Sample read

One outbound Medicare Advantage script,
read against the record.

This is the shape of what comes back when a script is sent to Comprenda. Nine agent lines. For each one: what the record knows, on which people and channel, where the evidence stops, what would resolve it, and the shortest next step. Then a ranking of what to test first, and a plain statement of what this read did not do.

Illustrative script

The script below was written for this page. It belongs to no client, no carrier and no call floor, and it has never been dialled. Every finding it is read against was measured on our corpus of senior calls: outbound calls in a senior pharmacy programme and neighbouring senior verticals, with an AI screening stage and a human licensed-agent stage, graded on settled enrolment or on a billed transfer. None of it was measured on a Medicare Advantage script. That is the first thing the read says, and it is why several verdicts below are “not enough evidence” rather than a number.

The script as sent

Nine agent lines, two caller turns for context.

A1“Hi, this is Dana on a recorded line from a licensed Medicare Advantage helpline. Am I speaking with Mr. Alvarez?”
A2“I’m calling because Medicare Advantage plans in your area change for next year, and in about two minutes I can check whether a plan covers your doctors and your prescriptions.”
A3“Do you currently have Medicare Parts A and B, and which company is your current plan with?”
A4“There’s no cost to check, and most people in your area pay nothing extra for the plan itself.”
Caller“I already have a plan and I’m fine with it.”
A5“I understand. Since you already have a plan, a quick comparison is exactly what this is for. What do you like most about your current plan?”
Caller“Sorry, what was that?”
A6“Of course. Can you hear me all right? I’ll slow down a little. I asked what you like most about your current plan.”
A7“Great. I’m going to bring a licensed agent onto the line now. His name is Marcus and he’ll pick up where we left off. Please hold for one moment.”
A8“Mr. Alvarez, this is Marcus. Dana told me you take three medications and see Dr. Patel. I’ve checked both against the plan. They’re covered.”
A9“So we’ve confirmed your doctors and your medications. Shall I go ahead and complete the enrollment with you now?”

Every line below is read on its own. The verdict on a line is deploy, do not deploy, or not enough evidence. A verdict is about the element, on the population named beside it. It is not a forecast of this script.

The read

What the record says about each element.

A1 · greeting with disclosureDeploy

“Hi, this is Dana on a recorded line from a licensed Medicare Advantage helpline. Am I speaking with Mr. Alvarez?”

What we know
The first agent turn is where the message most often fails to get through. 11.3 percent of seniors’ replies to the agent’s first turn show the greeting did not land, against 1 to 2 percent at every later turn. The bare “Hello?” a senior says on pickup is a habit that repeats across contacts (kappa +0.375 within a week, +0.392 at 8 to 30 days) and it carries no money: on the AI screening stage, calls with it billed at 49.8% against 50.1% without. Recording or AI disclosure is spoken on 0 to 8% of calls across 27 live senior agents, on 675 calls read.
Why we think that
Our corpus of senior calls. The 11.3 percent is an instrument read (a language-model annotator, kappa +0.688 against a blind second read, not a human coder). The “Hello?” habit: 36,492 outbound screening calls, same phone across contacts, billed transfer as the outcome. The disclosure count: agent-side text only, no audio, no outcome.
What we don’t know
Whether the wording of a disclosure changes the share of first turns that land. No assigned test of greeting wording exists in the record. The 0 to 8% is what the transcript shows, and a missed disclosure can be an engine miss.
What would resolve it
Two greeting wordings randomised at the first turn, graded first on whether the caller’s reply is an answer or a repair, then on settled enrolment.
Next action
Keep the disclosure; that is a compliance question, not an evidence one. Put the name and the question in the second sentence, after the line has settled. Expect about one in nine first replies to be a repair.

A2 · purpose lineNot enough evidence

“I’m calling because Medicare Advantage plans in your area change for next year, and in about two minutes I can check whether a plan covers your doctors and your prescriptions.”

What we know
On a neighbouring senior book, a screening package built from a no-cost line, the agent’s identity, a callback offer and qualify-then-transfer beat an assume-the-transfer family on billed transfers per engaged caller, RR 1.37 to 1.50 inbound and 1.22 to 1.67 outbound, over 161 windows with 9 decisive, while the assume family still won raw transfers. That result is about the package; no single sentence in it has been isolated. A doctor line of the kind in A2 runs on 0 of 27 live senior agents read.
Why we think that
Outbound senior pharmacy screening, AI stage, billed transfer as the outcome; a found split across script versions the operator ran, replicated across windows, not assigned by us. The doctor-line count is agent-side text on 675 calls.
What we don’t know
Unknown. The record does not cover this element on this population. There is no Medicare Advantage purpose line joined to a settled outcome in the record, and the Medicare-vertical outcome evidence that does exist is sign-only and from a campaign no longer running. Whether “plans change for next year” reads as help or as pressure to a 78-year-old is not something the record can say.
What would resolve it
A two-arm purpose line (plan-change framing against doctor-and-prescription framing) randomised at the screening stage on the buyer’s own traffic. The transfer read arrives in days; the settled read needs the enrolment to mature.
Next action
Keep the doctor-and-prescription half, which is the part the neighbouring evidence supports as a package. Treat the plan-change half as the thing to test, not the thing to assume.

