Rivvi

Every part of what Rivvi sells is model work. The workspace reads an uploaded roster or a 292,692 row payroll file, answers questions about it in plain language, drafts reports and remembers how a team writes; routines replay that work on a schedule; and the outreach agent holds phone calls, two way texts and website chats with patients in more than 30 languages.

The pieces around the models, connectors to Microsoft 365, Google and HubSpot, TRACE approval controls, credits and audit logs, exist to put that model work to use safely and have no value on their own. Remove the models and there is no workspace, no memory and no agent left to buy.

TRACE, Rivvi's tiered review layer, sets how far the AI may act. It runs on the server on every call and assigns each tool one of four levels: auto run, draft for a person to approve, propose only so a person makes the change, or never run. An organization can tighten a level but nothing can loosen one, an approval binds to the exact action approved, and writes to connected systems can require a signed approval.

On the patient side the agent does not diagnose, prescribe or give medical advice, hands off when a question needs a clinician, escalates crisis signals immediately, and transfers to a person on request. What counts as a crisis signal or a question needing a clinician is not listed, and which tools default to auto run is not stated.

The design around the models is clear: answers come from what a team has uploaded and connected rather than a blank prompt, memory of providers, sites and payers is built from the organization's website and public registries and can be viewed, corrected or removed, and routines turn repeated requests into scheduled jobs. TRACE sits between the model and every tool, so the model proposes an action and the server decides whether it runs.

Which language and voice models do the work and who supplies them are not stated, and there is no changelog or notice telling a team when an underlying model was swapped. Premium seats buy five times the assistant allowance, which describes capacity rather than a different model.

AWS is named as the host, with data in us-east-2 under KMS encryption, and HubSpot, Microsoft and Google appear as connected systems the customer chooses. Rivvi's trust center releases its subprocessor list only under a nondisclosure agreement, so no subprocessor is named in public.

Whoever supplies the language and voice models goes unnamed, as does any telephony provider placing calls and texts, along with the terms that govern data once it reaches them. The commitment that PHI never trains third party models implies outside models are in the chain, without saying whose.

Platform figures as of August 2026 describe scale: more than 2 million interactions, more than 500,000 patients reached and a live answer rate above 40 percent on outreach, across organizations in eight countries. One outcome is attributed to a named person: Jennifer Saddanathan, PharmD, a director of care management, says medication adherence rose 2 to 3 points within three months, without her organization, the measure or the baseline being given.

None of the figures comes with a method or denominator, the live answer rate measures reach rather than benefit, and nothing evaluates the workspace's answers against the source files or the agent's calls against a reference.

Training is ruled out in two phrasings: customer data never trains anyone else's model, and PHI never trains third party models. Neither sentence says whether Rivvi's own models learn from it. PHI stays inside HIPAA eligible infrastructure, k anonymity is on by default, and the audit trail records who touched PHI and every agent action without storing prompts or responses. How long conversations, call recordings and uploaded files are kept is not stated.

Safety engineering for the agent is spelled out: it stays inside the job it is given, gives no medical advice, escalates distress and crisis signals at once to a named person, and live edits to an outreach program cannot strip disclosure or TCPA rules. How a safety event is reviewed afterward is not described.

The tier question that trips up most buyers is answered outright: the free organization plan for up to three people and every paid plan include a HIPAA business associate agreement, while the free personal plan is limited to public information. The security page and pricing page both repeat that the agreement comes included, with no enterprise sales cycle needed to get it.

The trust center lists the agreement among its documents and releases it on request, so its terms on subcontractors cannot be read before contracting. It also states that patient data is processed only to deliver contracted services under the agreement. A privacy policy sits under the legal section, and the security page states that PHI stays inside HIPAA eligible infrastructure and is scoped to each organization.

Rivvi presents SOC 2 as its own. Its trust center lists Type I and Type II reports dated 2026, describes the audit as covering security, availability and confidentiality, and adds a penetration test report from March 2026. The reports, the test results and the subprocessor list are released under a nondisclosure agreement through the security contact, and the audit period and auditor are not shown.

