What one year in production produced
Evry Health is a fully-insured commercial large-group plan in Texas. In September 2025 it put its own voice AI platform into production against its own worst operational problem. Twelve months later the platform had placed 250,659 calls, held 112,003 conversations with a live human being, accumulated 7,088 hours of talk time, and analyzed every call that connected against a written compliance rubric.
The problem it was pointed at is unglamorous and expensive. Practices, for a variety of reasons, hold incorrect information about the plans they are contracted with. A practice looking at an Evry card at the front desk often cannot find the plan in its practice-management system and registers the patient as self-pay. The member gets a bill as though they were uninsured. Evry gets the complaint, the reprocessed claim, and the escalation, and pays anyway.
The legacy answer is a vast on-the-ground campaign, with staff sent out on office visits. It is exceedingly expensive, and the problem recurs as soon as the front-desk staff turns over. The alternative is a phone call to every practice in the network. Evry’s directory runs to roughly 117,000 providers across 292,000 directory records and 42,000 distinct practice phone numbers. No plan of Evry’s size staffs that, which is why the work had never been done at scale before.
Monthly volume, September 2025 to the cutoff
Usage is campaign-shaped, not steady-state. The quiet months are the gaps between metro campaigns while the next contact list is built and the agent revised. The last column is the cutoff day alone rather than a month, and the two earliest months ran on a retired call table that did not record whether a person answered.
Six things worth your attention
The calls audited the network as a side effect
13.8% of conversations that reached a person produced a correction to Evry's own provider record. Extrapolated across the year, roughly 15,000 corrections. The most common single finding was that the physician the directory placed at a practice was not there. The campaign was not designed to do this.
Every call is graded against a written rubric, and the record is reviewable
Seventeen criteria, two of them hard compliance floors. Across a 9,000-conversation sample the agent held the patient-information floor on 99.5% of calls and the network-route floor on 97.0%. Every exception carries a call ID and a verbatim quote. That is a different artifact from an assurance that a system is compliant.
The agent that places the call is graded on it, every time
After every connected call a second model, separate from the one that held the conversation, grades the transcript against the campaign's written rubric and records a result, its reasoning and the verbatim evidence for each criterion. That happened on 181,594 calls this year, whether or not anyone was ever going to read them. Where a conventional quality program reviews a 2% to 5% sample after the fact, this is a complete record produced by the system itself.
Most of the difficulty is the phone system, not the conversation
76.4% of calls that reached a person went through an IVR first. The agent escaped 97.7% of them, navigating by tone. Getting to the front desk at all is most of the engineering.
It did the work of four functions out of one program
Dialing, quality assurance, directory verification and reporting all came out of the same program. Staffed conventionally that is 14 full-time people, or $743,000 a year at US outsourced rates, and it had never been run because nobody was going to fund it. In twelve months the platform held a real conversation with 13,482 distinct practice numbers across four Texas metros.
The agent improved 23 times without a compliance re-review each time
Behavior is versioned, tested against an adversarial scenario library, and graded in production against the same rubric before and after every change. That is what let the agent go through 23 distinct production revisions in a year while the compliance floors held above 97%. Rapid iteration and auditability are usually a trade. Here they are not.
Three things the calls did
1. The directory audit nobody ordered
9,000-call sample · Aug 2026
The August campaigns dial a practice, name one or two physicians the directory says work there, and ask the front desk to look Evry up in their system. In 1,241 of 9,000 conversations the practice corrected the record instead. 837 of those were affiliation corrections, 251 were name corrections, and the rest were phone, email and title.
Provider directory accuracy is not a data-hygiene problem for a health plan. It is a network adequacy filing, a surprise-billing exposure, and a member abrasion problem at the same time.
Read them in bulk and the same sentence keeps coming back: "We don't have any of those providers here." One physician had retired and the other had moved. In one case the agent called the number the record gave for a named emergency care center and a different hospital answered.
Evry now has an additional tool supporting its other provider directory integrity systems. Vocalis audits the directory by calling it, with a timestamp and a verbatim quote attached to every correction. The alternative is buying a directory audit from a vendor by the record, and getting no provider education alongside it.
Why it matters. The analysis layer finds value the campaign was not designed to look for.