A3 · qualifying questionDeploy

“Do you currently have Medicare Parts A and B, and which company is your current plan with?”

What we know
A senior who names an insurance carrier early is further along, and the marker survives the usual artefacts. On a fixed cohort of 4,311 calls that all ran 30 turns or more, calls where the senior named a carrier by turn 10 settled at 61.5% against 48.0%, and the gap holds inside all five eventual-duration quintiles. It is not a closer behaviour: across 12 closers the share who elicit it does not correlate with settling (rho 0.0). On live senior agents the first qualifying question lands at 17 to 24 seconds. Confirming with the caller before connecting them is settled on transfers.
Why we think that
Human licensed-agent stage of a senior pharmacy programme, settled enrolment, closer held out; the carrier marker is observational and duration-controlled. Timing: agent-side text on 675 calls, 27 agents. Confirm-before-connecting: transfers on the screening stage; its settled leg is unresolved on the corrected label.
What we don’t know
Whether asking for the carrier earlier causes anything. The marker tells you where the conversation is; it is not shown to be a lever the agent pulls. On the platforms read, no field carries the caller’s answer to the licensed agent, so today the answer is collected and dropped.
What would resolve it
An insurance question for every qualified caller, collect-only, registered as an arm, with the answer passed to the licensed agent. Registered on a neighbouring book; not yet live.
Next action
Deploy the question. It is cheap and it is information. Do not deploy any claim that asking it lifts enrolment, and make sure the answer reaches the person in A8.

A4 · cost lineDeploy the framingNot enough evidence on the claim

“There’s no cost to check, and most people in your area pay nothing extra for the plan itself.”

What we know
Cost talk from the senior is engagement, not refusal. Calls where the senior raised a cost objection settled at 0.286 [0.196, 0.381] against 0.260 without; humour settled at 0.453 [0.358, 0.556] against 0.242 (1,085 calls). On the AI screening stage a senior asking about cost transferred 73.6% of the time [70.6, 76.4] on 891 such calls, against a book base of 51.6%. On 8,084 opening calls cost was raised by 2.3% of customers and its effect did not separate from zero [−9.8 to +2.8]. The agent’s side runs the other way: leading with cost early goes with fewer settles, RR 0.80 [0.70, 0.90] on the 7% of calls where it happens, while a no-cost line inside the benefit sentence goes with more; high-retention closers reassure on cost far more often (13% against 2% of closers).
Why we think that
Senior pharmacy programme, both stages, settled enrolment and billed transfer; every row observational and read by an instrument, not a human coder. The engagement states are redundant with each other, so two of them on one call are not two lifts.
What we don’t know
Whether the agent’s no-cost line causes the difference or marks the agents who use it. And for the second half of A4: Unknown. The record does not cover this element on this population. No premium claim about Medicare Advantage plans has been read against an outcome, and whether “most people pay nothing extra” is sayable at all is a compliance question the record does not answer.
What would resolve it
Cost-before-the-ask against a no-cost line inside the benefit sentence, randomised at the screening stage. The billed read arrives in days on ordinary volume; the settled read follows.
Next action
Deploy “no cost to check” inside the benefit sentence and never as the opener. Send the premium claim to compliance, not to the record. Treat a caller who asks what it costs as the strongest thing that can happen on the line.

A5 · the “already have a plan” branchDeploy the questionNot enough evidence on the claim

“I understand. Since you already have a plan, a quick comparison is exactly what this is for. What do you like most about your current plan?”