Controls described on the security page include enforced multi factor authentication per organization, role based permissions, KMS encryption at rest, server side checks that fail closed when a tool and user do not match PHI clearance, signed approvals for connector writes, and audit logs of every agent action touching PHI. The trust center adds single sign on through SAML 2.0 or OIDC and round the clock monitoring, and says no formal bug bounty runs yet.

The intended use is stated as a limit on the agent: it does not diagnose, prescribe or override clinical judgment, gives no medical advice, stays inside the job it is given and hands off to a clinician when a question calls for one. The marketed uses match that scope. Care gap outreach, medication adherence and refill follow up, calls after discharge, inbound call routing, trial pre screening and team data analysis are administrative and engagement work rather than clinical decisions.

No device position is taken beyond that statement, and none of the marketed work appears to need one. TCPA consent and do not call rules run on every outreach path, a separate regulatory regime the product addresses directly.

Language coverage is sold as a strength, with calls in Spanish and more than 30 other languages for patient outreach and front desk lines, and no results are given for any of them. A caller the agent misunderstands in a less common language leaves no complaint behind, and nothing shows how recognition or answer quality varies by language or accent.

Governance machinery does exist. Every outbound program must pass safety evaluations before it can run, TRACE enforces approval levels on the server, a white paper describes that control layer, and consent and policy changes are logged for seven years. The evaluations themselves, their pass criteria and their results are not released.

Controls make Rivvi's actions traceable: TRACE logs every call with who, what and when, an approval binds to exactly what runs, consent and policy changes are kept for seven years, and outreach results are written to HubSpot where staff can review them. A person is one transfer away on every call.

Nothing commits Rivvi to the correctness of what the workspace answers or what the agent says. No error rate is given for answers drawn from uploaded files or for call outcomes, and no guarantee, indemnity or remediation tied to output appears on the site. The audit log deliberately excludes prompts and responses, which protects privacy and leaves the content of a disputed answer outside the record.

No EHR connection exists, and Rivvi says so: nothing writes to the EHR. Patient lists reach it as uploaded files, such as a roster, a discharge list or a payer gap report, and outreach results land in HubSpot as notes, tasks and call logs.

Named connectors cover Microsoft 365 (Power BI, SharePoint, OneDrive, Excel and Outlook calendar), Google Drive, Sheets and Analytics, HubSpot and LeadingReach, a referral management system, with Salesforce listed but not yet self serve. Writes to those systems can require approval. A health system using Rivvi for follow up after discharge would move lists in and results out by file or CRM rather than through the chart.

Data rests in AWS us-east-2, Ohio, encrypted with AWS KMS, and PHI is said never to leave HIPAA eligible infrastructure. Each organization's data is scoped to it, with every read and write checked against that scope, and sharing defaults to the individual user.

Rivvi is a single hosted service with no private or regional deployment option. Customers in eight countries appear to share the Ohio region, and GDPR is acknowledged for people in the European Economic Area without a statement of how their data crosses the border. Where model and voice processing happens, inside the same boundary or with an outside provider's service, is implied rather than spelled out.

Prices are public and self serve. Free covers up to three people with workspace chat, memory, Microsoft 365, Google and Teams connections, 500MB of data and a business associate agreement, with no card needed. Team is $79 a seat a month, or $63 billed annually, with a three seat minimum and calls, texts and website chat; Business is $149 a seat with a ten seat minimum, audit log and admin tools; Premium seats with five times the assistant allowance run $199 on Team and $299 on Business.

Patient conversations draw on a shared credit balance: 40 credits a minute of calling, 4 a text and 10 a website chat, with blocks from $100 for 10,000 credits down to $5,000 for 1,000,000. Enterprise adds committed volume at a negotiated rate.

Settings are named in two groups. Self serve plans target value based care organizations, community pharmacies, Medicare brokers and FMOs, and specialist practices; enterprise covers health systems, payers and PBMs, life sciences patient support and trial recruitment, and government programs. Jobs include care gap closure, Star Ratings and HEDIS work, medication adherence, follow up after discharge, inbound calls and plan comparison for brokers.

Rivvi runs in eight countries, chats in the language a user writes in and speaks Spanish and more than 30 other languages on calls. The adherence result from a care management director is the only outcome tied to a setting, so most of the range is stated rather than shown.