2. The compliance floor that had to be written twice
17-criterion rubric · Aug 2026
The rubric's PHI floor does not say "the agent did not volunteer patient information." It says the agent must also decline correctly when asked. That clause is there because practices do ask: the analysis evidence has front-desk staff asking for a patient name or account number so they can pull up the call, which is the reasonable thing for them to ask.
The agent has to say no and explain that this is an administrative call about the practice's network status, and it is graded on the explanation. It held on 99.5% of conversations, and every exception carries a call ID and a verbatim quote, so it can be pulled up and reviewed rather than argued about.
The second floor is subtler. Evry reaches different practices by different routes, some direct and some through a rented network. Telling one practice it is reachable by two routes produces a claims error downstream. So the prompt carries a hard instruction, "Never name a second network, never offer alternatives", and the rubric grades whether it held. It held on 97.0%.
Why it matters. Compliance stated as a gradeable criterion, not as a policy document.
3. The refusals nobody said out loud
222 detections · full year
Every analyzed call is screened for do-not-call requests, legal threats and crisis indicators, independent of whatever the campaign was trying to achieve. Detections are written to a separate issues table and mailed to a named person the same day.
187 do-not-call requests were captured over the year, and 76 of them were implicit. Nobody said "do not call me." They said they were at work, thanked the agent, and hung up. A human dialler would very likely have logged those as a soft no and dialed again. Under TCPA the line between a soft no and a revocation is exactly where the liability sits, and 41% of the captured revocations were never stated plainly.
The screen runs on every campaign on the platform, not only provider education, and the implicit rate moves with who is on the other end. Of the 67 requests captured on the provider campaigns, 15% were implicit; on the consumer-facing campaigns running alongside them the implicit share was 60%. The same screen caught 28 crisis indicators over the year, 13 of them at critical severity.
Why it matters. The safety layer catches what a script would not have logged.
An agent that produces its own audit trail
The part that matters to a regulated buyer is not that the calls happen. It is that the system grades its own work on every call, against criteria written in advance, and writes the grade down with the evidence.
The grading is not done by the model that held the conversation. A separate analysis model reads the transcript after the call ends, scores each criterion True or False, and attaches the lines of dialogue and their timestamps that justify the score. The rubric it scores against is authored per campaign, in plain language, before the first call is placed. The standard the agent is held to is a document a compliance officer can read and amend, not a setting.
A conventional quality program reviews 2% to 5% of calls. That is not a methodology choice, it is a staffing limit: a reviewer works through a handful of calls an hour, so the sample is whatever the headcount allows. Everything outside the sample is an inference. Vocalis grades all of them: 181,594 of the 182,915 calls that connected carry a complete analysis, and the balance was still in the queue at the cutoff. For any call in the year, the rubric result, the reasoning and the verbatim evidence can be pulled up on request. That is the difference between assurance and an estimate.
Rubric pass rates, random 9,000 of the graded August conversations
The two criteria in green are marked FLOOR in the rubric: hard compliance failures, not quality targets. The platform graded every conversation; these rates are read from a random 9,000 of those graded records, because that is what the report's authors pulled, not because the platform stopped at 9,000.
The compliance and identification criteria sit at the top, which is where they are supposed to sit. The two lowest bars are the optional close: offering to send material and reciting the billing route. They are not compliance criteria, and they mostly go unspoken because the person on the other end already had what they needed. Because every call is graded rather than sampled, they are visible as a target for the next revision.
Before it ships
Agents run against a simulated-call harness before they reach production: an AI tester playing a scripted counterpart, graded on eight weighted criteria. 48 scenarios are on file, and the list is deliberately adversarial. It reads like a list of the things a compliance officer worries about at night: prompt injection, out-of-scope medical advice, controlled substance requests, identity verification failure, hostile escalation, do-not-call handling. Scenarios are written to be failed. An agent that clears them is one that has already been attacked in a lab rather than in front of a provider.
Caveat. Pass rates are produced by a model grading each transcript against the written rubric, at scale and consistently. They have not been separately validated against human review.
