Evergrove
Evergrove sells voice agents for workers' compensation, built to run the phone work that moves an injured worker's claim from first report to paid bill. Nineteen agents cover five workflows: intake (four agents) opens the case and sets up every downstream party, clinical coordination (four) handles referrals and evaluations, status (six) chases updates, billing (three) follows up on bills, and intel (two) gathers information. Inbound, the agents sit on the customer's phone system, triage the reason for the call, route it and hand off to a live representative on completion or escalation, with a callback if the line drops. Outbound, they run in parallel, bundle calls about the same injured worker, navigate phone menus, wait through holds and transfers, and time calls to a facility's working hours and past activity.
The agents read and write case details in claims management systems, Salesforce or a homegrown system of record, and an immutable log records every action, call, decision and escalation. Customers set approval rules, escalation rules and intervention thresholds per workflow. Buyers are managed care organizations, third party administrators, specialist provider networks, ancillary providers, facilities and investigation firms.
Nony AI Inc. operates Evergrove from New York. Sid Unnithan and Jacob Chia, who built AI insurance placement at Newfront, founded it in 2026, and it is in Y Combinator's Summer 2026 batch. The company reports 500,000+ calls handled in production, three weeks to integrate, 40% higher workflow completion and 60% lower operating spend.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
The agents are the product. Workers' compensation organizations buy them to place and take calls, track where each call stands and update the case afterward, and the rest of the platform, from the audit log to the claims system connectors, exists to support those conversations. The work splits into 19 agents across intake, clinical coordination, status, billing and intel workflows.
Outbound agents navigate phone menus, hold through transfers, retry with backoff and bundle several questions about one injured worker into a single call. Inbound agents identify why someone is calling and either finish the task or hand off to a live representative. Remove the voice models and nothing remains to sell, since the pitch is one seat turned into many parallel agents at the same cost.
Each workflow carries its own approval rules, escalation rules, intervention thresholds and permissions, set by the customer. The sample case log shows the structure working: a scheduling action is routed to an adjuster, an approval event follows, and the agent then confirms the appointment. Inbound calls hand off to a live representative on completion or when escalation is needed, and a dropped call triggers a callback.
Outbound agents retry, wait through transfers and continue on their own under what the company calls call supervision. The control points are real and named. What the agents may complete without a person and what always needs approval are left to each deployment's configuration, with no default rule set or threshold stated for a new customer.
The platform assigns calls to specialized agents for intake, clinical coordination, status, billing and information gathering, 19 in all. Voice agents run in parallel, use prior call activity and facility hours to decide when to call back, back off between retries and record each resulting action in the case history. Inbound calls are triaged by purpose and routed to a workflow or a person. That explains what the agents do inside a claim.
What produces the conversation is not described: no speech recognition, voice synthesis or language model provider is named, nothing says whether the models are adapted to workers' compensation terminology, and there is no version history showing when an agent's behavior changed.
The voice service connects to the customer's telephony and moves claim information between the call and systems such as Salesforce or a claims platform. The privacy notice says an SMS provider receives a phone number to send messages, without naming it, and that notice covers messaging rather than voice calls. Claim level access controls govern who sees what inside the platform. The rest of the chain is open.
No speech, language model or telephony provider is named, nothing states which party receives call audio and transcripts, and no retention or training terms are set for the parties that process a call about an injured worker's treatment.
The company reports more than 500,000 calls handled in production, three weeks on average to integrate, a 40% increase in workflow completion, a tenfold drop in workflow cycle time and more than 60% lower operating spend. Its Y Combinator profile frames the call figure as an annual pace and credits the gains to large managed care organizations and specialty physical therapy networks, none of them named.
No figure carries a denominator, a baseline, a measurement window or a client behind it, and the results are stated as averages across customers. The founders' earlier work at Newfront, an AI placement service they say cut renewal times by 40%, belongs to a different product and does not carry over.
Access is scoped at the claim level, and every action is tied in the immutable log to the staff member or agent responsible. The enterprise service describes encryption in transit and at rest, role based access and configurable retention, and the sample log separates access to prior imaging from the later notification and appointment confirmation. The public privacy notice covers phone numbers used for SMS and how to stop messages.
Nothing addresses the questions a voice agent raises: how long call recordings and transcripts are kept, whether medical details in a call reach a model provider, or whether call content trains any model. No process for an AI safety incident, such as a wrong appointment sent to an injured worker, is set out.
Evergrove calls its platform HIPAA compliant and says business associate agreements are available. Its voice workflows handle appointment, referral, imaging and claim details, and the sample case log shows restricted access to protected information. The only privacy document is the SMS notice, which covers phone numbers used for text messages and the option to opt out, and it sets no terms for how the voice product handles protected health information.
The agreement itself, the deployments it covers and the obligations it passes to subcontractors in the call processing chain are available only through a sales conversation. Compliance is claimed, with no document behind it that reaches the voice service.
Controls are built into the product: an immutable audit history of every action, communication, decision and escalation, permissions and approval points per workflow, encryption in transit and at rest, role based access and configurable retention. The sample case log shows who can view a record before a scheduling step goes to a human approver. All of it rests on the company's own description.