What we know
On a senior pharmacy book the analogous moment, “I already have a pharmacy,” is the hardest one in the call: callers in that state reach five minutes 14.0% of the time and settle 7.1%, against 52.3% and 39.2% for callers who arrive ready, on 9,283 labelled rows. What the agent does next moves with money. Asking an open question at that moment carries a duration-controlled difference of +0.054 [+0.015, +0.085] in settled enrolment, one of seven moves that survived independence, duration and matched-placebo guards out of 141 candidates. Agents already ask a question 39.4% of the time at this state, the highest of the four states studied.
Why we think that
Human licensed-agent stage, settled enrolment, 677 decision moments, observational; the agent’s move was not assigned. The state detector behind it was hand-audited at 40% precision (20 of 50), so the association sits on a partly mis-detected moment and its prevalence will fall after repair.
What we don’t know
Whether “I already have a plan” in Medicare Advantage behaves like “I already have a pharmacy.” The analogue is not in the record. “Exactly what this is for” is a claim no row supports or contradicts.
What would resolve it
A passport for the state on the buyer’s own calls first (definition, positive and negative examples, a hand audit at 85% precision or better), then an open question against the current mix, randomised at the moment.
Next action
Deploy the open question as the default move at this moment; it is the move the record favours and it costs nothing. Cut “exactly what this is for” or test it, but do not count on it.

A6 · the hearing-repair momentDo not deploy a hearing check

“Of course. Can you hear me all right? I’ll slow down a little. I asked what you like most about your current plan.”

What we know
The passive hearing screen from repair behaviour does not survive. 85.0% of what the first instrument counted as a request to repeat on the AI screening stage was “Hello?” on pickup; it fired on 15.1% of 36,492 calls, and stripped of the greeting it fires on 1.2%. On the human stage, repeat requests are flat in age, 5.5% / 5.6% / 6.8% at 50 to 64 / 65 to 79 / 80 and over, and not stable within a person across contacts (kappa 0.00 to 0.12). The one age-graded signal is the senior saying so: explicit hearing disclosure runs 0.43% / 0.90% / 1.94% across the same bands, 1.3% of calls in all. On a blind second read of 288 calls, an explicit disclosure was a real hearing signal 83% of the time; a bare repeat request was a coin flip, 45 to 50%. No live senior agent runs a hearing check: 0 of 60, across 2,000,718 engaged sessions. And when human agents do accommodate after a hearing signal, settlement is lower, 26.9% against 35.1% on 52 of 334 such calls, because they accommodate when hearing is worst; that number cannot be read causally.
Why we think that
Both stages of a senior pharmacy programme; 20,063 human-stage calls exposure-matched at 5 to 15 senior turns; two language-model readers at reader-to-reader kappa 0.92 and 0.97; no human-coded row. The agent count is a count of configurations and does not depend on the instrument. No rate of hearing difficulty may be stated from any of it, and none is stated here.
What we don’t know
Whether an assigned repair (slower, shorter, verbatim) at the moment of a repeat request changes anything settled. The record has no such arm. On the AI screening stage, of 15 repeat requests in 675 calls the agent simply moved on in 11, so the record also cannot say what a good repair looks like there.
What would resolve it
One clean test: at a detected repeat request, randomise the agent’s next turn between a verbatim repeat, a shorter rephrase and the current behaviour, graded on whether the next caller turn is an answer and then on settled enrolment. Nothing else in the record substitutes for it.
Next action
Do not deploy “Can you hear me all right?” as a screen, a routing step or a scored behaviour. At a bare repeat request, repeat the sentence and carry on. When a caller says they are hard of hearing, believe them and route on it; that is the one precise signal the record has.

A7 · the handoffDeploy confirm-before-connectingNot enough evidence on the naming

“Great. I’m going to bring a licensed agent onto the line now. His name is Marcus and he’ll pick up where we left off. Please hold for one moment.”

What we know
Confirming with the caller before connecting is settled on transfers. Two tests of handoff-announcement wording came back null. On the human stage a call that opens with the senior’s “Hello?” marks a failed handoff and runs shorter, median 124 seconds against 164, and its lower settle rate vanishes inside turn bands. On two senior pharmacy books the agent names the next human on 1% of long calls on one and 49% on the other; that is prevalence only, with no outcome contrast.
Why we think that
Confirm-before-connecting: transfers on the screening stage; unresolved on settled sale. The failed-handoff read: 454 human-stage calls, observational.
What we don’t know
Whether naming the licensed agent, or “pick up where we left off,” changes anything that settles. The record can say the handoff is where calls die; it cannot yet say which sentence saves them.
What would resolve it
Confirm-before-connecting against an assumed connect, randomised at the transfer line. On a neighbouring book the transfer read arrives in under two days and the billed read at day 14; settled sale is not powered inside 14 days at any number of arms, so the settled verdict waits for the enrolments to mature.
Next action
Deploy the confirm. Keep the name and the promise; they are harmless and untested. Grade the handoff on what settles, never on transfers alone.