What the platform sustained
A quarter of a million calls in twelve months, without expanding staff. The program ran as a series of metro campaigns rather than a steady dial tone, so volume is lumpy by design: a metro is loaded, worked through over two to three weeks, then the list is rebuilt and the agent revised before the next one opens.
| Measure | Twelve months |
|---|---|
| Calls placed | 250,659 |
| Conversations with a live person | 112,003 |
| Hours of live conversation | 7,088 |
| Hours of connected audio, including IVR and hold | 9,912 |
| Unique numbers dialed | 26,327 |
| Unique numbers eventually reached | 13,482 |
| Calls that connected | 182,915 |
| Connected calls carrying a complete graded analysis | 181,594 |
Throughput
Peak observed day was 31 August 2026: 34,761 calls placed, 18,342 conversations with a live person, and 2,163 hours of connected audio inside a single day. To put 2,163 hours in one day another way: a 270-seat call center, fully staffed and running an eight-hour shift at realistic occupancy, would not have matched it. No queue was formed, no shift was scheduled, and no seat was added. The same program scaled back to a few hundred calls a day the following week without anyone being stood down.
Reach rate by hour, Central time
The 1pm dip is the Texas lunch hour. Reach falls off after 4pm, and 17% of all calls were placed into that falling window. It is the most obvious scheduling improvement available in the dataset.
What the year was worth
The headline is not that Evry cut a cost. It is that Evry ran a program it could not otherwise have run, and got four department functions done out of one program. Sizing it against people is the right way to understand the scale, but the dialing is only the visible part: the platform also graded every call, corrected the provider directory as it went, and produced the reporting on top.
| Component | Hours | Basis |
|---|---|---|
| Live conversation | 7,088 | Measured |
| IVR navigation, hold, voicemail | 2,824 | Measured: total audio less live conversation |
| Attempts that reached nobody | 2,311 | Estimate: 138,656 attempts at 1.0 min |
| After-call work | 3,733 | Estimate: 112,003 conversations at 2.0 min |
| Quality assurance | 933 | Estimate: a 5% review sample at 6 calls reviewed per hour |
| Directory verification and correction | 3,000 | Estimate: 15,000 corrections at 12 min to confirm and update |
| Supervision and team overhead | 2,983 | Estimate: 15% of front-line hours |
| Reporting and analytics | 850 | Estimate: half an analytics engineer to build and run what the platform emits |
| Total | 23,722 | 14.0 FTE at 1,700 productive hours |
Quality assurance
The 933 hours above fund a 5% review sample, which is roughly what a regulated outbound operation actually staffs. Vocalis graded 100% of connected calls against the same 17-criterion rubric, twenty times the coverage, and it is included rather than staffed. Matching that with human reviewers would take about 18,700 hours on its own, or eleven full-time analysts. That figure is deliberately left out of the table, because no operation would ever fund it.
Directory verification
The 15,000 corrections are counted above as internal labor. As purchased data they are worth more: phone-verified provider records run roughly $4 to $8 each in the vendor market, which puts the same output at $60,000 to $120,000. Either basis is a real line item Evry did not have to fund at all.
| Basis | Loaded rate | Labor | Reporting | Seats and tooling | Annual value |
|---|---|---|---|---|---|
| Offshore BPO | $14/hr | $320,000 | $38,000 | $23,000 | $381,000 |
| US outsourced provider relations | $28/hr | $640,000 | $72,000 | $31,000 | $743,000 |
| In-house specialist | $42/hr | $961,000 | $81,000 | $39,000 | $1,081,000 |
On the central basis the program delivered $743,000 of staffing-equivalent value in its first year, roughly $6.60 for every conversation that reached a person, and it did so without a single seat being added. Run in-house at specialist rates the same work passes a million dollars. That figure still understates the platform twice over: a staffed operation at any of those rates reviews 5% of its calls where Vocalis reviews all of them, and it produces no directory corrections at all.
How to read the table
Evry did not remove fourteen people from a payroll. The program had never been staffed, so there was nobody to remove. Evry does not use AI to cut people. The principle is the opposite one: keep licensed and experienced staff working at the top of their license and their skill. Nurses, provider relations managers and claims specialists are expensive precisely because of what only they can do. Dialing a list of 26,000 practice numbers to read the same paragraph about network status is not that work. The table is the capacity the platform created, and the value is that it created it without pulling a single clinician or specialist away from the work that needs them.
