answer is the platform, then a seller, a price and a design partner precede any further model work, and the first hire is not a PM.
2. Convert "FDA pathway in progress" into a stated classification. Under the 6 January 2026 guidance, a reviewable, guideline-grounded renewal recommendation is very likely non-device CDS and needs no submission — which would be excellent news and should be said confidently rather than hinted at [FDA-CDS]. The change that gets you there is to stop marketing the proprietary corpus as the grounding and start showing the clinician the guideline the recommendation came from. That is a product change, it is small, and it is the criterion-4 fix.
3. Pursue a state regulatory-relief agreement as a product artefact, not a legal chore. A public, named, phased authorisation with an audited agreement rate is worth more to a health-system buyer than any accuracy claim on a landing page, and it is the only thing on this list a competitor already has [UT-OAIP].
4. Reprice the non-controlled renewal before automating it. §1.11 shows the feature destroys $2–6M of visit revenue and saves under $1.5M of clinician cost if it simply removes the visit. The decision is a price, not a model.
5. Spend one afternoon on ezcareclinic.io. A live MEDvidi-branded page promising a personal licensed therapist and 60-minute therapy sessions, in a 20-state list the company does not have, is free regulatory exposure (§1.13).

Five questions this pass earns

Ranked. Each one is unanswerable by a candidate who did not do this work, and none is a gotcha.

1. "Your roadmap had the AI Prescriber at Est. 2027 with 6.67× of the 10×. It launched in April. What did you cut to get there, and is the thing that shipped the 2027 thing?"
2. "Every public AI-prescribing authorisation in the US excludes controlled substances. What share of your follow-up book is Schedule II–V — because that is what sizes the Prescribing Assistant, and I could not find it anywhere."
3. "Your site says FDA pathway in progress. Under the January 2026 CDS guidance I'd expect a reviewable, guideline-grounded recommendation to be non-device — no submission at all. Are we pursuing a classification, or is that phrase doing marketing work?"
4. "RR1 went 41% to 53%, which means roughly half your ADHD patients never come back. Is 3.43 visits a churn number or a days-supply number, and which one does the AI Clinic Track get measured on?"
5. "There are 65 open roles and not one of them sells. If the AI Clinic Track is a product for other clinics, who sells it — and if nobody does yet, is medvidi.ai for partners or for investors?"

Question 5 is the one to hold until late, and to ask plainly. It is the question the whole artefact is built around, and the answer defines the job.