Senior Intelligence

What older customers actually do
is often not what people expect.

Comprenda has built a behavioural evidence base from real senior decisions. Here are some of the things it has taught us, including the ones that contradicted us.

Think you understand the 65+ customer? Two of these you can try yourself.

CurrentSay vs do

What people said changed the answer.

Before opening the outcome, we froze 17 business conclusions from stated evidence. After the outcome was opened:

5Held
4Reversed
8Disappeared
The finding
Conclusions a competent analyst would have published from stated evidence alone survived at less than one in three once the same people’s behavior and outcomes were joined in. The conclusions did not merely weaken, four of them pointed the other way.
Why it matters
Stated evidence and behavioral evidence answer different questions. Treating one as a proxy for the other is not a rounding error; it changes the business decision in a knowable fraction of cases.
Where it holds
One outcome-linked senior acquisition program, conversation stage, graded against settled enrollment. Conclusions were frozen in writing before the outcome was opened.
The receipt
Each of the 17 conclusions is recorded with the stated evidence that produced it, the frozen prediction, the outcome it was graded against, and the verdict. The freeze is what makes the count meaningful. A conclusion written after the outcome is not a prediction.
What it doesn’t establish
It does not establish that stated evidence is worthless, that the same ratio holds in another market, or that any specific reversal is causal. It establishes that stated evidence alone was incomplete on this population, at this stage, against this outcome.
What comes next
Replicate the freeze on a second program and a second outcome. The design is portable; only the outcome join has to exist.
CurrentMeasurement

The KPI leaderboard changed when we ranked by the outcome.

In one outcome-linked senior acquisition program, the fast operational KPI and settled enrollment ranked five acquisition sources almost independently.

Rank correlation: −0.10

Why it matters
A metric can be internally consistent, move reliably, and still point the business toward the wrong priority. The team optimizing the fast KPI was not making a mistake. They were making a well-executed decision on the wrong ranking.
Where it holds
Five acquisition sources in one program, over a period long enough for enrollment to settle.
What it doesn’t establish
Five sources is a small ranking. The correlation is a point estimate, and near-zero is compatible with a range of true values. What it does establish is that the two rankings are not interchangeable here. Which is the operational question.
What comes next
Run the same comparison wherever a fast metric is standing in for a slow outcome. The check is cheap and only needs both columns.
CurrentAudience

Purchase readiness mostly belonged to the interaction, not the person.

As repeated decisions were separated in time, most of the apparent person-level signal disappeared.

The finding
Pooled across contacts, purchase readiness looked like a stable property of the person. Once the same people’s decisions were separated in time, the person-level component collapsed and the great majority of the variance belonged to the interaction. A different observed state, price sensitivity, persisted much better over the same separation.
Why it matters
A momentary state should not quietly become a permanent customer label. Systems that write “purchase-ready” onto a profile and keep it are storing noise under a confident name.
Where it holds
People with repeated, time-separated contacts in one outcome-linked program.
What it doesn’t establish
It does not say readiness is never person-level, or that the same half-life applies to every state. The point is that half-lives differ, and have to be measured rather than assumed.
What comes next
Measure a half-life per stored property, and refuse to write a profile field that hasn’t got one.
CurrentJourney

The same signal changed meaning across the journey.

In one outcome-linked journey, four of five expressed states changed direction between two different stages.

Why it matters
Pooling customer touchpoints can create a very precise answer to the wrong question. The pooled estimate is not noisy, it is confidently the opposite, with a large sample and a tight interval behind it.
Where it holds
Two identified stages of one outcome-linked journey, same people, same outcome definition.
What it doesn’t establish
It does not identify which stage is “right”, both readings are correct about their own stage. It establishes that the stage is part of the claim, not context around it.
What comes next
Attach a stage coordinate to every expressed-state claim, and refuse the pooled write.
UpdatedAudience

The pattern was real. The explanation wasn’t.

Distrust rose sharply across successive contacts, roughly a six-fold rise when read across the customer base at one moment in time. The obvious reading is that repeated contact wears people down.