A8 · the licensed agent’s first turnDeploy

“Mr. Alvarez, this is Marcus. Dana told me you take three medications and see Dr. Patel. I’ve checked both against the plan. They’re covered.”

What we know
The information A8 hands back is the information A3 collected, and today that hand-back mostly does not happen: on the platforms read, no field carries the caller’s screening answers to the licensed agent. On one senior pharmacy book the caller has to ask what their own part is on 80% of the calls where the topic comes up at all, twice the rate of a peer book; that book asks the senior to begin the transfer themselves on 28% of long calls. Telling the senior what happens next, before they ask, is the cheapest move the comparison surfaced.
Why we think that
Two senior pharmacy books compared on who raises each concept first, explicit speaker labels, prevalence and timing only. No outcome contrast sits under this row.
What we don’t know
Whether continuity across the handoff changes settlement. It has never been assigned. The record shows the gap, not the value of closing it.
What would resolve it
Pass the screening answers through, then randomise whether the licensed agent opens by repeating them or by re-asking, graded on settled enrolment.
Next action
Deploy as written, and build the hand-off of A3’s answers that makes it true. Say what happens next and who does it before the caller has to ask.

A9 · the closeDeploy the askDo not deploy the yes as the outcome

“So we’ve confirmed your doctors and your medications. Shall I go ahead and complete the enrollment with you now?”

What we know
On one senior pharmacy book, about half of the seniors an AI reads as positively engaged do not settle. Where the closing ceremony completed on tape, details taken and confirmation given, 68.2% settled [57.7, 77.2], and payment-capture language ran below the base rate, 13.3 percent against 16.1 percent. Readiness belongs to the interaction, not the person: on a fixed cohort of 3,895 calls that all ran 30 turns or more, nothing before the call and nothing in the first five turns ranks who enrols; the text reaches 0.607 at turn 10 and 0.772 at turn 30, after the eventual length of the call is held fixed. And the same senior, contacted twice 30 or more days apart, both calls connected, agrees with their own earlier enrolment decision at kappa +0.022 [−0.147, +0.190] on 191 pairs.
Why we think that
Human licensed-agent stage of a senior pharmacy programme, settled enrolment 30 or more days out. The test-retest interval is wide: the data exclude kappa above about 0.19 and cannot separate 0.02 from 0.10.
What we don’t know
Which close wording moves settlement. No assigned close test exists in the record. The record is clear that the yes is not the close, and silent on what is.
What would resolve it
Grade every close on settled enrolment 30 or more days out, never on the yes, then randomise two close wordings inside that grading. A yes-rate KPI will pick the wrong wording about as often as the right one.
Next action
Deploy the confirm-then-ask close. Do not deploy a “yes” as the number the floor is managed on. Expect many yeses not to become members, and build the follow-up on that expectation.

What to test first

Three elements, in the order the record ranks them.

1 · A4, where the cost line sitsCost-before-the-ask against a no-cost line inside the benefit sentence. Largest supported direction in the read, on both sides of the conversation, and the billed read arrives in days on ordinary volume. The state it depends on (a cost sentence) needs no detector.

2 · A7, confirm before connectingConfirm-before-connecting against an assumed connect at the transfer line. Settled on transfers already; the open question is whether it holds on the sale, and the transfer leg reads out fastest of anything here. The settled verdict waits for enrolments to mature.

3 · A5, the open question at “already have a plan”An open question against the current mix at the moment the caller says they already have a plan. The strongest moment-level association in the record, and the one that most needs a passport first: the state detector it rests on was 40% precise, and the Medicare Advantage analogue has never been measured.

Not on the list: the hearing check (A6), which the record says not to run; the purpose line (A2), which the record cannot see on this population; and the close wording (A9), which needs the grading fixed before any wording is worth testing.

What this read did not do

The boundaries, stated.

It did not forecast this script. Nothing above is an estimate of how this script would perform if dialled. It is evidence about the elements the script is built from, each with the population, channel and outcome it was measured on printed beside it. It did not read a Medicare Advantage call: every row comes from a senior pharmacy programme and neighbouring senior verticals, and the read says so on each line rather than once at the top. It did not randomise anything; every association here is observational unless the row names the assignment, and the two that do (the five-arm opener test and the confirm-before-connecting result) were graded on transfers or on a label that drops unobservable outcomes. It did not consult a panel, a survey or a simulated older adult, and it did not ask a model for its opinion of the lines. It did not check compliance: two claims in the script were sent to compliance rather than to the record, and the record has no view on them. Where a line could not be read, the read says so and names the step that would read it.

Send a script or a spot What the record already knows →