There is no SOC 2 report, ISO/IEC 27001 certificate, penetration test or trust center, and the security section of the site carries no outside check. The company listed a team of two on Y Combinator, and its security program appears only as the set of controls above.
The software handles communication around an injured worker's claim: referral intake, arranging an evaluation, confirming appointments, dispatch, retrieving notes and calling for updates. Inbound triage classifies the purpose of a call and routes it to a workflow or a person, so the triage concerns the reason for the call rather than the injury. In the sample case a human adjuster approves the treatment plan, and the voice agent handles messaging and scheduling afterward. That is administrative and care coordination work. No FDA device claim is made, and nothing suggests the agents diagnose an injury, choose a treatment or decide whether care is needed.
Customers decide which agents may take actions and where an approval or a human handoff sits in each workflow, and the case trail records what happened and who acted, including a scheduling escalation in the sample. Those controls let a case owner trace a decision after the fact. Evaluation of the voice agents themselves is missing.
The completion and cost figures carry no breakdown by caller language, accent, hearing or speech impairment, or by whether the caller is an injured worker, an adjuster or a clinic's front desk. There is no protocol for testing those differences, or for checking that escalation fires consistently for the people calling about a claim, many of them injured and speaking to an unfamiliar system.
The immutable case log captures every agent action, communication, decision and escalation, and the platform can make an agent transfer a call or pause a step for a person to approve. A representative can review what happened and step in when a case falls outside routine handling. Nothing stands behind the output itself.
The only published terms are for SMS, covering message support and the carrier's responsibility for delayed or undelivered texts, and they say nothing about a wrong appointment, a mistaken claim update or an action taken on the wrong case. No warranty, correction commitment or remedy is offered to the customer or to the injured worker on the other end of the call.
The agents connect to the customer's telephony and read or write case details in claims management systems, Salesforce or a homegrown system of record. Inbound calls hand off to an existing representative, outbound workflows collect status information and notes for the case, and integration averages three weeks. Salesforce is the one platform named.
On the clinical side the agents call provider facilities rather than reading their systems: no EHR is named, no chart access route is described, and no claims system vendor is identified by name either. The integration evidence covers claims and contact center operations, with systems of record described by type rather than by product.
Deployment means linking the agents to a customer's telephone system and system of record, one workflow at a time or across the whole contact center, in about three weeks. Permissions and retention are configurable, and data is encrypted at rest and in transit. Where the service runs is not stated.
Nothing names a hosting provider or region for call processing, recordings, transcripts or backups, the New York headquarters says nothing about infrastructure, and there is no statement on tenant isolation between the managed care organizations, administrators and provider networks that may share the platform, or on any option to run it in a customer's own environment.
Sales start with a demo booked from the homepage, and the founders take inquiries directly. A customer can begin with one workflow or deploy agents across the contact center, and the pitch is economic: one seat becomes many parallel agents at the same cost, replacing contact center staff for a fraction of the spend. That frames the value without giving a price.
Nothing states whether charges run per minute, per call, per case, per agent or as a subscription, and there is no public rate, minimum commitment or setup charge. The SMS terms mention possible carrier charges, which are separate from what the voice service costs.
Workflows cover the communication between injured workers, managed care organizations, third party administrators and the providers who treat those workers: the first report of injury, getting a referral to the right place, scheduling an evaluation, dispatch, confirmations, note retrieval and following up on bills or case status. Specialist physical therapy networks, ancillary service providers, provider facilities and investigation firms are part of the audience alongside payers.
The setting is narrow and stated clearly: workers' compensation in the United States. The reported figures pool every deployment, with no separate result for intake against billing or clinical coordination, and nothing compares performance between therapy networks and the insurer side of the same workflow.
Alternatives to Evergrove
Other products that do the same kind of work as Evergrove, and what each one does.
Wisedocs turns claim files into sorted, deduplicated medical chronologies, summaries and insights for the people who decide claims: insurance carriers, third party administrators, government claims programs, independent medical evaluators, claims adjusters and defense lawyers.
Tennr automates inbound document and referral operations for the organizations that receive them.
Synthpop sells AI agents for healthcare administrative work, now led by durable medical equipment and home care.
Voice AI agent platform for the calls that stand between a patient and a prescribed treatment.
Arini provides an AI receptionist for dental offices and dental service organizations. It answers inbound calls, checks schedules and books against the appointment types, providers and slots a practice has set up.
Qventus automates hospital operations.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Custom quote
|
Sold through a demo and deployment discussion, one workflow at a time or across the contact center. The unit of charge and any commitment are set in the agreement. | Business associate agreements are available. Which deployments they cover, and their terms, are settled in contracting. | Integration takes three weeks on average. No onboarding or integration fee is listed. | Vendor Published |
The deployment links to the customer's telephony and systems of record. There is no list price, call allowance, usage rate, minimum spend or per workflow charge. Carrier charges under the SMS terms are separate from the price of the voice agents.