The next measurement
The premise of the program is that reaching a practice prevents a self-pay misregistration downstream. Linking a provider-education call to the claims and complaints that follow it is the next piece of instrumentation, and it is the measurement that will put a hard number on the avoided cost.
Sizing it for your own network
The variable that decides everything is reach. Across the year 51.2% of unique numbers eventually reached a live person, and 57% were settled within two attempts. Size it by provider record, not by phone number: Evry’s list carries one row per provider, so a hospital switchboard listed against several hundred physicians is dialed once for each of them. That is why the mean sits at 9.4 attempts per number while the median is 2.
What it took to run
The campaign is not the interesting artifact. The agent is. The agent that ran the August campaigns reached version 23 in a year, on a prompt that grew past 35,000 characters. It has three tools: play DTMF tones, leave a voicemail, hang up. Everything else is conversation. Across all agents, 568 versions are on file, and the most-iterated agents have run through 40 to 50 revisions each.
Most of that length is not instruction. It is the accumulated enumeration of specific failures observed in production and written down so they stop happening, each one traceable to a call that went wrong. Behavior that a demo never surfaces is what fills the document. Specificity of that kind is what separates a system that survives 100,000 production calls from one that presents well.
Sentiment across 20,000 sampled conversations
For unsolicited outbound calls into busy front desks, a two percent negative rate is the number worth noticing.
What the practices did
34.5% of conversations ended with the agent making a commitment. In the August sample, 44% of those were information to send, 32% documents, 12% callbacks and 9% general follow-ups. Every commitment is logged with its text, which means the follow-through is auditable. 16.0% asked to speak to a human and 10.2% asked for a callback; both route to a queue rather than being refused, which is why the agent’s transfer behavior is one of the graded criteria.
The retry tail is structural
10,258 numbers were settled after one attempt and 4,659 after two. At the other end, 2,671 shared numbers absorbed 72% of all calls, because the contact list is per provider and a hospital main line listed against hundreds of physicians is dialed once per physician. Collapsing those to one call per number is the single largest efficiency available in the program.
What organizations could benefit
Any health plan or TPA with a contracted network
Educating and managing a provider network is a constant, recurring challenge, and the self-pay misregistration problem it produces is structural rather than local. Nothing about the workflow is specific to Texas or to commercial large group.
Any organization with a call or customer support center
This case study covers the provider network work at Evry Health. The same platform is separately applied there to Tier 1 customer support.
Any organization carrying a directory it cannot verify
The mechanism that found 15,000 corrections does not care that the records were physicians. It calls a list, asks a question, and writes down what came back with evidence attached.
Regulated outbound of any kind
The parts that took the longest to build are the parts a regulated buyer needs and cannot skip: gradeable compliance floors, implicit do-not-call detection, crisis screening, and a record of the failures rather than an assurance there were none.
Anyone currently sampling calls for quality
The economics of listening to 2% of calls were dictated by supervisor time, not by what good assurance requires. When grading every call is included rather than staffed, the sampling decision stops being a constraint.
How the numbers were produced
Source. The Vocalis production database, read directly: 250,659 call records across the current and retired call tables, 181,594 completed post-call analyzes, and the campaign, agent, evaluation and issue tables behind them.
Census versus sample. Call counts, durations, reach rates, unique numbers and retry distributions are full-census figures across every call placed, and so is the grading itself. The pass rates, sentiment and commitment figures are read from random samples of those graded records: 9,000 August conversations for the rubric and commitment breakdown, 20,000 for sentiment.
Measured versus estimated. Everything in the program totals, the volume charts, the rubric and the reach rates is measured. The human-equivalent workload table is an estimate, labeled line by line; loaded labor rates, the 5% quality-assurance baseline and the half-FTE reporting allowance are market assumptions, not Evry figures. The 15,000 directory corrections figure applies the observed August correction rate of 13.8% to the 112,003 conversations that reached a live person across the year. No figure in this report is a survey response.
Redaction. The underlying records contain provider names, practice names, national provider identifiers and verbatim call quotes. None of it appears here. The quoted call evidence is reproduced only where it identifies no one. Results were measured by Evry Health and Pendral from their own production records and have not been audited or independently verified by a third party.