The finding
Followed within the people who actually reached four contacts, distrust was flat. The rise across the base was selection: those people’s first calls were already far more distrustful than average. Contact wasn’t creating distrust, distrustful people were the ones who stayed reachable.
Why it matters
Same data, same chart, opposite instruction. The cross-sectional read says “call less.” The within-person read says the contact policy is not the lever, and points at who is in the pool instead.
Where it holds
People who reached four or more contacts in one outcome-linked program.
What it doesn’t establish
It does not establish that repeated contact never affects trust, only that this six-fold rise is explained by who remains in the pool, not by what contact did to them.
What comes next
Any claim about a trend across contacts ships with the within-person read beside it. This one is the reason that rule exists.
CurrentSimulation & AI

We measured the exact shape our simulator gets wrong. So we can use it.

We compared what a simulated 65+ persona says it would ask about against what real people in the same situation actually raised. An unmeasured generator is a liability; a measured one is an instrument.

The finding
The gaps run in both directions and they are large. Concerns the model is confident about are rarer in reality than it assumes; concerns real people raise constantly, including whether they can hear the person speaking, are almost absent from the simulation. The simulation is not randomly wrong. It is wrong in a consistent, describable shape.
Why it matters
This is the argument for grounding labels rather than against simulation. A generated read that says where it came from is useful. One that presents a model prior as a customer insight is not.
What it doesn’t establish
It does not establish that simulation is useless for exploration. It is the cheapest way to widen a search. It establishes that the generator does not get to grade itself.
What comes next
Every simulated read renders with its grounding: person-grounded, cohort-grounded, or model prior.

Thirty seconds

Think you understand
the 65+ customer?

Four objections older customers raise on real calls. One of them costs far more than the others. Most people, including the floors that hear them every day, pick wrong.

Try it

Four objections older customers raise. Which one costs you the most enrollments when it comes up?

Try it yourself

Six things the industry coaches.
Which of them survive older adults?

Every call-analytics product a floor can buy learned its playbook on business-to-business software sales. That playbook is published. We put twenty-nine of those findings to our own tape, 14,903 recorded senior sessions, graded on the settled sale. Thirteen could be graded. Six of them are below.

The vendors are not named here. Their findings are almost certainly true where they were measured; the question is only whether they transfer.

Finding 1 of 6
Every figure here is an association on our own tape, not a randomised effect, and each is measured on the leg named beside it. Anything worth acting on becomes a test.

Three things the record says

Findings a senior floor
would not have guessed.

Found · the word yes

Engaged is not enrolled.

On one senior pharmacy book, about half of the seniors an AI reads as positively engaged do not settle.

Found · the completed close

About a third of finished sales never settled.

Where the closing ceremony completed on tape, details taken, confirmation given:68.2% settled [57.7, 77.2]. Payment-capture language ran below the base rate rather than above it.

A 31.8-point gap between the ceremony and the money

Found · the check nobody runs

Not one live agent checks whether it was heard.

Across a fleet of live senior agents, zero ask whether the person can hear them. Confirmed three independent ways on the same 150 calls. Two frontier readers and a registered detector, all returning zero, on a sample deliberately enriched for hearing-related words.

Corroboration, not one instrument's opinion

Observed rates on real outcomes, not treatment effects. And one that did not survive: a state we had built a coaching page around separated in no turn band at all once call length was held fixed. It came off the page.

What the instruments found

Four things we can measure
that nobody was measuring.

Observed · comprehension

Visible confusion is a poor proxy for understanding.

In a randomised study of televised risk statements, faster delivery reduced recall. In our senior phone corpus, faster delivery did not produce more requests to repeat. Two settings, two measures. And together they make “nobody sounded confused” a weak basis for concluding a message landed.

Measured · advertising

We read the spot, not the script.

Pace, placement and where the safety copy sits, from word-level timings, 193.0 wpm median against a published 190.2.

Settled · outcomes

You do not have to hand us an outcome feed.

A certified reader at AUC 0.9725. And, pointed at a book it was not certified on, it recognises 0 of 27. The abstention is the feature.

Updated · living evidence

Answers expire.

Walking a dated record forward, 19 of 52 first verdicts were contradicted by the next decisive window.

How each was validated →

Evidence states

Not every finding gets the same label.

Supported
Current evidence supports the claim, within its stated scope.
Ready to prove
Strong enough to justify a prospective test.
Unknown
An important question with insufficient evidence.
Contradicted
New evidence runs against the previous conclusion.
Retired
No longer used as a current claim, and kept visible rather than deleted.

No silent supersession.

Start anywhere

A hunch is enough to start with.

See what Comprenda knows Read the methods →